src/hotspot/cpu/sparc/stubGenerator_sparc.cpp
author shade
Thu, 07 Dec 2017 13:05:31 +0100
branchepsilon-gc-branch
changeset 55974 06122633fead
parent 55767 8e22715afabc
parent 48104 62d5973082e3
child 56021 864ee22719af
permissions -rw-r--r--
Merge
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
     1
/*
46381
020219e46c86 8150388: Remove SPARC 32-bit support
gtriantafill
parents: 38246
diff changeset
     2
 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
     3
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
489c9b5090e2 Initial load
duke
parents:
diff changeset
     4
 *
489c9b5090e2 Initial load
duke
parents:
diff changeset
     5
 * This code is free software; you can redistribute it and/or modify it
489c9b5090e2 Initial load
duke
parents:
diff changeset
     6
 * under the terms of the GNU General Public License version 2 only, as
489c9b5090e2 Initial load
duke
parents:
diff changeset
     7
 * published by the Free Software Foundation.
489c9b5090e2 Initial load
duke
parents:
diff changeset
     8
 *
489c9b5090e2 Initial load
duke
parents:
diff changeset
     9
 * This code is distributed in the hope that it will be useful, but WITHOUT
489c9b5090e2 Initial load
duke
parents:
diff changeset
    10
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
489c9b5090e2 Initial load
duke
parents:
diff changeset
    11
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
489c9b5090e2 Initial load
duke
parents:
diff changeset
    12
 * version 2 for more details (a copy is included in the LICENSE file that
489c9b5090e2 Initial load
duke
parents:
diff changeset
    13
 * accompanied this code).
489c9b5090e2 Initial load
duke
parents:
diff changeset
    14
 *
489c9b5090e2 Initial load
duke
parents:
diff changeset
    15
 * You should have received a copy of the GNU General Public License version
489c9b5090e2 Initial load
duke
parents:
diff changeset
    16
 * 2 along with this work; if not, write to the Free Software Foundation,
489c9b5090e2 Initial load
duke
parents:
diff changeset
    17
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
489c9b5090e2 Initial load
duke
parents:
diff changeset
    18
 *
5547
f4b087cbb361 6941466: Oracle rebranding changes for Hotspot repositories
trims
parents: 5419
diff changeset
    19
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
f4b087cbb361 6941466: Oracle rebranding changes for Hotspot repositories
trims
parents: 5419
diff changeset
    20
 * or visit www.oracle.com if you need additional information or have any
f4b087cbb361 6941466: Oracle rebranding changes for Hotspot repositories
trims
parents: 5419
diff changeset
    21
 * questions.
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    22
 *
489c9b5090e2 Initial load
duke
parents:
diff changeset
    23
 */
489c9b5090e2 Initial load
duke
parents:
diff changeset
    24
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    25
#include "precompiled.hpp"
14631
526804361522 8003250: SPARC: move MacroAssembler into separate file
twisti
parents: 13952
diff changeset
    26
#include "asm/macroAssembler.inline.hpp"
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    27
#include "interpreter/interpreter.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    28
#include "nativeInst_sparc.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    29
#include "oops/instanceOop.hpp"
13728
882756847a04 6964458: Reimplement class meta-data storage to use native memory
coleenp
parents: 13391
diff changeset
    30
#include "oops/method.hpp"
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    31
#include "oops/objArrayKlass.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    32
#include "oops/oop.inline.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    33
#include "prims/methodHandles.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    34
#include "runtime/frame.inline.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    35
#include "runtime/handles.inline.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    36
#include "runtime/sharedRuntime.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    37
#include "runtime/stubCodeGenerator.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    38
#include "runtime/stubRoutines.hpp"
14583
d70ee55535f4 8003935: Simplify the needed includes for using Thread::current()
stefank
parents: 13952
diff changeset
    39
#include "runtime/thread.inline.hpp"
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    40
#ifdef COMPILER2
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    41
#include "opto/runtime.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6770
diff changeset
    42
#endif
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    43
489c9b5090e2 Initial load
duke
parents:
diff changeset
    44
// Declaration and definition of StubGenerator (no .hpp file).
489c9b5090e2 Initial load
duke
parents:
diff changeset
    45
// For a more detailed description of the stub routine structure
489c9b5090e2 Initial load
duke
parents:
diff changeset
    46
// see the comment in stubRoutines.hpp.
489c9b5090e2 Initial load
duke
parents:
diff changeset
    47
489c9b5090e2 Initial load
duke
parents:
diff changeset
    48
#define __ _masm->
489c9b5090e2 Initial load
duke
parents:
diff changeset
    49
489c9b5090e2 Initial load
duke
parents:
diff changeset
    50
#ifdef PRODUCT
489c9b5090e2 Initial load
duke
parents:
diff changeset
    51
#define BLOCK_COMMENT(str) /* nothing */
489c9b5090e2 Initial load
duke
parents:
diff changeset
    52
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
    53
#define BLOCK_COMMENT(str) __ block_comment(str)
489c9b5090e2 Initial load
duke
parents:
diff changeset
    54
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
    55
489c9b5090e2 Initial load
duke
parents:
diff changeset
    56
#define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
489c9b5090e2 Initial load
duke
parents:
diff changeset
    57
489c9b5090e2 Initial load
duke
parents:
diff changeset
    58
// Note:  The register L7 is used as L7_thread_cache, and may not be used
489c9b5090e2 Initial load
duke
parents:
diff changeset
    59
//        any other way within this module.
489c9b5090e2 Initial load
duke
parents:
diff changeset
    60
489c9b5090e2 Initial load
duke
parents:
diff changeset
    61
static const Register& Lstub_temp = L2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
    62
489c9b5090e2 Initial load
duke
parents:
diff changeset
    63
// -------------------------------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
    64
// Stub Code definitions
489c9b5090e2 Initial load
duke
parents:
diff changeset
    65
489c9b5090e2 Initial load
duke
parents:
diff changeset
    66
class StubGenerator: public StubCodeGenerator {
489c9b5090e2 Initial load
duke
parents:
diff changeset
    67
 private:
489c9b5090e2 Initial load
duke
parents:
diff changeset
    68
489c9b5090e2 Initial load
duke
parents:
diff changeset
    69
#ifdef PRODUCT
24326
d3fdd5c16fe0 8022070: Compilation error in stubGenerator_sparc.cpp with some compilers
mikael
parents: 22505
diff changeset
    70
#define inc_counter_np(a,b,c)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    71
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
    72
#define inc_counter_np(counter, t1, t2) \
489c9b5090e2 Initial load
duke
parents:
diff changeset
    73
  BLOCK_COMMENT("inc_counter " #counter); \
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
    74
  __ inc_counter(&counter, t1, t2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    75
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
    76
489c9b5090e2 Initial load
duke
parents:
diff changeset
    77
  //----------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
    78
  // Call stubs are used to call Java from C
489c9b5090e2 Initial load
duke
parents:
diff changeset
    79
489c9b5090e2 Initial load
duke
parents:
diff changeset
    80
  address generate_call_stub(address& return_pc) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
    81
    StubCodeMark mark(this, "StubRoutines", "call_stub");
489c9b5090e2 Initial load
duke
parents:
diff changeset
    82
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
    83
489c9b5090e2 Initial load
duke
parents:
diff changeset
    84
    // Incoming arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
    85
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
    86
    // o0         : call wrapper address
489c9b5090e2 Initial load
duke
parents:
diff changeset
    87
    // o1         : result (address)
489c9b5090e2 Initial load
duke
parents:
diff changeset
    88
    // o2         : result type
489c9b5090e2 Initial load
duke
parents:
diff changeset
    89
    // o3         : method
489c9b5090e2 Initial load
duke
parents:
diff changeset
    90
    // o4         : (interpreter) entry point
489c9b5090e2 Initial load
duke
parents:
diff changeset
    91
    // o5         : parameters (address)
489c9b5090e2 Initial load
duke
parents:
diff changeset
    92
    // [sp + 0x5c]: parameter size (in words)
489c9b5090e2 Initial load
duke
parents:
diff changeset
    93
    // [sp + 0x60]: thread
489c9b5090e2 Initial load
duke
parents:
diff changeset
    94
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
    95
    // +---------------+ <--- sp + 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
    96
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
    97
    // . reg save area .
489c9b5090e2 Initial load
duke
parents:
diff changeset
    98
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
    99
    // +---------------+ <--- sp + 0x40
489c9b5090e2 Initial load
duke
parents:
diff changeset
   100
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   101
    // . extra 7 slots .
489c9b5090e2 Initial load
duke
parents:
diff changeset
   102
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   103
    // +---------------+ <--- sp + 0x5c
489c9b5090e2 Initial load
duke
parents:
diff changeset
   104
    // |  param. size  |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   105
    // +---------------+ <--- sp + 0x60
489c9b5090e2 Initial load
duke
parents:
diff changeset
   106
    // |    thread     |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   107
    // +---------------+
489c9b5090e2 Initial load
duke
parents:
diff changeset
   108
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   109
489c9b5090e2 Initial load
duke
parents:
diff changeset
   110
    // note: if the link argument position changes, adjust
489c9b5090e2 Initial load
duke
parents:
diff changeset
   111
    //       the code in frame::entry_frame_call_wrapper()
489c9b5090e2 Initial load
duke
parents:
diff changeset
   112
489c9b5090e2 Initial load
duke
parents:
diff changeset
   113
    const Argument link           = Argument(0, false); // used only for GC
489c9b5090e2 Initial load
duke
parents:
diff changeset
   114
    const Argument result         = Argument(1, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   115
    const Argument result_type    = Argument(2, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   116
    const Argument method         = Argument(3, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   117
    const Argument entry_point    = Argument(4, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   118
    const Argument parameters     = Argument(5, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   119
    const Argument parameter_size = Argument(6, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   120
    const Argument thread         = Argument(7, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   121
489c9b5090e2 Initial load
duke
parents:
diff changeset
   122
    // setup thread register
489c9b5090e2 Initial load
duke
parents:
diff changeset
   123
    __ ld_ptr(thread.as_address(), G2_thread);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
   124
    __ reinit_heapbase();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   125
489c9b5090e2 Initial load
duke
parents:
diff changeset
   126
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   127
    // make sure we have no pending exceptions
489c9b5090e2 Initial load
duke
parents:
diff changeset
   128
    { const Register t = G3_scratch;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   129
      Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   130
      __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), t);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   131
      __ br_null_short(t, Assembler::pt, L);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   132
      __ stop("StubRoutines::call_stub: entered with pending exception");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   133
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   134
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   135
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   136
489c9b5090e2 Initial load
duke
parents:
diff changeset
   137
    // create activation frame & allocate space for parameters
489c9b5090e2 Initial load
duke
parents:
diff changeset
   138
    { const Register t = G3_scratch;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   139
      __ ld_ptr(parameter_size.as_address(), t);                // get parameter size (in words)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   140
      __ add(t, frame::memory_parameter_word_sp_offset, t);     // add space for save area (in words)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   141
      __ round_to(t, WordsPerLong);                             // make sure it is multiple of 2 (in words)
5419
f2e8cc8c12ea 6943304: remove tagged stack interpreter
twisti
parents: 5416
diff changeset
   142
      __ sll(t, Interpreter::logStackElementSize, t);           // compute number of bytes
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   143
      __ neg(t);                                                // negate so it can be used with save
489c9b5090e2 Initial load
duke
parents:
diff changeset
   144
      __ save(SP, t, SP);                                       // setup new frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
   145
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   146
489c9b5090e2 Initial load
duke
parents:
diff changeset
   147
    // +---------------+ <--- sp + 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   148
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   149
    // . reg save area .
489c9b5090e2 Initial load
duke
parents:
diff changeset
   150
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   151
    // +---------------+ <--- sp + 0x40
489c9b5090e2 Initial load
duke
parents:
diff changeset
   152
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   153
    // . extra 7 slots .
489c9b5090e2 Initial load
duke
parents:
diff changeset
   154
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   155
    // +---------------+ <--- sp + 0x5c
489c9b5090e2 Initial load
duke
parents:
diff changeset
   156
    // |  empty slot   |      (only if parameter size is even)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   157
    // +---------------+
489c9b5090e2 Initial load
duke
parents:
diff changeset
   158
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   159
    // .  parameters   .
489c9b5090e2 Initial load
duke
parents:
diff changeset
   160
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   161
    // +---------------+ <--- fp + 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   162
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   163
    // . reg save area .
489c9b5090e2 Initial load
duke
parents:
diff changeset
   164
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   165
    // +---------------+ <--- fp + 0x40
489c9b5090e2 Initial load
duke
parents:
diff changeset
   166
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   167
    // . extra 7 slots .
489c9b5090e2 Initial load
duke
parents:
diff changeset
   168
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   169
    // +---------------+ <--- fp + 0x5c
489c9b5090e2 Initial load
duke
parents:
diff changeset
   170
    // |  param. size  |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   171
    // +---------------+ <--- fp + 0x60
489c9b5090e2 Initial load
duke
parents:
diff changeset
   172
    // |    thread     |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   173
    // +---------------+
489c9b5090e2 Initial load
duke
parents:
diff changeset
   174
    // |               |
489c9b5090e2 Initial load
duke
parents:
diff changeset
   175
489c9b5090e2 Initial load
duke
parents:
diff changeset
   176
    // pass parameters if any
489c9b5090e2 Initial load
duke
parents:
diff changeset
   177
    BLOCK_COMMENT("pass parameters if any");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   178
    { const Register src = parameters.as_in().as_register();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   179
      const Register dst = Lentry_args;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   180
      const Register tmp = G3_scratch;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   181
      const Register cnt = G4_scratch;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   182
489c9b5090e2 Initial load
duke
parents:
diff changeset
   183
      // test if any parameters & setup of Lentry_args
489c9b5090e2 Initial load
duke
parents:
diff changeset
   184
      Label exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   185
      __ ld_ptr(parameter_size.as_in().as_address(), cnt);      // parameter counter
489c9b5090e2 Initial load
duke
parents:
diff changeset
   186
      __ add( FP, STACK_BIAS, dst );
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   187
      __ cmp_zero_and_br(Assembler::zero, cnt, exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   188
      __ delayed()->sub(dst, BytesPerWord, dst);                 // setup Lentry_args
489c9b5090e2 Initial load
duke
parents:
diff changeset
   189
489c9b5090e2 Initial load
duke
parents:
diff changeset
   190
      // copy parameters if any
489c9b5090e2 Initial load
duke
parents:
diff changeset
   191
      Label loop;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   192
      __ BIND(loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   193
      // Store parameter value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   194
      __ ld_ptr(src, 0, tmp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   195
      __ add(src, BytesPerWord, src);
5419
f2e8cc8c12ea 6943304: remove tagged stack interpreter
twisti
parents: 5416
diff changeset
   196
      __ st_ptr(tmp, dst, 0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   197
      __ deccc(cnt);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   198
      __ br(Assembler::greater, false, Assembler::pt, loop);
5419
f2e8cc8c12ea 6943304: remove tagged stack interpreter
twisti
parents: 5416
diff changeset
   199
      __ delayed()->sub(dst, Interpreter::stackElementSize, dst);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   200
489c9b5090e2 Initial load
duke
parents:
diff changeset
   201
      // done
489c9b5090e2 Initial load
duke
parents:
diff changeset
   202
      __ BIND(exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   203
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   204
489c9b5090e2 Initial load
duke
parents:
diff changeset
   205
    // setup parameters, method & call Java function
489c9b5090e2 Initial load
duke
parents:
diff changeset
   206
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   207
    // layout_activation_impl checks it's notion of saved SP against
489c9b5090e2 Initial load
duke
parents:
diff changeset
   208
    // this register, so if this changes update it as well.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   209
    const Register saved_SP = Lscratch;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   210
    __ mov(SP, saved_SP);                               // keep track of SP before call
489c9b5090e2 Initial load
duke
parents:
diff changeset
   211
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   212
489c9b5090e2 Initial load
duke
parents:
diff changeset
   213
    // setup parameters
489c9b5090e2 Initial load
duke
parents:
diff changeset
   214
    const Register t = G3_scratch;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   215
    __ ld_ptr(parameter_size.as_in().as_address(), t); // get parameter size (in words)
5419
f2e8cc8c12ea 6943304: remove tagged stack interpreter
twisti
parents: 5416
diff changeset
   216
    __ sll(t, Interpreter::logStackElementSize, t);    // compute number of bytes
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   217
    __ sub(FP, t, Gargs);                              // setup parameter pointer
489c9b5090e2 Initial load
duke
parents:
diff changeset
   218
    __ add( Gargs, STACK_BIAS, Gargs );                // Account for LP64 stack bias
489c9b5090e2 Initial load
duke
parents:
diff changeset
   219
    __ mov(SP, O5_savedSP);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   220
489c9b5090e2 Initial load
duke
parents:
diff changeset
   221
489c9b5090e2 Initial load
duke
parents:
diff changeset
   222
    // do the call
489c9b5090e2 Initial load
duke
parents:
diff changeset
   223
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   224
    // the following register must be setup:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   225
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   226
    // G2_thread
489c9b5090e2 Initial load
duke
parents:
diff changeset
   227
    // G5_method
489c9b5090e2 Initial load
duke
parents:
diff changeset
   228
    // Gargs
489c9b5090e2 Initial load
duke
parents:
diff changeset
   229
    BLOCK_COMMENT("call Java function");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   230
    __ jmpl(entry_point.as_in().as_register(), G0, O7);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   231
    __ delayed()->mov(method.as_in().as_register(), G5_method);   // setup method
489c9b5090e2 Initial load
duke
parents:
diff changeset
   232
489c9b5090e2 Initial load
duke
parents:
diff changeset
   233
    BLOCK_COMMENT("call_stub_return_address:");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   234
    return_pc = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   235
489c9b5090e2 Initial load
duke
parents:
diff changeset
   236
    // The callee, if it wasn't interpreted, can return with SP changed so
489c9b5090e2 Initial load
duke
parents:
diff changeset
   237
    // we can no longer assert of change of SP.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   238
489c9b5090e2 Initial load
duke
parents:
diff changeset
   239
    // store result depending on type
489c9b5090e2 Initial load
duke
parents:
diff changeset
   240
    // (everything that is not T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE
489c9b5090e2 Initial load
duke
parents:
diff changeset
   241
    //  is treated as T_INT)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   242
    { const Register addr = result     .as_in().as_register();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   243
      const Register type = result_type.as_in().as_register();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   244
      Label is_long, is_float, is_double, is_object, exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   245
      __            cmp(type, T_OBJECT);  __ br(Assembler::equal, false, Assembler::pn, is_object);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   246
      __ delayed()->cmp(type, T_FLOAT);   __ br(Assembler::equal, false, Assembler::pn, is_float);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   247
      __ delayed()->cmp(type, T_DOUBLE);  __ br(Assembler::equal, false, Assembler::pn, is_double);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   248
      __ delayed()->cmp(type, T_LONG);    __ br(Assembler::equal, false, Assembler::pn, is_long);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   249
      __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   250
489c9b5090e2 Initial load
duke
parents:
diff changeset
   251
      // store int result
489c9b5090e2 Initial load
duke
parents:
diff changeset
   252
      __ st(O0, addr, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   253
489c9b5090e2 Initial load
duke
parents:
diff changeset
   254
      __ BIND(exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   255
      __ ret();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   256
      __ delayed()->restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   257
489c9b5090e2 Initial load
duke
parents:
diff changeset
   258
      __ BIND(is_object);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   259
      __ ba(exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   260
      __ delayed()->st_ptr(O0, addr, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   261
489c9b5090e2 Initial load
duke
parents:
diff changeset
   262
      __ BIND(is_float);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   263
      __ ba(exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   264
      __ delayed()->stf(FloatRegisterImpl::S, F0, addr, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   265
489c9b5090e2 Initial load
duke
parents:
diff changeset
   266
      __ BIND(is_double);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   267
      __ ba(exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   268
      __ delayed()->stf(FloatRegisterImpl::D, F0, addr, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   269
489c9b5090e2 Initial load
duke
parents:
diff changeset
   270
      __ BIND(is_long);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   271
      __ ba(exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   272
      __ delayed()->st_long(O0, addr, G0);      // store entire long
489c9b5090e2 Initial load
duke
parents:
diff changeset
   273
     }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   274
     return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   275
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   276
489c9b5090e2 Initial load
duke
parents:
diff changeset
   277
489c9b5090e2 Initial load
duke
parents:
diff changeset
   278
  //----------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
   279
  // Return point for a Java call if there's an exception thrown in Java code.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   280
  // The exception is caught and transformed into a pending exception stored in
489c9b5090e2 Initial load
duke
parents:
diff changeset
   281
  // JavaThread that can be tested from within the VM.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   282
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   283
  // Oexception: exception oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
   284
489c9b5090e2 Initial load
duke
parents:
diff changeset
   285
  address generate_catch_exception() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   286
    StubCodeMark mark(this, "StubRoutines", "catch_exception");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   287
489c9b5090e2 Initial load
duke
parents:
diff changeset
   288
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   289
    // verify that thread corresponds
489c9b5090e2 Initial load
duke
parents:
diff changeset
   290
    __ verify_thread();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   291
489c9b5090e2 Initial load
duke
parents:
diff changeset
   292
    const Register& temp_reg = Gtemp;
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   293
    Address pending_exception_addr    (G2_thread, Thread::pending_exception_offset());
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   294
    Address exception_file_offset_addr(G2_thread, Thread::exception_file_offset   ());
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   295
    Address exception_line_offset_addr(G2_thread, Thread::exception_line_offset   ());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   296
489c9b5090e2 Initial load
duke
parents:
diff changeset
   297
    // set pending exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   298
    __ verify_oop(Oexception);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   299
    __ st_ptr(Oexception, pending_exception_addr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   300
    __ set((intptr_t)__FILE__, temp_reg);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   301
    __ st_ptr(temp_reg, exception_file_offset_addr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   302
    __ set((intptr_t)__LINE__, temp_reg);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   303
    __ st(temp_reg, exception_line_offset_addr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   304
489c9b5090e2 Initial load
duke
parents:
diff changeset
   305
    // complete return to VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
   306
    assert(StubRoutines::_call_stub_return_address != NULL, "must have been generated before");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   307
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   308
    AddressLiteral stub_ret(StubRoutines::_call_stub_return_address);
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   309
    __ jump_to(stub_ret, temp_reg);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   310
    __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   311
489c9b5090e2 Initial load
duke
parents:
diff changeset
   312
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   313
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   314
489c9b5090e2 Initial load
duke
parents:
diff changeset
   315
489c9b5090e2 Initial load
duke
parents:
diff changeset
   316
  //----------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
   317
  // Continuation point for runtime calls returning with a pending exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   318
  // The pending exception check happened in the runtime or native call stub
489c9b5090e2 Initial load
duke
parents:
diff changeset
   319
  // The pending exception in Thread is converted into a Java-level exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   320
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   321
  // Contract with Java-level exception handler: O0 = exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   322
  //                                             O1 = throwing pc
489c9b5090e2 Initial load
duke
parents:
diff changeset
   323
489c9b5090e2 Initial load
duke
parents:
diff changeset
   324
  address generate_forward_exception() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   325
    StubCodeMark mark(this, "StubRoutines", "forward_exception");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   326
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   327
489c9b5090e2 Initial load
duke
parents:
diff changeset
   328
    // Upon entry, O7 has the return address returning into Java
489c9b5090e2 Initial load
duke
parents:
diff changeset
   329
    // (interpreted or compiled) code; i.e. the return address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   330
    // becomes the throwing pc.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   331
489c9b5090e2 Initial load
duke
parents:
diff changeset
   332
    const Register& handler_reg = Gtemp;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   333
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   334
    Address exception_addr(G2_thread, Thread::pending_exception_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   335
489c9b5090e2 Initial load
duke
parents:
diff changeset
   336
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   337
    // make sure that this code is only executed if there is a pending exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   338
    { Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   339
      __ ld_ptr(exception_addr, Gtemp);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   340
      __ br_notnull_short(Gtemp, Assembler::pt, L);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   341
      __ stop("StubRoutines::forward exception: no pending exception (1)");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   342
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   343
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   344
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   345
489c9b5090e2 Initial load
duke
parents:
diff changeset
   346
    // compute exception handler into handler_reg
489c9b5090e2 Initial load
duke
parents:
diff changeset
   347
    __ get_thread();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   348
    __ ld_ptr(exception_addr, Oexception);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   349
    __ verify_oop(Oexception);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   350
    __ save_frame(0);             // compensates for compiler weakness
489c9b5090e2 Initial load
duke
parents:
diff changeset
   351
    __ add(O7->after_save(), frame::pc_return_offset, Lscratch); // save the issuing PC
489c9b5090e2 Initial load
duke
parents:
diff changeset
   352
    BLOCK_COMMENT("call exception_handler_for_return_address");
5046
27e801a857cb 6919934: JSR 292 needs to support x86 C1
twisti
parents: 4645
diff changeset
   353
    __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), G2_thread, Lscratch);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   354
    __ mov(O0, handler_reg);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   355
    __ restore();                 // compensates for compiler weakness
489c9b5090e2 Initial load
duke
parents:
diff changeset
   356
489c9b5090e2 Initial load
duke
parents:
diff changeset
   357
    __ ld_ptr(exception_addr, Oexception);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   358
    __ add(O7, frame::pc_return_offset, Oissuing_pc); // save the issuing PC
489c9b5090e2 Initial load
duke
parents:
diff changeset
   359
489c9b5090e2 Initial load
duke
parents:
diff changeset
   360
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   361
    // make sure exception is set
489c9b5090e2 Initial load
duke
parents:
diff changeset
   362
    { Label L;
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   363
      __ br_notnull_short(Oexception, Assembler::pt, L);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   364
      __ stop("StubRoutines::forward exception: no pending exception (2)");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   365
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   366
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   367
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   368
    // jump to exception handler
489c9b5090e2 Initial load
duke
parents:
diff changeset
   369
    __ jmp(handler_reg, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   370
    // clear pending exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   371
    __ delayed()->st_ptr(G0, exception_addr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   372
489c9b5090e2 Initial load
duke
parents:
diff changeset
   373
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   374
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   375
18740
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   376
  // Safefetch stubs.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   377
  void generate_safefetch(const char* name, int size, address* entry,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   378
                          address* fault_pc, address* continuation_pc) {
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   379
    // safefetch signatures:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   380
    //   int      SafeFetch32(int*      adr, int      errValue);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   381
    //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   382
    //
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   383
    // arguments:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   384
    //   o0 = adr
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   385
    //   o1 = errValue
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   386
    //
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   387
    // result:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   388
    //   o0  = *adr or errValue
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   389
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   390
    StubCodeMark mark(this, "StubRoutines", name);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   391
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   392
    // Entry point, pc or function descriptor.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   393
    __ align(CodeEntryAlignment);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   394
    *entry = __ pc();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   395
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   396
    __ mov(O0, G1);  // g1 = o0
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   397
    __ mov(O1, O0);  // o0 = o1
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   398
    // Load *adr into c_rarg1, may fault.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   399
    *fault_pc = __ pc();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   400
    switch (size) {
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   401
      case 4:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   402
        // int32_t
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   403
        __ ldsw(G1, 0, O0);  // o0 = [g1]
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   404
        break;
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   405
      case 8:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   406
        // int64_t
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   407
        __ ldx(G1, 0, O0);   // o0 = [g1]
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   408
        break;
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   409
      default:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   410
        ShouldNotReachHere();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   411
    }
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   412
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   413
    // return errValue or *adr
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   414
    *continuation_pc = __ pc();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   415
    // By convention with the trap handler we ensure there is a non-CTI
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   416
    // instruction in the trap shadow.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   417
    __ nop();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   418
    __ retl();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   419
    __ delayed()->nop();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
   420
  }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   421
489c9b5090e2 Initial load
duke
parents:
diff changeset
   422
  //------------------------------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
   423
  // Continuation point for throwing of implicit exceptions that are not handled in
489c9b5090e2 Initial load
duke
parents:
diff changeset
   424
  // the current activation. Fabricates an exception oop and initiates normal
489c9b5090e2 Initial load
duke
parents:
diff changeset
   425
  // exception dispatching in this frame. Only callee-saved registers are preserved
489c9b5090e2 Initial load
duke
parents:
diff changeset
   426
  // (through the normal register window / RegisterMap handling).
489c9b5090e2 Initial load
duke
parents:
diff changeset
   427
  // If the compiler needs all registers to be preserved between the fault
489c9b5090e2 Initial load
duke
parents:
diff changeset
   428
  // point and the exception handler then it must assume responsibility for that in
489c9b5090e2 Initial load
duke
parents:
diff changeset
   429
  // AbstractCompiler::continuation_for_implicit_null_exception or
489c9b5090e2 Initial load
duke
parents:
diff changeset
   430
  // continuation_for_implicit_division_by_zero_exception. All other implicit
489c9b5090e2 Initial load
duke
parents:
diff changeset
   431
  // exceptions (e.g., NullPointerException or AbstractMethodError on entry) are
489c9b5090e2 Initial load
duke
parents:
diff changeset
   432
  // either at call sites or otherwise assume that stack unwinding will be initiated,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   433
  // so caller saved registers were assumed volatile in the compiler.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   434
489c9b5090e2 Initial load
duke
parents:
diff changeset
   435
  // Note that we generate only this stub into a RuntimeStub, because it needs to be
489c9b5090e2 Initial load
duke
parents:
diff changeset
   436
  // properly traversed and ignored during GC, so we change the meaning of the "__"
489c9b5090e2 Initial load
duke
parents:
diff changeset
   437
  // macro within this method.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   438
#undef __
489c9b5090e2 Initial load
duke
parents:
diff changeset
   439
#define __ masm->
489c9b5090e2 Initial load
duke
parents:
diff changeset
   440
10545
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10512
diff changeset
   441
  address generate_throw_exception(const char* name, address runtime_entry,
10004
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   442
                                   Register arg1 = noreg, Register arg2 = noreg) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   443
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   444
    int insts_size = VerifyThread ? 1 * K : 600;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   445
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   446
    int insts_size = VerifyThread ? 1 * K : 256;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   447
#endif /* ASSERT */
489c9b5090e2 Initial load
duke
parents:
diff changeset
   448
    int locs_size  = 32;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   449
489c9b5090e2 Initial load
duke
parents:
diff changeset
   450
    CodeBuffer      code(name, insts_size, locs_size);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   451
    MacroAssembler* masm = new MacroAssembler(&code);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   452
489c9b5090e2 Initial load
duke
parents:
diff changeset
   453
    __ verify_thread();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   454
489c9b5090e2 Initial load
duke
parents:
diff changeset
   455
    // This is an inlined and slightly modified version of call_VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
   456
    // which has the ability to fetch the return PC out of thread-local storage
489c9b5090e2 Initial load
duke
parents:
diff changeset
   457
    __ assert_not_delayed();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   458
489c9b5090e2 Initial load
duke
parents:
diff changeset
   459
    // Note that we always push a frame because on the SPARC
489c9b5090e2 Initial load
duke
parents:
diff changeset
   460
    // architecture, for all of our implicit exception kinds at call
489c9b5090e2 Initial load
duke
parents:
diff changeset
   461
    // sites, the implicit exception is taken before the callee frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
   462
    // is pushed.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   463
    __ save_frame(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   464
489c9b5090e2 Initial load
duke
parents:
diff changeset
   465
    int frame_complete = __ offset();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   466
489c9b5090e2 Initial load
duke
parents:
diff changeset
   467
    // Note that we always have a runtime stub frame on the top of stack by this point
489c9b5090e2 Initial load
duke
parents:
diff changeset
   468
    Register last_java_sp = SP;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   469
    // 64-bit last_java_sp is biased!
489c9b5090e2 Initial load
duke
parents:
diff changeset
   470
    __ set_last_Java_frame(last_java_sp, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   471
    if (VerifyThread)  __ mov(G2_thread, O0); // about to be smashed; pass early
489c9b5090e2 Initial load
duke
parents:
diff changeset
   472
    __ save_thread(noreg);
10004
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   473
    if (arg1 != noreg) {
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   474
      assert(arg2 != O1, "clobbered");
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   475
      __ mov(arg1, O1);
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   476
    }
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   477
    if (arg2 != noreg) {
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   478
      __ mov(arg2, O2);
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8498
diff changeset
   479
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   480
    // do the call
489c9b5090e2 Initial load
duke
parents:
diff changeset
   481
    BLOCK_COMMENT("call runtime_entry");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   482
    __ call(runtime_entry, relocInfo::runtime_call_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   483
    if (!VerifyThread)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   484
      __ delayed()->mov(G2_thread, O0);  // pass thread as first argument
489c9b5090e2 Initial load
duke
parents:
diff changeset
   485
    else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   486
      __ delayed()->nop();             // (thread already passed)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   487
    __ restore_thread(noreg);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   488
    __ reset_last_Java_frame();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   489
489c9b5090e2 Initial load
duke
parents:
diff changeset
   490
    // check for pending exceptions. use Gtemp as scratch register.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   491
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   492
    Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   493
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   494
    Address exception_addr(G2_thread, Thread::pending_exception_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   495
    Register scratch_reg = Gtemp;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   496
    __ ld_ptr(exception_addr, scratch_reg);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   497
    __ br_notnull_short(scratch_reg, Assembler::pt, L);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   498
    __ should_not_reach_here();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   499
    __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   500
#endif // ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   501
    BLOCK_COMMENT("call forward_exception_entry");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   502
    __ call(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   503
    // we use O7 linkage so that forward_exception_entry has the issuing PC
489c9b5090e2 Initial load
duke
parents:
diff changeset
   504
    __ delayed()->restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   505
489c9b5090e2 Initial load
duke
parents:
diff changeset
   506
    RuntimeStub* stub = RuntimeStub::new_runtime_stub(name, &code, frame_complete, masm->total_frame_size_in_bytes(0), NULL, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   507
    return stub->entry_point();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   508
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   509
489c9b5090e2 Initial load
duke
parents:
diff changeset
   510
#undef __
489c9b5090e2 Initial load
duke
parents:
diff changeset
   511
#define __ _masm->
489c9b5090e2 Initial load
duke
parents:
diff changeset
   512
489c9b5090e2 Initial load
duke
parents:
diff changeset
   513
489c9b5090e2 Initial load
duke
parents:
diff changeset
   514
  // Generate a routine that sets all the registers so we
489c9b5090e2 Initial load
duke
parents:
diff changeset
   515
  // can tell if the stop routine prints them correctly.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   516
  address generate_test_stop() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   517
    StubCodeMark mark(this, "StubRoutines", "test_stop");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   518
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   519
489c9b5090e2 Initial load
duke
parents:
diff changeset
   520
    int i;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   521
489c9b5090e2 Initial load
duke
parents:
diff changeset
   522
    __ save_frame(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   523
489c9b5090e2 Initial load
duke
parents:
diff changeset
   524
    static jfloat zero = 0.0, one = 1.0;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   525
489c9b5090e2 Initial load
duke
parents:
diff changeset
   526
    // put addr in L0, then load through L0 to F0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   527
    __ set((intptr_t)&zero, L0);  __ ldf( FloatRegisterImpl::S, L0, 0, F0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   528
    __ set((intptr_t)&one,  L0);  __ ldf( FloatRegisterImpl::S, L0, 0, F1); // 1.0 to F1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   529
489c9b5090e2 Initial load
duke
parents:
diff changeset
   530
    // use add to put 2..18 in F2..F18
489c9b5090e2 Initial load
duke
parents:
diff changeset
   531
    for ( i = 2;  i <= 18;  ++i ) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   532
      __ fadd( FloatRegisterImpl::S, F1, as_FloatRegister(i-1),  as_FloatRegister(i));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   533
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   534
489c9b5090e2 Initial load
duke
parents:
diff changeset
   535
    // Now put double 2 in F16, double 18 in F18
489c9b5090e2 Initial load
duke
parents:
diff changeset
   536
    __ ftof( FloatRegisterImpl::S, FloatRegisterImpl::D, F2, F16 );
489c9b5090e2 Initial load
duke
parents:
diff changeset
   537
    __ ftof( FloatRegisterImpl::S, FloatRegisterImpl::D, F18, F18 );
489c9b5090e2 Initial load
duke
parents:
diff changeset
   538
489c9b5090e2 Initial load
duke
parents:
diff changeset
   539
    // use add to put 20..32 in F20..F32
489c9b5090e2 Initial load
duke
parents:
diff changeset
   540
    for (i = 20; i < 32; i += 2) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   541
      __ fadd( FloatRegisterImpl::D, F16, as_FloatRegister(i-2),  as_FloatRegister(i));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   542
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   543
489c9b5090e2 Initial load
duke
parents:
diff changeset
   544
    // put 0..7 in i's, 8..15 in l's, 16..23 in o's, 24..31 in g's
489c9b5090e2 Initial load
duke
parents:
diff changeset
   545
    for ( i = 0; i < 8; ++i ) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   546
      if (i < 6) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   547
        __ set(     i, as_iRegister(i));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   548
        __ set(16 + i, as_oRegister(i));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   549
        __ set(24 + i, as_gRegister(i));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   550
      }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   551
      __ set( 8 + i, as_lRegister(i));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   552
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   553
489c9b5090e2 Initial load
duke
parents:
diff changeset
   554
    __ stop("testing stop");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   555
489c9b5090e2 Initial load
duke
parents:
diff changeset
   556
489c9b5090e2 Initial load
duke
parents:
diff changeset
   557
    __ ret();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   558
    __ delayed()->restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   559
489c9b5090e2 Initial load
duke
parents:
diff changeset
   560
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   561
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   562
489c9b5090e2 Initial load
duke
parents:
diff changeset
   563
489c9b5090e2 Initial load
duke
parents:
diff changeset
   564
  address generate_stop_subroutine() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   565
    StubCodeMark mark(this, "StubRoutines", "stop_subroutine");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   566
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   567
489c9b5090e2 Initial load
duke
parents:
diff changeset
   568
    __ stop_subroutine();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   569
489c9b5090e2 Initial load
duke
parents:
diff changeset
   570
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   571
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   572
489c9b5090e2 Initial load
duke
parents:
diff changeset
   573
  address generate_flush_callers_register_windows() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   574
    StubCodeMark mark(this, "StubRoutines", "flush_callers_register_windows");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   575
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   576
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   577
    __ flushw();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   578
    __ retl(false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   579
    __ delayed()->add( FP, STACK_BIAS, O0 );
489c9b5090e2 Initial load
duke
parents:
diff changeset
   580
    // The returned value must be a stack pointer whose register save area
489c9b5090e2 Initial load
duke
parents:
diff changeset
   581
    // is flushed, and will stay flushed while the caller executes.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   582
489c9b5090e2 Initial load
duke
parents:
diff changeset
   583
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   584
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   585
489c9b5090e2 Initial load
duke
parents:
diff changeset
   586
  // Support for jint Atomic::xchg(jint exchange_value, volatile jint* dest).
489c9b5090e2 Initial load
duke
parents:
diff changeset
   587
  //
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   588
  // Arguments:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   589
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   590
  //      exchange_value: O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   591
  //      dest:           O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   592
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   593
  // Results:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   594
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   595
  //     O0: the value previously stored in dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   596
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   597
  address generate_atomic_xchg() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   598
    StubCodeMark mark(this, "StubRoutines", "atomic_xchg");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   599
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   600
489c9b5090e2 Initial load
duke
parents:
diff changeset
   601
    if (UseCASForSwap) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   602
      // Use CAS instead of swap, just in case the MP hardware
489c9b5090e2 Initial load
duke
parents:
diff changeset
   603
      // prefers to work with just one kind of synch. instruction.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   604
      Label retry;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   605
      __ BIND(retry);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   606
      __ mov(O0, O3);       // scratch copy of exchange value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   607
      __ ld(O1, 0, O2);     // observe the previous value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   608
      // try to replace O2 with O3
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   609
      __ cas(O1, O2, O3);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
   610
      __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pn, retry);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   611
489c9b5090e2 Initial load
duke
parents:
diff changeset
   612
      __ retl(false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   613
      __ delayed()->mov(O2, O0);  // report previous value to caller
489c9b5090e2 Initial load
duke
parents:
diff changeset
   614
    } else {
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   615
      __ retl(false);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   616
      __ delayed()->swap(O1, 0, O0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   617
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   618
489c9b5090e2 Initial load
duke
parents:
diff changeset
   619
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   620
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   621
489c9b5090e2 Initial load
duke
parents:
diff changeset
   622
489c9b5090e2 Initial load
duke
parents:
diff changeset
   623
  // Support for jint Atomic::cmpxchg(jint exchange_value, volatile jint* dest, jint compare_value)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   624
  //
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   625
  // Arguments:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   626
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   627
  //      exchange_value: O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   628
  //      dest:           O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   629
  //      compare_value:  O2
489c9b5090e2 Initial load
duke
parents:
diff changeset
   630
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   631
  // Results:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   632
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   633
  //     O0: the value previously stored in dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   634
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   635
  address generate_atomic_cmpxchg() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   636
    StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   637
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   638
489c9b5090e2 Initial load
duke
parents:
diff changeset
   639
    // cmpxchg(dest, compare_value, exchange_value)
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   640
    __ cas(O1, O2, O0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   641
    __ retl(false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   642
    __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   643
489c9b5090e2 Initial load
duke
parents:
diff changeset
   644
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   645
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   646
489c9b5090e2 Initial load
duke
parents:
diff changeset
   647
  // Support for jlong Atomic::cmpxchg(jlong exchange_value, volatile jlong *dest, jlong compare_value)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   648
  //
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   649
  // Arguments:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   650
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   651
  //      exchange_value: O1:O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   652
  //      dest:           O2
489c9b5090e2 Initial load
duke
parents:
diff changeset
   653
  //      compare_value:  O4:O3
489c9b5090e2 Initial load
duke
parents:
diff changeset
   654
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   655
  // Results:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   656
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   657
  //     O1:O0: the value previously stored in dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   658
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   659
  // Overwrites: G1,G2,G3
489c9b5090e2 Initial load
duke
parents:
diff changeset
   660
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   661
  address generate_atomic_cmpxchg_long() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   662
    StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   663
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   664
489c9b5090e2 Initial load
duke
parents:
diff changeset
   665
    __ sllx(O0, 32, O0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   666
    __ srl(O1, 0, O1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   667
    __ or3(O0,O1,O0);      // O0 holds 64-bit value from compare_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   668
    __ sllx(O3, 32, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   669
    __ srl(O4, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   670
    __ or3(O3,O4,O3);     // O3 holds 64-bit value from exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   671
    __ casx(O2, O3, O0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   672
    __ srl(O0, 0, O1);    // unpacked return value in O1:O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   673
    __ retl(false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   674
    __ delayed()->srlx(O0, 32, O0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   675
489c9b5090e2 Initial load
duke
parents:
diff changeset
   676
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   677
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   678
489c9b5090e2 Initial load
duke
parents:
diff changeset
   679
489c9b5090e2 Initial load
duke
parents:
diff changeset
   680
  // Support for jint Atomic::add(jint add_value, volatile jint* dest).
489c9b5090e2 Initial load
duke
parents:
diff changeset
   681
  //
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   682
  // Arguments:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   683
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   684
  //      add_value: O0   (e.g., +1 or -1)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   685
  //      dest:      O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   686
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   687
  // Results:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   688
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   689
  //     O0: the new value stored in dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   690
  //
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   691
  // Overwrites: O3
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   692
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   693
  address generate_atomic_add() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   694
    StubCodeMark mark(this, "StubRoutines", "atomic_add");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   695
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   696
    __ BIND(_atomic_add_stub);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   697
18097
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   698
    Label(retry);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   699
    __ BIND(retry);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   700
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   701
    __ lduw(O1, 0, O2);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   702
    __ add(O0, O2, O3);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   703
    __ cas(O1, O2, O3);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   704
    __ cmp_and_br_short(O2, O3, Assembler::notEqual, Assembler::pn, retry);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   705
    __ retl(false);
acd70736bd60 8008407: remove SPARC V8 support
morris
parents: 14633
diff changeset
   706
    __ delayed()->add(O0, O2, O0); // note that cas made O2==O3
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   707
489c9b5090e2 Initial load
duke
parents:
diff changeset
   708
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   709
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   710
  Label _atomic_add_stub;  // called from other stubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
   711
489c9b5090e2 Initial load
duke
parents:
diff changeset
   712
489c9b5090e2 Initial load
duke
parents:
diff changeset
   713
  // Support for uint StubRoutine::Sparc::partial_subtype_check( Klass sub, Klass super );
489c9b5090e2 Initial load
duke
parents:
diff changeset
   714
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   715
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   716
  //      ret  : O0, returned
489c9b5090e2 Initial load
duke
parents:
diff changeset
   717
  //      icc/xcc: set as O0 (depending on wordSize)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   718
  //      sub  : O1, argument, not changed
489c9b5090e2 Initial load
duke
parents:
diff changeset
   719
  //      super: O2, argument, not changed
489c9b5090e2 Initial load
duke
parents:
diff changeset
   720
  //      raddr: O7, blown by call
489c9b5090e2 Initial load
duke
parents:
diff changeset
   721
  address generate_partial_subtype_check() {
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
   722
    __ align(CodeEntryAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   723
    StubCodeMark mark(this, "StubRoutines", "partial_subtype_check");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   724
    address start = __ pc();
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   725
    Label miss;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   726
489c9b5090e2 Initial load
duke
parents:
diff changeset
   727
    __ save_frame(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   728
    Register Rret   = I0;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   729
    Register Rsub   = I1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   730
    Register Rsuper = I2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   731
489c9b5090e2 Initial load
duke
parents:
diff changeset
   732
    Register L0_ary_len = L0;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   733
    Register L1_ary_ptr = L1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   734
    Register L2_super   = L2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   735
    Register L3_index   = L3;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   736
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   737
    __ check_klass_subtype_slow_path(Rsub, Rsuper,
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   738
                                     L0, L1, L2, L3,
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   739
                                     NULL, &miss);
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   740
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   741
    // Match falls through here.
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
   742
    __ addcc(G0,0,Rret);        // set Z flags, Z result
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   743
489c9b5090e2 Initial load
duke
parents:
diff changeset
   744
    __ ret();                   // Result in Rret is zero; flags set to Z
489c9b5090e2 Initial load
duke
parents:
diff changeset
   745
    __ delayed()->restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   746
489c9b5090e2 Initial load
duke
parents:
diff changeset
   747
    __ BIND(miss);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   748
    __ addcc(G0,1,Rret);        // set NZ flags, NZ result
489c9b5090e2 Initial load
duke
parents:
diff changeset
   749
489c9b5090e2 Initial load
duke
parents:
diff changeset
   750
    __ ret();                   // Result in Rret is != 0; flags set to NZ
489c9b5090e2 Initial load
duke
parents:
diff changeset
   751
    __ delayed()->restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   752
489c9b5090e2 Initial load
duke
parents:
diff changeset
   753
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   754
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   755
489c9b5090e2 Initial load
duke
parents:
diff changeset
   756
489c9b5090e2 Initial load
duke
parents:
diff changeset
   757
  // Called from MacroAssembler::verify_oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
   758
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   759
  address generate_verify_oop_subroutine() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   760
    StubCodeMark mark(this, "StubRoutines", "verify_oop_stub");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   761
489c9b5090e2 Initial load
duke
parents:
diff changeset
   762
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   763
489c9b5090e2 Initial load
duke
parents:
diff changeset
   764
    __ verify_oop_subroutine();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   765
489c9b5090e2 Initial load
duke
parents:
diff changeset
   766
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   767
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   768
489c9b5090e2 Initial load
duke
parents:
diff changeset
   769
489c9b5090e2 Initial load
duke
parents:
diff changeset
   770
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   771
  // Verify that a register contains clean 32-bits positive value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   772
  // (high 32-bits are 0) so it could be used in 64-bits shifts (sllx, srax).
489c9b5090e2 Initial load
duke
parents:
diff changeset
   773
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   774
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   775
  //    Rint  -  32-bits value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   776
  //    Rtmp  -  scratch
489c9b5090e2 Initial load
duke
parents:
diff changeset
   777
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   778
  void assert_clean_int(Register Rint, Register Rtmp) {
46381
020219e46c86 8150388: Remove SPARC 32-bit support
gtriantafill
parents: 38246
diff changeset
   779
  #if defined(ASSERT)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   780
    __ signx(Rint, Rtmp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   781
    __ cmp(Rint, Rtmp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   782
    __ breakpoint_trap(Assembler::notEqual, Assembler::xcc);
46381
020219e46c86 8150388: Remove SPARC 32-bit support
gtriantafill
parents: 38246
diff changeset
   783
  #endif
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   784
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   785
489c9b5090e2 Initial load
duke
parents:
diff changeset
   786
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   787
  //  Generate overlap test for array copy stubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
   788
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   789
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   790
  //    O0    -  array1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   791
  //    O1    -  array2
489c9b5090e2 Initial load
duke
parents:
diff changeset
   792
  //    O2    -  element count
489c9b5090e2 Initial load
duke
parents:
diff changeset
   793
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   794
  //  Kills temps:  O3, O4
489c9b5090e2 Initial load
duke
parents:
diff changeset
   795
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   796
  void array_overlap_test(address no_overlap_target, int log2_elem_size) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   797
    assert(no_overlap_target != NULL, "must be generated");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   798
    array_overlap_test(no_overlap_target, NULL, log2_elem_size);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   799
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   800
  void array_overlap_test(Label& L_no_overlap, int log2_elem_size) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   801
    array_overlap_test(NULL, &L_no_overlap, log2_elem_size);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   802
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   803
  void array_overlap_test(address no_overlap_target, Label* NOLp, int log2_elem_size) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   804
    const Register from       = O0;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   805
    const Register to         = O1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   806
    const Register count      = O2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   807
    const Register to_from    = O3; // to - from
489c9b5090e2 Initial load
duke
parents:
diff changeset
   808
    const Register byte_count = O4; // count << log2_elem_size
489c9b5090e2 Initial load
duke
parents:
diff changeset
   809
489c9b5090e2 Initial load
duke
parents:
diff changeset
   810
      __ subcc(to, from, to_from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   811
      __ sll_ptr(count, log2_elem_size, byte_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   812
      if (NOLp == NULL)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   813
        __ brx(Assembler::lessEqualUnsigned, false, Assembler::pt, no_overlap_target);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   814
      else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   815
        __ brx(Assembler::lessEqualUnsigned, false, Assembler::pt, (*NOLp));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   816
      __ delayed()->cmp(to_from, byte_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   817
      if (NOLp == NULL)
6057
e660446e0804 6962569: assembler_sparc.cpp:1969: assert(false) failed: error
tonyp
parents: 5702
diff changeset
   818
        __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, no_overlap_target);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   819
      else
6057
e660446e0804 6962569: assembler_sparc.cpp:1969: assert(false) failed: error
tonyp
parents: 5702
diff changeset
   820
        __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, (*NOLp));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   821
      __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   822
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   823
489c9b5090e2 Initial load
duke
parents:
diff changeset
   824
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   825
  //  Generate pre-write barrier for array.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   826
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   827
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   828
  //     addr     - register containing starting address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   829
  //     count    - register containing element count
489c9b5090e2 Initial load
duke
parents:
diff changeset
   830
  //     tmp      - scratch register
489c9b5090e2 Initial load
duke
parents:
diff changeset
   831
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   832
  //  The input registers are overwritten.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   833
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   834
  void gen_write_ref_array_pre_barrier(Register addr, Register count, bool dest_uninitialized) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   835
    BarrierSet* bs = Universe::heap()->barrier_set();
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   836
    switch (bs->kind()) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   837
      case BarrierSet::G1SATBCTLogging:
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   838
        // With G1, don't generate the call if we statically know that the target in uninitialized
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   839
        if (!dest_uninitialized) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   840
          __ save_frame(0);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   841
          // Save the necessary global regs... will be used after.
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   842
          if (addr->is_global()) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   843
            __ mov(addr, L0);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   844
          }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   845
          if (count->is_global()) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   846
            __ mov(count, L1);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   847
          }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   848
          __ mov(addr->after_save(), O0);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   849
          // Get the count into O1
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   850
          __ call(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre));
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   851
          __ delayed()->mov(count->after_save(), O1);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   852
          if (addr->is_global()) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   853
            __ mov(L0, addr);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   854
          }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   855
          if (count->is_global()) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   856
            __ mov(L1, count);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   857
          }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   858
          __ restore();
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   859
        }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   860
        break;
32596
8feecdee3156 8072817: CardTableExtension kind() should be BarrierSet::CardTableExtension
kbarrett
parents: 31515
diff changeset
   861
      case BarrierSet::CardTableForRS:
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   862
      case BarrierSet::CardTableExtension:
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   863
      case BarrierSet::ModRef:
55767
8e22715afabc Initial import of Epsilon sources from jdk10/sandbox
shade
parents: 47561
diff changeset
   864
      case BarrierSet::Epsilon:
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   865
        break;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   866
      default:
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   867
        ShouldNotReachHere();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   868
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   869
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   870
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   871
  //  Generate post-write barrier for array.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   872
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   873
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   874
  //     addr     - register containing starting address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   875
  //     count    - register containing element count
489c9b5090e2 Initial load
duke
parents:
diff changeset
   876
  //     tmp      - scratch register
489c9b5090e2 Initial load
duke
parents:
diff changeset
   877
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   878
  //  The input registers are overwritten.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   879
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   880
  void gen_write_ref_array_post_barrier(Register addr, Register count,
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
   881
                                        Register tmp) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   882
    BarrierSet* bs = Universe::heap()->barrier_set();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   883
489c9b5090e2 Initial load
duke
parents:
diff changeset
   884
    switch (bs->kind()) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   885
      case BarrierSet::G1SATBCTLogging:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   886
        {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   887
          // Get some new fresh output registers.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   888
          __ save_frame(0);
1374
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 360
diff changeset
   889
          __ mov(addr->after_save(), O0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   890
          __ call(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post));
1374
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 360
diff changeset
   891
          __ delayed()->mov(count->after_save(), O1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   892
          __ restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   893
        }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   894
        break;
32596
8feecdee3156 8072817: CardTableExtension kind() should be BarrierSet::CardTableExtension
kbarrett
parents: 31515
diff changeset
   895
      case BarrierSet::CardTableForRS:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   896
      case BarrierSet::CardTableExtension:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   897
        {
29325
0e86e64c66e5 8069016: Add BarrierSet downcast support
kbarrett
parents: 27691
diff changeset
   898
          CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   899
          assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   900
          assert_different_registers(addr, count, tmp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   901
48104
62d5973082e3 8185591: guarantee(_byte_map[_guard_index] == last_card) failed: card table guard has been modified
aharlap
parents: 47561
diff changeset
   902
          Label L_loop, L_done;
62d5973082e3 8185591: guarantee(_byte_map[_guard_index] == last_card) failed: card table guard has been modified
aharlap
parents: 47561
diff changeset
   903
62d5973082e3 8185591: guarantee(_byte_map[_guard_index] == last_card) failed: card table guard has been modified
aharlap
parents: 47561
diff changeset
   904
          __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_done); // zero count - nothing to do
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   905
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
   906
          __ sll_ptr(count, LogBytesPerHeapOop, count);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
   907
          __ sub(count, BytesPerHeapOop, count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   908
          __ add(count, addr, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   909
          // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   910
          __ srl_ptr(addr, CardTableModRefBS::card_shift, addr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   911
          __ srl_ptr(count, CardTableModRefBS::card_shift, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   912
          __ sub(count, addr, count);
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   913
          AddressLiteral rs(ct->byte_map_base);
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   914
          __ set(rs, tmp);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   915
        __ BIND(L_loop);
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   916
          __ stb(G0, tmp, addr);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   917
          __ subcc(count, 1, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   918
          __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   919
          __ delayed()->add(addr, 1, addr);
48104
62d5973082e3 8185591: guarantee(_byte_map[_guard_index] == last_card) failed: card table guard has been modified
aharlap
parents: 47561
diff changeset
   920
        __ BIND(L_done);
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   921
        }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   922
        break;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   923
      case BarrierSet::ModRef:
55767
8e22715afabc Initial import of Epsilon sources from jdk10/sandbox
shade
parents: 47561
diff changeset
   924
      case BarrierSet::Epsilon:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   925
        break;
2571
d602ad6538bd 6822110: Add AddressLiteral class on SPARC
twisti
parents: 2338
diff changeset
   926
      default:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   927
        ShouldNotReachHere();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   928
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   929
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   930
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   931
  //
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   932
  // Generate main code for disjoint arraycopy
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   933
  //
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   934
  typedef void (StubGenerator::*CopyLoopFunc)(Register from, Register to, Register count, int count_dec,
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   935
                                              Label& L_loop, bool use_prefetch, bool use_bis);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   936
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   937
  void disjoint_copy_core(Register from, Register to, Register count, int log2_elem_size,
24326
d3fdd5c16fe0 8022070: Compilation error in stubGenerator_sparc.cpp with some compilers
mikael
parents: 22505
diff changeset
   938
                          int iter_size, StubGenerator::CopyLoopFunc copy_loop_func) {
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   939
    Label L_copy;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   940
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   941
    assert(log2_elem_size <= 3, "the following code should be changed");
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   942
    int count_dec = 16>>log2_elem_size;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   943
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   944
    int prefetch_dist = MAX2(ArraycopySrcPrefetchDistance, ArraycopyDstPrefetchDistance);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   945
    assert(prefetch_dist < 4096, "invalid value");
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   946
    prefetch_dist = (prefetch_dist + (iter_size-1)) & (-iter_size); // round up to one iteration copy size
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   947
    int prefetch_count = (prefetch_dist >> log2_elem_size); // elements count
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   948
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   949
    if (UseBlockCopy) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   950
      Label L_block_copy, L_block_copy_prefetch, L_skip_block_copy;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   951
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   952
      // 64 bytes tail + bytes copied in one loop iteration
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   953
      int tail_size = 64 + iter_size;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   954
      int block_copy_count = (MAX2(tail_size, (int)BlockCopyLowLimit)) >> log2_elem_size;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   955
      // Use BIS copy only for big arrays since it requires membar.
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   956
      __ set(block_copy_count, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   957
      __ cmp_and_br_short(count, O4, Assembler::lessUnsigned, Assembler::pt, L_skip_block_copy);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   958
      // This code is for disjoint source and destination:
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   959
      //   to <= from || to >= from+count
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   960
      // but BIS will stomp over 'from' if (to > from-tail_size && to <= from)
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   961
      __ sub(from, to, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   962
      __ srax(O4, 4, O4); // divide by 16 since following short branch have only 5 bits for imm.
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   963
      __ cmp_and_br_short(O4, (tail_size>>4), Assembler::lessEqualUnsigned, Assembler::pn, L_skip_block_copy);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   964
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   965
      __ wrasi(G0, Assembler::ASI_ST_BLKINIT_PRIMARY);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   966
      // BIS should not be used to copy tail (64 bytes+iter_size)
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   967
      // to avoid zeroing of following values.
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   968
      __ sub(count, (tail_size>>log2_elem_size), count); // count is still positive >= 0
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   969
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   970
      if (prefetch_count > 0) { // rounded up to one iteration count
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   971
        // Do prefetching only if copy size is bigger
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   972
        // than prefetch distance.
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   973
        __ set(prefetch_count, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   974
        __ cmp_and_brx_short(count, O4, Assembler::less, Assembler::pt, L_block_copy);
46422
6ecc7cc67619 8179019: Correct range checks for command-line options ArraycopySrcPrefetchDistance and ArraycopyDstPrefetchDistance
zmajo
parents: 46381
diff changeset
   975
        __ sub(count, O4, count);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   976
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   977
        (this->*copy_loop_func)(from, to, count, count_dec, L_block_copy_prefetch, true, true);
46422
6ecc7cc67619 8179019: Correct range checks for command-line options ArraycopySrcPrefetchDistance and ArraycopyDstPrefetchDistance
zmajo
parents: 46381
diff changeset
   978
        __ set(prefetch_count, O4);
6ecc7cc67619 8179019: Correct range checks for command-line options ArraycopySrcPrefetchDistance and ArraycopyDstPrefetchDistance
zmajo
parents: 46381
diff changeset
   979
        __ add(count, O4, count);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   980
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   981
      } // prefetch_count > 0
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   982
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   983
      (this->*copy_loop_func)(from, to, count, count_dec, L_block_copy, false, true);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   984
      __ add(count, (tail_size>>log2_elem_size), count); // restore count
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   985
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   986
      __ wrasi(G0, Assembler::ASI_PRIMARY_NOFAULT);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   987
      // BIS needs membar.
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   988
      __ membar(Assembler::StoreLoad);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   989
      // Copy tail
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   990
      __ ba_short(L_copy);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   991
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   992
      __ BIND(L_skip_block_copy);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   993
    } // UseBlockCopy
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   994
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   995
    if (prefetch_count > 0) { // rounded up to one iteration count
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   996
      // Do prefetching only if copy size is bigger
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   997
      // than prefetch distance.
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   998
      __ set(prefetch_count, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
   999
      __ cmp_and_brx_short(count, O4, Assembler::lessUnsigned, Assembler::pt, L_copy);
46422
6ecc7cc67619 8179019: Correct range checks for command-line options ArraycopySrcPrefetchDistance and ArraycopyDstPrefetchDistance
zmajo
parents: 46381
diff changeset
  1000
      __ sub(count, O4, count);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1001
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1002
      Label L_copy_prefetch;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1003
      (this->*copy_loop_func)(from, to, count, count_dec, L_copy_prefetch, true, false);
46422
6ecc7cc67619 8179019: Correct range checks for command-line options ArraycopySrcPrefetchDistance and ArraycopyDstPrefetchDistance
zmajo
parents: 46381
diff changeset
  1004
      __ set(prefetch_count, O4);
6ecc7cc67619 8179019: Correct range checks for command-line options ArraycopySrcPrefetchDistance and ArraycopyDstPrefetchDistance
zmajo
parents: 46381
diff changeset
  1005
      __ add(count, O4, count);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1006
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1007
    } // prefetch_count > 0
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1008
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1009
    (this->*copy_loop_func)(from, to, count, count_dec, L_copy, false, false);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1010
  }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1011
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1012
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1013
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1014
  //
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1015
  // Helper methods for copy_16_bytes_forward_with_shift()
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1016
  //
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1017
  void copy_16_bytes_shift_loop(Register from, Register to, Register count, int count_dec,
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1018
                                Label& L_loop, bool use_prefetch, bool use_bis) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1019
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1020
    const Register left_shift  = G1; // left  shift bit counter
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1021
    const Register right_shift = G5; // right shift bit counter
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1022
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1023
    __ align(OptoLoopAlignment);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1024
    __ BIND(L_loop);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1025
    if (use_prefetch) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1026
      if (ArraycopySrcPrefetchDistance > 0) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1027
        __ prefetch(from, ArraycopySrcPrefetchDistance, Assembler::severalReads);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1028
      }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1029
      if (ArraycopyDstPrefetchDistance > 0) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1030
        __ prefetch(to, ArraycopyDstPrefetchDistance, Assembler::severalWritesAndPossiblyReads);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1031
      }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1032
    }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1033
    __ ldx(from, 0, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1034
    __ ldx(from, 8, G4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1035
    __ inc(to, 16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1036
    __ inc(from, 16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1037
    __ deccc(count, count_dec); // Can we do next iteration after this one?
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1038
    __ srlx(O4, right_shift, G3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1039
    __ bset(G3, O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1040
    __ sllx(O4, left_shift,  O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1041
    __ srlx(G4, right_shift, G3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1042
    __ bset(G3, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1043
    if (use_bis) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1044
      __ stxa(O3, to, -16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1045
      __ stxa(O4, to, -8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1046
    } else {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1047
      __ stx(O3, to, -16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1048
      __ stx(O4, to, -8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1049
    }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1050
    __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1051
    __ delayed()->sllx(G4, left_shift,  O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1052
  }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1053
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1054
  // Copy big chunks forward with shift
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1055
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1056
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1057
  //   from      - source arrays
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1058
  //   to        - destination array aligned to 8-bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1059
  //   count     - elements count to copy >= the count equivalent to 16 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1060
  //   count_dec - elements count's decrement equivalent to 16 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1061
  //   L_copy_bytes - copy exit label
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1062
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1063
  void copy_16_bytes_forward_with_shift(Register from, Register to,
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1064
                     Register count, int log2_elem_size, Label& L_copy_bytes) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1065
    Label L_aligned_copy, L_copy_last_bytes;
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1066
    assert(log2_elem_size <= 3, "the following code should be changed");
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1067
    int count_dec = 16>>log2_elem_size;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1068
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1069
    // if both arrays have the same alignment mod 8, do 8 bytes aligned copy
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1070
    __ andcc(from, 7, G1); // misaligned bytes
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1071
    __ br(Assembler::zero, false, Assembler::pt, L_aligned_copy);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1072
    __ delayed()->nop();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1073
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1074
    const Register left_shift  = G1; // left  shift bit counter
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1075
    const Register right_shift = G5; // right shift bit counter
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1076
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1077
    __ sll(G1, LogBitsPerByte, left_shift);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1078
    __ mov(64, right_shift);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1079
    __ sub(right_shift, left_shift, right_shift);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1080
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1081
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1082
    // Load 2 aligned 8-bytes chunks and use one from previous iteration
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1083
    // to form 2 aligned 8-bytes chunks to store.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1084
    //
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1085
    __ dec(count, count_dec);   // Pre-decrement 'count'
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1086
    __ andn(from, 7, from);     // Align address
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1087
    __ ldx(from, 0, O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1088
    __ inc(from, 8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1089
    __ sllx(O3, left_shift,  O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1090
24326
d3fdd5c16fe0 8022070: Compilation error in stubGenerator_sparc.cpp with some compilers
mikael
parents: 22505
diff changeset
  1091
    disjoint_copy_core(from, to, count, log2_elem_size, 16, &StubGenerator::copy_16_bytes_shift_loop);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1092
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1093
    __ inccc(count, count_dec>>1 ); // + 8 bytes
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1094
    __ brx(Assembler::negative, true, Assembler::pn, L_copy_last_bytes);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1095
    __ delayed()->inc(count, count_dec>>1); // restore 'count'
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1096
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1097
    // copy 8 bytes, part of them already loaded in O3
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1098
    __ ldx(from, 0, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1099
    __ inc(to, 8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1100
    __ inc(from, 8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1101
    __ srlx(O4, right_shift, G3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1102
    __ bset(O3, G3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1103
    __ stx(G3, to, -8);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1104
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1105
    __ BIND(L_copy_last_bytes);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1106
    __ srl(right_shift, LogBitsPerByte, right_shift); // misaligned bytes
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1107
    __ br(Assembler::always, false, Assembler::pt, L_copy_bytes);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1108
    __ delayed()->sub(from, right_shift, from);       // restore address
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1109
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1110
    __ BIND(L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1111
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1112
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1113
  // Copy big chunks backward with shift
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1114
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1115
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1116
  //   end_from  - source arrays end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1117
  //   end_to    - destination array end address aligned to 8-bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1118
  //   count     - elements count to copy >= the count equivalent to 16 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1119
  //   count_dec - elements count's decrement equivalent to 16 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1120
  //   L_aligned_copy - aligned copy exit label
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1121
  //   L_copy_bytes   - copy exit label
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1122
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1123
  void copy_16_bytes_backward_with_shift(Register end_from, Register end_to,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1124
                     Register count, int count_dec,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1125
                     Label& L_aligned_copy, Label& L_copy_bytes) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1126
    Label L_loop, L_copy_last_bytes;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1127
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1128
    // if both arrays have the same alignment mod 8, do 8 bytes aligned copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1129
      __ andcc(end_from, 7, G1); // misaligned bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1130
      __ br(Assembler::zero, false, Assembler::pt, L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1131
      __ delayed()->deccc(count, count_dec); // Pre-decrement 'count'
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1132
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1133
    const Register left_shift  = G1; // left  shift bit counter
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1134
    const Register right_shift = G5; // right shift bit counter
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1135
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1136
      __ sll(G1, LogBitsPerByte, left_shift);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1137
      __ mov(64, right_shift);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1138
      __ sub(right_shift, left_shift, right_shift);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1139
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1140
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1141
    // Load 2 aligned 8-bytes chunks and use one from previous iteration
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1142
    // to form 2 aligned 8-bytes chunks to store.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1143
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1144
      __ andn(end_from, 7, end_from);     // Align address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1145
      __ ldx(end_from, 0, O3);
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  1146
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1147
    __ BIND(L_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1148
      __ ldx(end_from, -8, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1149
      __ deccc(count, count_dec); // Can we do next iteration after this one?
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1150
      __ ldx(end_from, -16, G4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1151
      __ dec(end_to, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1152
      __ dec(end_from, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1153
      __ srlx(O3, right_shift, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1154
      __ sllx(O4, left_shift,  G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1155
      __ bset(G3, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1156
      __ stx(O3, end_to, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1157
      __ srlx(O4, right_shift, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1158
      __ sllx(G4, left_shift,  G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1159
      __ bset(G3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1160
      __ stx(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1161
      __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1162
      __ delayed()->mov(G4, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1163
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1164
      __ inccc(count, count_dec>>1 ); // + 8 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1165
      __ brx(Assembler::negative, true, Assembler::pn, L_copy_last_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1166
      __ delayed()->inc(count, count_dec>>1); // restore 'count'
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1167
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1168
      // copy 8 bytes, part of them already loaded in O3
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1169
      __ ldx(end_from, -8, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1170
      __ dec(end_to, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1171
      __ dec(end_from, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1172
      __ srlx(O3, right_shift, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1173
      __ sllx(O4, left_shift,  G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1174
      __ bset(O3, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1175
      __ stx(G3, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1176
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1177
    __ BIND(L_copy_last_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1178
      __ srl(left_shift, LogBitsPerByte, left_shift);    // misaligned bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1179
      __ br(Assembler::always, false, Assembler::pt, L_copy_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1180
      __ delayed()->add(end_from, left_shift, end_from); // restore address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1181
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1182
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1183
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1184
  //  Generate stub for disjoint byte copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1185
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1186
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1187
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1188
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1189
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1190
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1191
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1192
  address generate_disjoint_byte_copy(bool aligned, address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1193
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1194
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1195
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1196
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1197
    Label L_skip_alignment, L_align;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1198
    Label L_copy_byte, L_copy_byte_loop, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1199
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1200
    const Register from      = O0;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1201
    const Register to        = O1;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1202
    const Register count     = O2;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1203
    const Register offset    = O5;   // offset from start of arrays
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1204
    // O3, O4, G3, G4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1205
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1206
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1207
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1208
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1209
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1210
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1211
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1212
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1213
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1214
    // for short arrays, just do single element copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1215
    __ cmp(count, 23); // 16 + 7
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1216
    __ brx(Assembler::less, false, Assembler::pn, L_copy_byte);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1217
    __ delayed()->mov(G0, offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1218
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1219
    if (aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1220
      // 'aligned' == true when it is known statically during compilation
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1221
      // of this arraycopy call site that both 'from' and 'to' addresses
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1222
      // are HeapWordSize aligned (see LibraryCallKit::basictype2arraycopy()).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1223
      //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1224
      // Aligned arrays have 4 bytes alignment in 32-bits VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1225
      // and 8 bytes - in 64-bits VM. So we do it only for 32-bits VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1226
      //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1227
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1228
      // copy bytes to align 'to' on 8 byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1229
      __ andcc(to, 7, G1); // misaligned bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1230
      __ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1231
      __ delayed()->neg(G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1232
      __ inc(G1, 8);       // bytes need to copy to next 8-bytes alignment
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1233
      __ sub(count, G1, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1234
    __ BIND(L_align);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1235
      __ ldub(from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1236
      __ deccc(G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1237
      __ inc(from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1238
      __ stb(O3, to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1239
      __ br(Assembler::notZero, false, Assembler::pt, L_align);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1240
      __ delayed()->inc(to);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1241
    __ BIND(L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1242
    }
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1243
    if (!aligned) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1244
      // Copy with shift 16 bytes per iteration if arrays do not have
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1245
      // the same alignment mod 8, otherwise fall through to the next
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1246
      // code for aligned copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1247
      // The compare above (count >= 23) guarantes 'count' >= 16 bytes.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1248
      // Also jump over aligned copy after the copy with shift completed.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1249
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1250
      copy_16_bytes_forward_with_shift(from, to, count, 0, L_copy_byte);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1251
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1252
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1253
    // Both array are 8 bytes aligned, copy 16 bytes at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1254
      __ and3(count, 7, G4); // Save count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1255
      __ srl(count, 3, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1256
     generate_disjoint_long_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1257
      __ mov(G4, count);     // Restore count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1258
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1259
    // copy tailing bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1260
    __ BIND(L_copy_byte);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  1261
      __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  1262
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1263
    __ BIND(L_copy_byte_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1264
      __ ldub(from, offset, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1265
      __ deccc(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1266
      __ stb(O3, to, offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1267
      __ brx(Assembler::notZero, false, Assembler::pt, L_copy_byte_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1268
      __ delayed()->inc(offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1269
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1270
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1271
      // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1272
      inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1273
      __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1274
      __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1275
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1276
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1277
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1278
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1279
  //  Generate stub for conjoint byte copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1280
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1281
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1282
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1283
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1284
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1285
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1286
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1287
  address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1288
                                      address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1289
    // Do reverse copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1290
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1291
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1292
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1293
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1294
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1295
    Label L_skip_alignment, L_align, L_aligned_copy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1296
    Label L_copy_byte, L_copy_byte_loop, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1297
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1298
    const Register from      = O0;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1299
    const Register to        = O1;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1300
    const Register count     = O2;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1301
    const Register end_from  = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1302
    const Register end_to    = to;   // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1303
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1304
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1305
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1306
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1307
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1308
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1309
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1310
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1311
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1312
    array_overlap_test(nooverlap_target, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1313
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1314
    __ add(to, count, end_to);       // offset after last copied element
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1315
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1316
    // for short arrays, just do single element copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1317
    __ cmp(count, 23); // 16 + 7
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1318
    __ brx(Assembler::less, false, Assembler::pn, L_copy_byte);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1319
    __ delayed()->add(from, count, end_from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1320
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1321
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1322
      // Align end of arrays since they could be not aligned even
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1323
      // when arrays itself are aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1324
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1325
      // copy bytes to align 'end_to' on 8 byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1326
      __ andcc(end_to, 7, G1); // misaligned bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1327
      __ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1328
      __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1329
      __ sub(count, G1, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1330
    __ BIND(L_align);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1331
      __ dec(end_from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1332
      __ dec(end_to);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1333
      __ ldub(end_from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1334
      __ deccc(G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1335
      __ brx(Assembler::notZero, false, Assembler::pt, L_align);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1336
      __ delayed()->stb(O3, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1337
    __ BIND(L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1338
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1339
    if (aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1340
      // Both arrays are aligned to 8-bytes in 64-bits VM.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1341
      // The 'count' is decremented in copy_16_bytes_backward_with_shift()
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1342
      // in unaligned case.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1343
      __ dec(count, 16);
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1344
    } else {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1345
      // Copy with shift 16 bytes per iteration if arrays do not have
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1346
      // the same alignment mod 8, otherwise jump to the next
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1347
      // code for aligned copy (and substracting 16 from 'count' before jump).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1348
      // The compare above (count >= 11) guarantes 'count' >= 16 bytes.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1349
      // Also jump over aligned copy after the copy with shift completed.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1350
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1351
      copy_16_bytes_backward_with_shift(end_from, end_to, count, 16,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1352
                                        L_aligned_copy, L_copy_byte);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1353
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1354
    // copy 4 elements (16 bytes) at a time
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  1355
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1356
    __ BIND(L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1357
      __ dec(end_from, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1358
      __ ldx(end_from, 8, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1359
      __ ldx(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1360
      __ dec(end_to, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1361
      __ deccc(count, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1362
      __ stx(O3, end_to, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1363
      __ brx(Assembler::greaterEqual, false, Assembler::pt, L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1364
      __ delayed()->stx(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1365
      __ inc(count, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1366
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1367
    // copy 1 element (2 bytes) at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1368
    __ BIND(L_copy_byte);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  1369
      __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  1370
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1371
    __ BIND(L_copy_byte_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1372
      __ dec(end_from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1373
      __ dec(end_to);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1374
      __ ldub(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1375
      __ deccc(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1376
      __ brx(Assembler::greater, false, Assembler::pt, L_copy_byte_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1377
      __ delayed()->stb(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1378
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1379
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1380
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1381
    inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1382
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1383
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1384
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1385
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1386
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1387
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1388
  //  Generate stub for disjoint short copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1389
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1390
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1391
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1392
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1393
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1394
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1395
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1396
  address generate_disjoint_short_copy(bool aligned, address *entry, const char * name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1397
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1398
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1399
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1400
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1401
    Label L_skip_alignment, L_skip_alignment2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1402
    Label L_copy_2_bytes, L_copy_2_bytes_loop, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1403
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1404
    const Register from      = O0;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1405
    const Register to        = O1;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1406
    const Register count     = O2;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1407
    const Register offset    = O5;   // offset from start of arrays
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1408
    // O3, O4, G3, G4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1409
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1410
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1411
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1412
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1413
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1414
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1415
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1416
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1417
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1418
    // for short arrays, just do single element copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1419
    __ cmp(count, 11); // 8 + 3  (22 bytes)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1420
    __ brx(Assembler::less, false, Assembler::pn, L_copy_2_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1421
    __ delayed()->mov(G0, offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1422
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1423
    if (aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1424
      // 'aligned' == true when it is known statically during compilation
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1425
      // of this arraycopy call site that both 'from' and 'to' addresses
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1426
      // are HeapWordSize aligned (see LibraryCallKit::basictype2arraycopy()).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1427
      //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1428
      // Aligned arrays have 4 bytes alignment in 32-bits VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1429
      // and 8 bytes - in 64-bits VM.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1430
      //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1431
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1432
      // copy 1 element if necessary to align 'to' on an 4 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1433
      __ andcc(to, 3, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1434
      __ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1435
      __ delayed()->lduh(from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1436
      __ inc(from, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1437
      __ inc(to, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1438
      __ dec(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1439
      __ sth(O3, to, -2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1440
    __ BIND(L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1441
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1442
      // copy 2 elements to align 'to' on an 8 byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1443
      __ andcc(to, 7, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1444
      __ br(Assembler::zero, false, Assembler::pn, L_skip_alignment2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1445
      __ delayed()->lduh(from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1446
      __ dec(count, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1447
      __ lduh(from, 2, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1448
      __ inc(from, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1449
      __ inc(to, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1450
      __ sth(O3, to, -4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1451
      __ sth(O4, to, -2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1452
    __ BIND(L_skip_alignment2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1453
    }
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1454
    if (!aligned) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1455
      // Copy with shift 16 bytes per iteration if arrays do not have
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1456
      // the same alignment mod 8, otherwise fall through to the next
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1457
      // code for aligned copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1458
      // The compare above (count >= 11) guarantes 'count' >= 16 bytes.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1459
      // Also jump over aligned copy after the copy with shift completed.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1460
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1461
      copy_16_bytes_forward_with_shift(from, to, count, 1, L_copy_2_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1462
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1463
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1464
    // Both array are 8 bytes aligned, copy 16 bytes at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1465
      __ and3(count, 3, G4); // Save
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1466
      __ srl(count, 2, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1467
     generate_disjoint_long_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1468
      __ mov(G4, count); // restore
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1469
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1470
    // copy 1 element at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1471
    __ BIND(L_copy_2_bytes);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  1472
      __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  1473
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1474
    __ BIND(L_copy_2_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1475
      __ lduh(from, offset, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1476
      __ deccc(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1477
      __ sth(O3, to, offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1478
      __ brx(Assembler::notZero, false, Assembler::pt, L_copy_2_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1479
      __ delayed()->inc(offset, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1480
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1481
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1482
      // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1483
      inc_counter_np(SharedRuntime::_jshort_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1484
      __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1485
      __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1486
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1487
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1488
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1489
  //
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1490
  //  Generate stub for disjoint short fill.  If "aligned" is true, the
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1491
  //  "to" address is assumed to be heapword aligned.
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1492
  //
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1493
  // Arguments for generated stub:
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1494
  //      to:    O0
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1495
  //      value: O1
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1496
  //      count: O2 treated as signed
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1497
  //
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1498
  address generate_fill(BasicType t, bool aligned, const char* name) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1499
    __ align(CodeEntryAlignment);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1500
    StubCodeMark mark(this, "StubRoutines", name);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1501
    address start = __ pc();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1502
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1503
    const Register to        = O0;   // source array address
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1504
    const Register value     = O1;   // fill value
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1505
    const Register count     = O2;   // elements count
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1506
    // O3 is used as a temp register
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1507
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1508
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1509
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1510
    Label L_exit, L_skip_align1, L_skip_align2, L_fill_byte;
6464
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1511
    Label L_fill_2_bytes, L_fill_elements, L_fill_32_bytes;
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1512
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1513
    int shift = -1;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1514
    switch (t) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1515
       case T_BYTE:
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1516
        shift = 2;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1517
        break;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1518
       case T_SHORT:
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1519
        shift = 1;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1520
        break;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1521
      case T_INT:
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1522
         shift = 0;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1523
        break;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1524
      default: ShouldNotReachHere();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1525
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1526
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1527
    BLOCK_COMMENT("Entry:");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1528
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1529
    if (t == T_BYTE) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1530
      // Zero extend value
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1531
      __ and3(value, 0xff, value);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1532
      __ sllx(value, 8, O3);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1533
      __ or3(value, O3, value);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1534
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1535
    if (t == T_SHORT) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1536
      // Zero extend value
6464
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1537
      __ sllx(value, 48, value);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1538
      __ srlx(value, 48, value);
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1539
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1540
    if (t == T_BYTE || t == T_SHORT) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1541
      __ sllx(value, 16, O3);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1542
      __ or3(value, O3, value);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1543
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1544
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1545
    __ cmp(count, 2<<shift); // Short arrays (< 8 bytes) fill by element
6464
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1546
    __ brx(Assembler::lessUnsigned, false, Assembler::pn, L_fill_elements); // use unsigned cmp
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1547
    __ delayed()->andcc(count, 1, G0);
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1548
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1549
    if (!aligned && (t == T_BYTE || t == T_SHORT)) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1550
      // align source address at 4 bytes address boundary
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1551
      if (t == T_BYTE) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1552
        // One byte misalignment happens only for byte arrays
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1553
        __ andcc(to, 1, G0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1554
        __ br(Assembler::zero, false, Assembler::pt, L_skip_align1);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1555
        __ delayed()->nop();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1556
        __ stb(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1557
        __ inc(to, 1);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1558
        __ dec(count, 1);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1559
        __ BIND(L_skip_align1);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1560
      }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1561
      // Two bytes misalignment happens only for byte and short (char) arrays
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1562
      __ andcc(to, 2, G0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1563
      __ br(Assembler::zero, false, Assembler::pt, L_skip_align2);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1564
      __ delayed()->nop();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1565
      __ sth(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1566
      __ inc(to, 2);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1567
      __ dec(count, 1 << (shift - 1));
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1568
      __ BIND(L_skip_align2);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1569
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1570
    if (!aligned) {
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1571
      // align to 8 bytes, we know we are 4 byte aligned to start
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1572
      __ andcc(to, 7, G0);
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1573
      __ br(Assembler::zero, false, Assembler::pt, L_fill_32_bytes);
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1574
      __ delayed()->nop();
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1575
      __ stw(value, to, 0);
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1576
      __ inc(to, 4);
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1577
      __ dec(count, 1 << shift);
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1578
      __ BIND(L_fill_32_bytes);
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1579
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1580
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1581
    if (t == T_INT) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1582
      // Zero extend value
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1583
      __ srl(value, 0, value);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1584
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1585
    if (t == T_BYTE || t == T_SHORT || t == T_INT) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1586
      __ sllx(value, 32, O3);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1587
      __ or3(value, O3, value);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1588
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1589
6452
cc624b341ab2 6981773: incorrect fill value with OptimizeFill
never
parents: 6433
diff changeset
  1590
    Label L_check_fill_8_bytes;
cc624b341ab2 6981773: incorrect fill value with OptimizeFill
never
parents: 6433
diff changeset
  1591
    // Fill 32-byte chunks
cc624b341ab2 6981773: incorrect fill value with OptimizeFill
never
parents: 6433
diff changeset
  1592
    __ subcc(count, 8 << shift, count);
cc624b341ab2 6981773: incorrect fill value with OptimizeFill
never
parents: 6433
diff changeset
  1593
    __ brx(Assembler::less, false, Assembler::pt, L_check_fill_8_bytes);
cc624b341ab2 6981773: incorrect fill value with OptimizeFill
never
parents: 6433
diff changeset
  1594
    __ delayed()->nop();
cc624b341ab2 6981773: incorrect fill value with OptimizeFill
never
parents: 6433
diff changeset
  1595
6464
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1596
    Label L_fill_32_bytes_loop, L_fill_4_bytes;
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1597
    __ align(16);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1598
    __ BIND(L_fill_32_bytes_loop);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1599
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1600
    __ stx(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1601
    __ stx(value, to, 8);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1602
    __ stx(value, to, 16);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1603
    __ stx(value, to, 24);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1604
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1605
    __ subcc(count, 8 << shift, count);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1606
    __ brx(Assembler::greaterEqual, false, Assembler::pt, L_fill_32_bytes_loop);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1607
    __ delayed()->add(to, 32, to);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1608
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1609
    __ BIND(L_check_fill_8_bytes);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1610
    __ addcc(count, 8 << shift, count);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1611
    __ brx(Assembler::zero, false, Assembler::pn, L_exit);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1612
    __ delayed()->subcc(count, 1 << (shift + 1), count);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1613
    __ brx(Assembler::less, false, Assembler::pn, L_fill_4_bytes);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1614
    __ delayed()->andcc(count, 1<<shift, G0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1615
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1616
    //
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1617
    // length is too short, just fill 8 bytes at a time
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1618
    //
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1619
    Label L_fill_8_bytes_loop;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1620
    __ BIND(L_fill_8_bytes_loop);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1621
    __ stx(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1622
    __ subcc(count, 1 << (shift + 1), count);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1623
    __ brx(Assembler::greaterEqual, false, Assembler::pn, L_fill_8_bytes_loop);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1624
    __ delayed()->add(to, 8, to);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1625
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1626
    // fill trailing 4 bytes
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1627
    __ andcc(count, 1<<shift, G0);  // in delay slot of branches
6464
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1628
    if (t == T_INT) {
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1629
      __ BIND(L_fill_elements);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1630
    }
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1631
    __ BIND(L_fill_4_bytes);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1632
    __ brx(Assembler::zero, false, Assembler::pt, L_fill_2_bytes);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1633
    if (t == T_BYTE || t == T_SHORT) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1634
      __ delayed()->andcc(count, 1<<(shift-1), G0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1635
    } else {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1636
      __ delayed()->nop();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1637
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1638
    __ stw(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1639
    if (t == T_BYTE || t == T_SHORT) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1640
      __ inc(to, 4);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1641
      // fill trailing 2 bytes
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1642
      __ andcc(count, 1<<(shift-1), G0); // in delay slot of branches
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1643
      __ BIND(L_fill_2_bytes);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1644
      __ brx(Assembler::zero, false, Assembler::pt, L_fill_byte);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1645
      __ delayed()->andcc(count, 1, count);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1646
      __ sth(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1647
      if (t == T_BYTE) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1648
        __ inc(to, 2);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1649
        // fill trailing byte
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1650
        __ andcc(count, 1, count);  // in delay slot of branches
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1651
        __ BIND(L_fill_byte);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1652
        __ brx(Assembler::zero, false, Assembler::pt, L_exit);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1653
        __ delayed()->nop();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1654
        __ stb(value, to, 0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1655
      } else {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1656
        __ BIND(L_fill_byte);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1657
      }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1658
    } else {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1659
      __ BIND(L_fill_2_bytes);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1660
    }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1661
    __ BIND(L_exit);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1662
    __ retl();
6464
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1663
    __ delayed()->nop();
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1664
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1665
    // Handle copies less than 8 bytes.  Int is handled elsewhere.
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1666
    if (t == T_BYTE) {
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1667
      __ BIND(L_fill_elements);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1668
      Label L_fill_2, L_fill_4;
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1669
      // in delay slot __ andcc(count, 1, G0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1670
      __ brx(Assembler::zero, false, Assembler::pt, L_fill_2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1671
      __ delayed()->andcc(count, 2, G0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1672
      __ stb(value, to, 0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1673
      __ inc(to, 1);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1674
      __ BIND(L_fill_2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1675
      __ brx(Assembler::zero, false, Assembler::pt, L_fill_4);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1676
      __ delayed()->andcc(count, 4, G0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1677
      __ stb(value, to, 0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1678
      __ stb(value, to, 1);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1679
      __ inc(to, 2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1680
      __ BIND(L_fill_4);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1681
      __ brx(Assembler::zero, false, Assembler::pt, L_exit);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1682
      __ delayed()->nop();
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1683
      __ stb(value, to, 0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1684
      __ stb(value, to, 1);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1685
      __ stb(value, to, 2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1686
      __ retl();
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1687
      __ delayed()->stb(value, to, 3);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1688
    }
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1689
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1690
    if (t == T_SHORT) {
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1691
      Label L_fill_2;
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1692
      __ BIND(L_fill_elements);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1693
      // in delay slot __ andcc(count, 1, G0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1694
      __ brx(Assembler::zero, false, Assembler::pt, L_fill_2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1695
      __ delayed()->andcc(count, 2, G0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1696
      __ sth(value, to, 0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1697
      __ inc(to, 2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1698
      __ BIND(L_fill_2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1699
      __ brx(Assembler::zero, false, Assembler::pt, L_exit);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1700
      __ delayed()->nop();
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1701
      __ sth(value, to, 0);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1702
      __ retl();
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1703
      __ delayed()->sth(value, to, 2);
cd40daf5b832 6982370: SIGBUS in jbyte_fill
never
parents: 6452
diff changeset
  1704
    }
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1705
    return start;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1706
  }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1707
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  1708
  //
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1709
  //  Generate stub for conjoint short copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1710
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1711
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1712
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1713
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1714
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1715
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1716
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1717
  address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1718
                                       address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1719
    // Do reverse copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1720
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1721
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1722
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1723
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1724
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1725
    Label L_skip_alignment, L_skip_alignment2, L_aligned_copy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1726
    Label L_copy_2_bytes, L_copy_2_bytes_loop, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1727
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1728
    const Register from      = O0;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1729
    const Register to        = O1;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1730
    const Register count     = O2;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1731
    const Register end_from  = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1732
    const Register end_to    = to;   // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1733
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1734
    const Register byte_count = O3;  // bytes count to copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1735
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1736
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1737
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1738
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1739
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1740
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1741
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1742
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1743
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1744
    array_overlap_test(nooverlap_target, 1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1745
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1746
    __ sllx(count, LogBytesPerShort, byte_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1747
    __ add(to, byte_count, end_to);  // offset after last copied element
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1748
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1749
    // for short arrays, just do single element copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1750
    __ cmp(count, 11); // 8 + 3  (22 bytes)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1751
    __ brx(Assembler::less, false, Assembler::pn, L_copy_2_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1752
    __ delayed()->add(from, byte_count, end_from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1753
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1754
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1755
      // Align end of arrays since they could be not aligned even
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1756
      // when arrays itself are aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1757
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1758
      // copy 1 element if necessary to align 'end_to' on an 4 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1759
      __ andcc(end_to, 3, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1760
      __ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1761
      __ delayed()->lduh(end_from, -2, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1762
      __ dec(end_from, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1763
      __ dec(end_to, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1764
      __ dec(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1765
      __ sth(O3, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1766
    __ BIND(L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1767
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1768
      // copy 2 elements to align 'end_to' on an 8 byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1769
      __ andcc(end_to, 7, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1770
      __ br(Assembler::zero, false, Assembler::pn, L_skip_alignment2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1771
      __ delayed()->lduh(end_from, -2, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1772
      __ dec(count, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1773
      __ lduh(end_from, -4, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1774
      __ dec(end_from, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1775
      __ dec(end_to, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1776
      __ sth(O3, end_to, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1777
      __ sth(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1778
    __ BIND(L_skip_alignment2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1779
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1780
    if (aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1781
      // Both arrays are aligned to 8-bytes in 64-bits VM.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1782
      // The 'count' is decremented in copy_16_bytes_backward_with_shift()
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1783
      // in unaligned case.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1784
      __ dec(count, 8);
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1785
    } else {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1786
      // Copy with shift 16 bytes per iteration if arrays do not have
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1787
      // the same alignment mod 8, otherwise jump to the next
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1788
      // code for aligned copy (and substracting 8 from 'count' before jump).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1789
      // The compare above (count >= 11) guarantes 'count' >= 16 bytes.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1790
      // Also jump over aligned copy after the copy with shift completed.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1791
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1792
      copy_16_bytes_backward_with_shift(end_from, end_to, count, 8,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1793
                                        L_aligned_copy, L_copy_2_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1794
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1795
    // copy 4 elements (16 bytes) at a time
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  1796
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1797
    __ BIND(L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1798
      __ dec(end_from, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1799
      __ ldx(end_from, 8, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1800
      __ ldx(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1801
      __ dec(end_to, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1802
      __ deccc(count, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1803
      __ stx(O3, end_to, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1804
      __ brx(Assembler::greaterEqual, false, Assembler::pt, L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1805
      __ delayed()->stx(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1806
      __ inc(count, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1807
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1808
    // copy 1 element (2 bytes) at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1809
    __ BIND(L_copy_2_bytes);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  1810
      __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1811
    __ BIND(L_copy_2_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1812
      __ dec(end_from, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1813
      __ dec(end_to, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1814
      __ lduh(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1815
      __ deccc(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1816
      __ brx(Assembler::greater, false, Assembler::pt, L_copy_2_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1817
      __ delayed()->sth(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1818
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1819
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1820
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1821
    inc_counter_np(SharedRuntime::_jshort_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1822
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1823
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1824
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1825
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1826
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1827
  //
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1828
  // Helper methods for generate_disjoint_int_copy_core()
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1829
  //
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1830
  void copy_16_bytes_loop(Register from, Register to, Register count, int count_dec,
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1831
                          Label& L_loop, bool use_prefetch, bool use_bis) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1832
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1833
    __ align(OptoLoopAlignment);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1834
    __ BIND(L_loop);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1835
    if (use_prefetch) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1836
      if (ArraycopySrcPrefetchDistance > 0) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1837
        __ prefetch(from, ArraycopySrcPrefetchDistance, Assembler::severalReads);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1838
      }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1839
      if (ArraycopyDstPrefetchDistance > 0) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1840
        __ prefetch(to, ArraycopyDstPrefetchDistance, Assembler::severalWritesAndPossiblyReads);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1841
      }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1842
    }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1843
    __ ldx(from, 4, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1844
    __ ldx(from, 12, G4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1845
    __ inc(to, 16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1846
    __ inc(from, 16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1847
    __ deccc(count, 4); // Can we do next iteration after this one?
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1848
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1849
    __ srlx(O4, 32, G3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1850
    __ bset(G3, O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1851
    __ sllx(O4, 32, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1852
    __ srlx(G4, 32, G3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1853
    __ bset(G3, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1854
    if (use_bis) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1855
      __ stxa(O3, to, -16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1856
      __ stxa(O4, to, -8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1857
    } else {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1858
      __ stx(O3, to, -16);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1859
      __ stx(O4, to, -8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1860
    }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1861
    __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1862
    __ delayed()->sllx(G4, 32,  O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1863
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1864
  }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1865
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1866
  //
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1867
  //  Generate core code for disjoint int copy (and oop copy on 32-bit).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1868
  //  If "aligned" is true, the "from" and "to" addresses are assumed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1869
  //  to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1870
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1871
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1872
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1873
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1874
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1875
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1876
  void generate_disjoint_int_copy_core(bool aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1877
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1878
    Label L_skip_alignment, L_aligned_copy;
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1879
    Label L_copy_4_bytes, L_copy_4_bytes_loop, L_exit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1880
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1881
    const Register from      = O0;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1882
    const Register to        = O1;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1883
    const Register count     = O2;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1884
    const Register offset    = O5;   // offset from start of arrays
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1885
    // O3, O4, G3, G4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1886
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1887
    // 'aligned' == true when it is known statically during compilation
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1888
    // of this arraycopy call site that both 'from' and 'to' addresses
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1889
    // are HeapWordSize aligned (see LibraryCallKit::basictype2arraycopy()).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1890
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1891
    // Aligned arrays have 4 bytes alignment in 32-bits VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1892
    // and 8 bytes - in 64-bits VM.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1893
    //
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  1894
    if (!aligned) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1895
      // The next check could be put under 'ifndef' since the code in
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1896
      // generate_disjoint_long_copy_core() has own checks and set 'offset'.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1897
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1898
      // for short arrays, just do single element copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1899
      __ cmp(count, 5); // 4 + 1 (20 bytes)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1900
      __ brx(Assembler::lessEqual, false, Assembler::pn, L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1901
      __ delayed()->mov(G0, offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1902
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1903
      // copy 1 element to align 'to' on an 8 byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1904
      __ andcc(to, 7, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1905
      __ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1906
      __ delayed()->ld(from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1907
      __ inc(from, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1908
      __ inc(to, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1909
      __ dec(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1910
      __ st(O3, to, -4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1911
    __ BIND(L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1912
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1913
    // if arrays have same alignment mod 8, do 4 elements copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1914
      __ andcc(from, 7, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1915
      __ br(Assembler::zero, false, Assembler::pt, L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1916
      __ delayed()->ld(from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1917
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1918
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1919
    // Load 2 aligned 8-bytes chunks and use one from previous iteration
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1920
    // to form 2 aligned 8-bytes chunks to store.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1921
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1922
    // copy_16_bytes_forward_with_shift() is not used here since this
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1923
    // code is more optimal.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1924
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1925
    // copy with shift 4 elements (16 bytes) at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1926
      __ dec(count, 4);   // The cmp at the beginning guaranty count >= 4
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1927
      __ sllx(O3, 32,  O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1928
24326
d3fdd5c16fe0 8022070: Compilation error in stubGenerator_sparc.cpp with some compilers
mikael
parents: 22505
diff changeset
  1929
      disjoint_copy_core(from, to, count, 2, 16, &StubGenerator::copy_16_bytes_loop);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1930
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1931
      __ br(Assembler::always, false, Assembler::pt, L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1932
      __ delayed()->inc(count, 4); // restore 'count'
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1933
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1934
    __ BIND(L_aligned_copy);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1935
    } // !aligned
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  1936
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1937
    // copy 4 elements (16 bytes) at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1938
      __ and3(count, 1, G4); // Save
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1939
      __ srl(count, 1, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1940
     generate_disjoint_long_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1941
      __ mov(G4, count);     // Restore
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1942
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1943
    // copy 1 element at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1944
    __ BIND(L_copy_4_bytes);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  1945
      __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1946
    __ BIND(L_copy_4_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1947
      __ ld(from, offset, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1948
      __ deccc(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1949
      __ st(O3, to, offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1950
      __ brx(Assembler::notZero, false, Assembler::pt, L_copy_4_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1951
      __ delayed()->inc(offset, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1952
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1953
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1954
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1955
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1956
  //  Generate stub for disjoint int copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1957
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1958
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1959
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1960
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1961
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1962
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1963
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1964
  address generate_disjoint_int_copy(bool aligned, address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1965
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1966
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1967
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1968
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1969
    const Register count = O2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1970
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1971
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1972
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1973
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1974
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1975
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  1976
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1977
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1978
    generate_disjoint_int_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1979
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1980
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1981
    inc_counter_np(SharedRuntime::_jint_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1982
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1983
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1984
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1985
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1986
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1987
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1988
  //  Generate core code for conjoint int copy (and oop copy on 32-bit).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1989
  //  If "aligned" is true, the "from" and "to" addresses are assumed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1990
  //  to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1991
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1992
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1993
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1994
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1995
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1996
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1997
  void generate_conjoint_int_copy_core(bool aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1998
    // Do reverse copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1999
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2000
    Label L_skip_alignment, L_aligned_copy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2001
    Label L_copy_16_bytes,  L_copy_4_bytes, L_copy_4_bytes_loop, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2002
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2003
    const Register from      = O0;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2004
    const Register to        = O1;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2005
    const Register count     = O2;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2006
    const Register end_from  = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2007
    const Register end_to    = to;   // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2008
    // O3, O4, O5, G3 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2009
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2010
    const Register byte_count = O3;  // bytes count to copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2011
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2012
      __ sllx(count, LogBytesPerInt, byte_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2013
      __ add(to, byte_count, end_to); // offset after last copied element
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2014
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2015
      __ cmp(count, 5); // for short arrays, just do single element copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2016
      __ brx(Assembler::lessEqual, false, Assembler::pn, L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2017
      __ delayed()->add(from, byte_count, end_from);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2018
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2019
    // copy 1 element to align 'to' on an 8 byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2020
      __ andcc(end_to, 7, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2021
      __ br(Assembler::zero, false, Assembler::pt, L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2022
      __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2023
      __ dec(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2024
      __ dec(end_from, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2025
      __ dec(end_to,   4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2026
      __ ld(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2027
      __ st(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2028
    __ BIND(L_skip_alignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2029
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2030
    // Check if 'end_from' and 'end_to' has the same alignment.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2031
      __ andcc(end_from, 7, G0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2032
      __ br(Assembler::zero, false, Assembler::pt, L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2033
      __ delayed()->dec(count, 4); // The cmp at the start guaranty cnt >= 4
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2034
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2035
    // copy with shift 4 elements (16 bytes) at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2036
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2037
    // Load 2 aligned 8-bytes chunks and use one from previous iteration
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2038
    // to form 2 aligned 8-bytes chunks to store.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2039
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2040
      __ ldx(end_from, -4, O3);
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  2041
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2042
    __ BIND(L_copy_16_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2043
      __ ldx(end_from, -12, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2044
      __ deccc(count, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2045
      __ ldx(end_from, -20, O5);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2046
      __ dec(end_to, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2047
      __ dec(end_from, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2048
      __ srlx(O3, 32, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2049
      __ sllx(O4, 32, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2050
      __ bset(G3, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2051
      __ stx(O3, end_to, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2052
      __ srlx(O4, 32, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2053
      __ sllx(O5, 32, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2054
      __ bset(O4, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2055
      __ stx(G3, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2056
      __ brx(Assembler::greaterEqual, false, Assembler::pt, L_copy_16_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2057
      __ delayed()->mov(O5, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2058
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2059
      __ br(Assembler::always, false, Assembler::pt, L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2060
      __ delayed()->inc(count, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2061
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2062
    // copy 4 elements (16 bytes) at a time
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  2063
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2064
    __ BIND(L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2065
      __ dec(end_from, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2066
      __ ldx(end_from, 8, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2067
      __ ldx(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2068
      __ dec(end_to, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2069
      __ deccc(count, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2070
      __ stx(O3, end_to, 8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2071
      __ brx(Assembler::greaterEqual, false, Assembler::pt, L_aligned_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2072
      __ delayed()->stx(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2073
      __ inc(count, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2074
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2075
    // copy 1 element (4 bytes) at a time
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2076
    __ BIND(L_copy_4_bytes);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  2077
      __ cmp_and_br_short(count, 0, Assembler::equal, Assembler::pt, L_exit);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2078
    __ BIND(L_copy_4_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2079
      __ dec(end_from, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2080
      __ dec(end_to, 4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2081
      __ ld(end_from, 0, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2082
      __ deccc(count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2083
      __ brx(Assembler::greater, false, Assembler::pt, L_copy_4_bytes_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2084
      __ delayed()->st(O4, end_to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2085
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2086
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2087
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2088
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2089
  //  Generate stub for conjoint int copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2090
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2091
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2092
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2093
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2094
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2095
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2096
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2097
  address generate_conjoint_int_copy(bool aligned, address nooverlap_target,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2098
                                     address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2099
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2100
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2101
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2102
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2103
    assert_clean_int(O2, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2104
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2105
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2106
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2107
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2108
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2109
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2110
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2111
    array_overlap_test(nooverlap_target, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2112
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2113
    generate_conjoint_int_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2114
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2115
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2116
    inc_counter_np(SharedRuntime::_jint_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2117
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2118
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2119
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2120
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2121
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2122
  //
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2123
  // Helper methods for generate_disjoint_long_copy_core()
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2124
  //
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2125
  void copy_64_bytes_loop(Register from, Register to, Register count, int count_dec,
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2126
                          Label& L_loop, bool use_prefetch, bool use_bis) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2127
    __ align(OptoLoopAlignment);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2128
    __ BIND(L_loop);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2129
    for (int off = 0; off < 64; off += 16) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2130
      if (use_prefetch && (off & 31) == 0) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2131
        if (ArraycopySrcPrefetchDistance > 0) {
10566
630c177ec580 7081933: Use zeroing elimination optimization for large array
kvn
parents: 10545
diff changeset
  2132
          __ prefetch(from, ArraycopySrcPrefetchDistance+off, Assembler::severalReads);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2133
        }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2134
        if (ArraycopyDstPrefetchDistance > 0) {
10566
630c177ec580 7081933: Use zeroing elimination optimization for large array
kvn
parents: 10545
diff changeset
  2135
          __ prefetch(to, ArraycopyDstPrefetchDistance+off, Assembler::severalWritesAndPossiblyReads);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2136
        }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2137
      }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2138
      __ ldx(from,  off+0, O4);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2139
      __ ldx(from,  off+8, O5);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2140
      if (use_bis) {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2141
        __ stxa(O4, to,  off+0);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2142
        __ stxa(O5, to,  off+8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2143
      } else {
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2144
        __ stx(O4, to,  off+0);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2145
        __ stx(O5, to,  off+8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2146
      }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2147
    }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2148
    __ deccc(count, 8);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2149
    __ inc(from, 64);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2150
    __ brx(Assembler::greaterEqual, false, Assembler::pt, L_loop);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2151
    __ delayed()->inc(to, 64);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2152
  }
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2153
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2154
  //
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2155
  //  Generate core code for disjoint long copy (and oop copy on 64-bit).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2156
  //  "aligned" is ignored, because we must make the stronger
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2157
  //  assumption that both addresses are always 64-bit aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2158
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2159
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2160
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2161
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2162
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2163
  //
5248
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2164
  // count -= 2;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2165
  // if ( count >= 0 ) { // >= 2 elements
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2166
  //   if ( count > 6) { // >= 8 elements
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2167
  //     count -= 6; // original count - 8
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2168
  //     do {
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2169
  //       copy_8_elements;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2170
  //       count -= 8;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2171
  //     } while ( count >= 0 );
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2172
  //     count += 6;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2173
  //   }
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2174
  //   if ( count >= 0 ) { // >= 2 elements
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2175
  //     do {
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2176
  //       copy_2_elements;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2177
  //     } while ( (count=count-2) >= 0 );
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2178
  //   }
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2179
  // }
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2180
  // count += 2;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2181
  // if ( count != 0 ) { // 1 element left
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2182
  //   copy_1_element;
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2183
  // }
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2184
  //
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2185
  void generate_disjoint_long_copy_core(bool aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2186
    Label L_copy_8_bytes, L_copy_16_bytes, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2187
    const Register from    = O0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2188
    const Register to      = O1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2189
    const Register count   = O2;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2190
    const Register offset0 = O4;  // element offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2191
    const Register offset8 = O5;  // next element offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2192
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2193
    __ deccc(count, 2);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2194
    __ mov(G0, offset0);   // offset from start of arrays (0)
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2195
    __ brx(Assembler::negative, false, Assembler::pn, L_copy_8_bytes );
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2196
    __ delayed()->add(offset0, 8, offset8);
5248
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2197
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2198
    // Copy by 64 bytes chunks
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2199
5248
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2200
    const Register from64 = O3;  // source address
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2201
    const Register to64   = G3;  // destination address
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2202
    __ subcc(count, 6, O3);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2203
    __ brx(Assembler::negative, false, Assembler::pt, L_copy_16_bytes );
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2204
    __ delayed()->mov(to,   to64);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2205
    // Now we can use O4(offset0), O5(offset8) as temps
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2206
    __ mov(O3, count);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2207
    // count >= 0 (original count - 8)
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2208
    __ mov(from, from64);
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2209
24326
d3fdd5c16fe0 8022070: Compilation error in stubGenerator_sparc.cpp with some compilers
mikael
parents: 22505
diff changeset
  2210
    disjoint_copy_core(from64, to64, count, 3, 64, &StubGenerator::copy_64_bytes_loop);
5248
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2211
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2212
      // Restore O4(offset0), O5(offset8)
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2213
      __ sub(from64, from, offset0);
10512
935fc9d89f08 7039731: arraycopy could use prefetch on SPARC
kvn
parents: 10501
diff changeset
  2214
      __ inccc(count, 6); // restore count
5248
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2215
      __ brx(Assembler::negative, false, Assembler::pn, L_copy_8_bytes );
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2216
      __ delayed()->add(offset0, 8, offset8);
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2217
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2218
      // Copy by 16 bytes chunks
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  2219
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2220
    __ BIND(L_copy_16_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2221
      __ ldx(from, offset0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2222
      __ ldx(from, offset8, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2223
      __ deccc(count, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2224
      __ stx(O3, to, offset0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2225
      __ inc(offset0, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2226
      __ stx(G3, to, offset8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2227
      __ brx(Assembler::greaterEqual, false, Assembler::pt, L_copy_16_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2228
      __ delayed()->inc(offset8, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2229
5248
5eddf88cc3c9 6940677: Use 64 bytes chunk copy for arraycopy on Sparc
kvn
parents: 5046
diff changeset
  2230
      // Copy last 8 bytes
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2231
    __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2232
      __ inccc(count, 2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2233
      __ brx(Assembler::zero, true, Assembler::pn, L_exit );
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2234
      __ delayed()->mov(offset0, offset8); // Set O5 used by other stubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2235
      __ ldx(from, offset0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2236
      __ stx(O3, to, offset0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2237
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2238
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2239
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2240
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2241
  //  Generate stub for disjoint long copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2242
  //  "aligned" is ignored, because we must make the stronger
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2243
  //  assumption that both addresses are always 64-bit aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2244
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2245
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2246
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2247
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2248
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2249
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2250
  address generate_disjoint_long_copy(bool aligned, address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2251
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2252
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2253
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2254
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2255
    assert_clean_int(O2, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2256
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2257
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2258
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2259
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2260
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2261
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2262
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2263
    generate_disjoint_long_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2264
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2265
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2266
    inc_counter_np(SharedRuntime::_jlong_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2267
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2268
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2269
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2270
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2271
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2272
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2273
  //  Generate core code for conjoint long copy (and oop copy on 64-bit).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2274
  //  "aligned" is ignored, because we must make the stronger
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2275
  //  assumption that both addresses are always 64-bit aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2276
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2277
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2278
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2279
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2280
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2281
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2282
  void generate_conjoint_long_copy_core(bool aligned) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2283
    // Do reverse copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2284
    Label L_copy_8_bytes, L_copy_16_bytes, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2285
    const Register from    = O0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2286
    const Register to      = O1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2287
    const Register count   = O2;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2288
    const Register offset8 = O4;  // element offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2289
    const Register offset0 = O5;  // previous element offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2290
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2291
      __ subcc(count, 1, count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2292
      __ brx(Assembler::lessEqual, false, Assembler::pn, L_copy_8_bytes );
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2293
      __ delayed()->sllx(count, LogBytesPerLong, offset8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2294
      __ sub(offset8, 8, offset0);
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  2295
      __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2296
    __ BIND(L_copy_16_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2297
      __ ldx(from, offset8, O2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2298
      __ ldx(from, offset0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2299
      __ stx(O2, to, offset8);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2300
      __ deccc(offset8, 16);      // use offset8 as counter
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2301
      __ stx(O3, to, offset0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2302
      __ brx(Assembler::greater, false, Assembler::pt, L_copy_16_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2303
      __ delayed()->dec(offset0, 16);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2304
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2305
    __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2306
      __ brx(Assembler::negative, false, Assembler::pn, L_exit );
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2307
      __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2308
      __ ldx(from, 0, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2309
      __ stx(O3, to, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2310
    __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2311
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2312
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2313
  //  Generate stub for conjoint long copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2314
  //  "aligned" is ignored, because we must make the stronger
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2315
  //  assumption that both addresses are always 64-bit aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2316
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2317
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2318
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2319
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2320
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2321
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2322
  address generate_conjoint_long_copy(bool aligned, address nooverlap_target,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2323
                                      address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2324
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2325
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2326
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2327
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2328
    assert(aligned, "Should always be aligned");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2329
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2330
    assert_clean_int(O2, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2331
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2332
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2333
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2334
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2335
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2336
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2337
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2338
    array_overlap_test(nooverlap_target, 3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2339
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2340
    generate_conjoint_long_copy_core(aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2341
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2342
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2343
    inc_counter_np(SharedRuntime::_jlong_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2344
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2345
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2346
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2347
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2348
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2349
  //  Generate stub for disjoint oop copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2350
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2351
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2352
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2353
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2354
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2355
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2356
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2357
  address generate_disjoint_oop_copy(bool aligned, address *entry, const char *name,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2358
                                     bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2359
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2360
    const Register from  = O0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2361
    const Register to    = O1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2362
    const Register count = O2;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2363
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2364
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2365
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2366
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2367
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2368
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2369
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2370
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2371
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2372
      // caller can pass a 64-bit byte count here
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2373
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2374
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2375
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2376
    // save arguments for barrier generation
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2377
    __ mov(to, G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2378
    __ mov(count, G5);
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2379
    gen_write_ref_array_pre_barrier(G1, G5, dest_uninitialized);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2380
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2381
    if (UseCompressedOops) {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2382
      generate_disjoint_int_copy_core(aligned);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2383
    } else {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2384
      generate_disjoint_long_copy_core(aligned);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2385
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2386
    // O0 is used as temp register
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2387
    gen_write_ref_array_post_barrier(G1, G5, O0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2388
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2389
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2390
    inc_counter_np(SharedRuntime::_oop_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2391
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2392
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2393
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2394
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2395
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2396
  //  Generate stub for conjoint oop copy.  If "aligned" is true, the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2397
  //  "from" and "to" addresses are assumed to be heapword aligned.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2398
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2399
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2400
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2401
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2402
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2403
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2404
  address generate_conjoint_oop_copy(bool aligned, address nooverlap_target,
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2405
                                     address *entry, const char *name,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2406
                                     bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2407
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2408
    const Register from  = O0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2409
    const Register to    = O1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2410
    const Register count = O2;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2411
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2412
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2413
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2414
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2415
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2416
    assert_clean_int(count, O3);     // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2417
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2418
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2419
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2420
      // caller can pass a 64-bit byte count here
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2421
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2422
    }
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2423
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2424
    array_overlap_test(nooverlap_target, LogBytesPerHeapOop);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2425
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2426
    // save arguments for barrier generation
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2427
    __ mov(to, G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2428
    __ mov(count, G5);
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2429
    gen_write_ref_array_pre_barrier(G1, G5, dest_uninitialized);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2430
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2431
    if (UseCompressedOops) {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2432
      generate_conjoint_int_copy_core(aligned);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2433
    } else {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2434
      generate_conjoint_long_copy_core(aligned);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2435
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2436
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2437
    // O0 is used as temp register
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2438
    gen_write_ref_array_post_barrier(G1, G5, O0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2439
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2440
    // O3, O4 are used as temp registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2441
    inc_counter_np(SharedRuntime::_oop_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2442
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2443
    __ delayed()->mov(G0, O0); // return 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2444
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2445
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2446
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2447
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2448
  // Helper for generating a dynamic type check.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2449
  // Smashes only the given temp registers.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2450
  void generate_type_check(Register sub_klass,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2451
                           Register super_check_offset,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2452
                           Register super_klass,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2453
                           Register temp,
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2454
                           Label& L_success) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2455
    assert_different_registers(sub_klass, super_check_offset, super_klass, temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2456
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2457
    BLOCK_COMMENT("type_check:");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2458
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2459
    Label L_miss, L_pop_to_miss;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2460
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2461
    assert_clean_int(super_check_offset, temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2462
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2463
    __ check_klass_subtype_fast_path(sub_klass, super_klass, temp, noreg,
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2464
                                     &L_success, &L_miss, NULL,
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2465
                                     super_check_offset);
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2466
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2467
    BLOCK_COMMENT("type_check_slow_path:");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2468
    __ save_frame(0);
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2469
    __ check_klass_subtype_slow_path(sub_klass->after_save(),
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2470
                                     super_klass->after_save(),
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2471
                                     L0, L1, L2, L4,
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2472
                                     NULL, &L_pop_to_miss);
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  2473
    __ ba(L_success);
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2474
    __ delayed()->restore();
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2475
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2476
    __ bind(L_pop_to_miss);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2477
    __ restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2478
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2479
    // Fall through on failure!
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2480
    __ BIND(L_miss);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2481
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2482
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2483
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2484
  //  Generate stub for checked oop copy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2485
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2486
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2487
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2488
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2489
  //      count: O2 treated as signed
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2490
  //      ckoff: O3 (super_check_offset)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2491
  //      ckval: O4 (super_klass)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2492
  //      ret:   O0 zero for success; (-1^K) where K is partial transfer count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2493
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2494
  address generate_checkcast_copy(const char *name, address *entry, bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2495
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2496
    const Register O0_from   = O0;      // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2497
    const Register O1_to     = O1;      // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2498
    const Register O2_count  = O2;      // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2499
    const Register O3_ckoff  = O3;      // super_check_offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2500
    const Register O4_ckval  = O4;      // super_klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2501
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2502
    const Register O5_offset = O5;      // loop var, with stride wordSize
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2503
    const Register G1_remain = G1;      // loop var, with stride -1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2504
    const Register G3_oop    = G3;      // actual oop copied
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2505
    const Register G4_klass  = G4;      // oop._klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2506
    const Register G5_super  = G5;      // oop._klass._primary_supers[ckval]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2507
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2508
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2509
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2510
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2511
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2512
#ifdef ASSERT
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2513
    // We sometimes save a frame (see generate_type_check below).
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2514
    // If this will cause trouble, let's fail now instead of later.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2515
    __ save_frame(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2516
    __ restore();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2517
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2518
6770
a67870aaedb3 6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents: 6464
diff changeset
  2519
    assert_clean_int(O2_count, G1);     // Make sure 'count' is clean int.
a67870aaedb3 6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents: 6464
diff changeset
  2520
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2521
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2522
    // caller guarantees that the arrays really are different
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2523
    // otherwise, we would have to make conjoint checks
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2524
    { Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2525
      __ mov(O3, G1);           // spill: overlap test smashes O3
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2526
      __ mov(O4, G4);           // spill: overlap test smashes O4
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2527
      array_overlap_test(L, LogBytesPerHeapOop);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2528
      __ stop("checkcast_copy within a single array");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2529
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2530
      __ mov(G1, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2531
      __ mov(G4, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2532
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2533
#endif //ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2534
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2535
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2536
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2537
      // caller can pass a 64-bit byte count here (from generic stub)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2538
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2539
    }
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2540
    gen_write_ref_array_pre_barrier(O1_to, O2_count, dest_uninitialized);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2541
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2542
    Label load_element, store_element, do_card_marks, fail, done;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2543
    __ addcc(O2_count, 0, G1_remain);   // initialize loop index, and test it
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2544
    __ brx(Assembler::notZero, false, Assembler::pt, load_element);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2545
    __ delayed()->mov(G0, O5_offset);   // offset from start of arrays
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2546
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2547
    // Empty array:  Nothing to do.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2548
    inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2549
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2550
    __ delayed()->set(0, O0);           // return 0 on (trivial) success
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2551
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2552
    // ======== begin loop ========
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2553
    // (Loop is rotated; its entry is load_element.)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2554
    // Loop variables:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2555
    //   (O5 = 0; ; O5 += wordSize) --- offset from src, dest arrays
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2556
    //   (O2 = len; O2 != 0; O2--) --- number of oops *remaining*
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2557
    //   G3, G4, G5 --- current oop, oop.klass, oop.klass.super
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5248
diff changeset
  2558
    __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2559
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2560
    __ BIND(store_element);
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2561
    __ deccc(G1_remain);                // decrement the count
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2562
    __ store_heap_oop(G3_oop, O1_to, O5_offset); // store the oop
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2563
    __ inc(O5_offset, heapOopSize);     // step to next offset
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2564
    __ brx(Assembler::zero, true, Assembler::pt, do_card_marks);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2565
    __ delayed()->set(0, O0);           // return -1 on success
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2566
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2567
    // ======== loop entry is here ========
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2568
    __ BIND(load_element);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2569
    __ load_heap_oop(O0_from, O5_offset, G3_oop);  // load the oop
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  2570
    __ br_null_short(G3_oop, Assembler::pt, store_element);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2571
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2572
    __ load_klass(G3_oop, G4_klass); // query the object klass
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2573
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2574
    generate_type_check(G4_klass, O3_ckoff, O4_ckval, G5_super,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2575
                        // branch to this on success:
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2576
                        store_element);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2577
    // ======== end loop ========
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2578
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2579
    // It was a real error; we must depend on the caller to finish the job.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2580
    // Register G1 has number of *remaining* oops, O2 number of *total* oops.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2581
    // Emit GC store barriers for the oops we have copied (O2 minus G1),
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2582
    // and report their number to the caller.
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2583
    __ BIND(fail);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2584
    __ subcc(O2_count, G1_remain, O2_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2585
    __ brx(Assembler::zero, false, Assembler::pt, done);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2586
    __ delayed()->not1(O2_count, O0);   // report (-1^K) to caller
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2587
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2588
    __ BIND(do_card_marks);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2589
    gen_write_ref_array_post_barrier(O1_to, O2_count, O3);   // store check on O1[0..O2]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2590
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 2254
diff changeset
  2591
    __ BIND(done);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2592
    inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr, O3, O4);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2593
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2594
    __ delayed()->nop();             // return value in 00
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2595
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2596
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2597
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2598
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2599
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2600
  //  Generate 'unsafe' array copy stub
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2601
  //  Though just as safe as the other stubs, it takes an unscaled
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2602
  //  size_t argument instead of an element count.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2603
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2604
  // Arguments for generated stub:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2605
  //      from:  O0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2606
  //      to:    O1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2607
  //      count: O2 byte count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2608
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2609
  // Examines the alignment of the operands and dispatches
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2610
  // to a long, int, short, or byte copy loop.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2611
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2612
  address generate_unsafe_copy(const char* name,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2613
                               address byte_copy_entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2614
                               address short_copy_entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2615
                               address int_copy_entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2616
                               address long_copy_entry) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2617
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2618
    const Register O0_from   = O0;      // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2619
    const Register O1_to     = O1;      // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2620
    const Register O2_count  = O2;      // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2621
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2622
    const Register G1_bits   = G1;      // test copy of low bits
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2623
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2624
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2625
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2626
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2627
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2628
    // bump this on entry, not on exit:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2629
    inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr, G1, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2630
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2631
    __ or3(O0_from, O1_to, G1_bits);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2632
    __ or3(O2_count,       G1_bits, G1_bits);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2633
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2634
    __ btst(BytesPerLong-1, G1_bits);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2635
    __ br(Assembler::zero, true, Assembler::pt,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2636
          long_copy_entry, relocInfo::runtime_call_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2637
    // scale the count on the way out:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2638
    __ delayed()->srax(O2_count, LogBytesPerLong, O2_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2639
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2640
    __ btst(BytesPerInt-1, G1_bits);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2641
    __ br(Assembler::zero, true, Assembler::pt,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2642
          int_copy_entry, relocInfo::runtime_call_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2643
    // scale the count on the way out:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2644
    __ delayed()->srax(O2_count, LogBytesPerInt, O2_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2645
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2646
    __ btst(BytesPerShort-1, G1_bits);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2647
    __ br(Assembler::zero, true, Assembler::pt,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2648
          short_copy_entry, relocInfo::runtime_call_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2649
    // scale the count on the way out:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2650
    __ delayed()->srax(O2_count, LogBytesPerShort, O2_count);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2651
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2652
    __ br(Assembler::always, false, Assembler::pt,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2653
          byte_copy_entry, relocInfo::runtime_call_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2654
    __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2655
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2656
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2657
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2658
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2659
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2660
  // Perform range checks on the proposed arraycopy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2661
  // Kills the two temps, but nothing else.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2662
  // Also, clean the sign bits of src_pos and dst_pos.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2663
  void arraycopy_range_checks(Register src,     // source array oop (O0)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2664
                              Register src_pos, // source position (O1)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2665
                              Register dst,     // destination array oo (O2)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2666
                              Register dst_pos, // destination position (O3)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2667
                              Register length,  // length of copy (O4)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2668
                              Register temp1, Register temp2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2669
                              Label& L_failed) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2670
    BLOCK_COMMENT("arraycopy_range_checks:");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2671
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2672
    //  if (src_pos + length > arrayOop(src)->length() ) FAIL;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2673
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2674
    const Register array_length = temp1;  // scratch
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2675
    const Register end_pos      = temp2;  // scratch
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2676
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2677
    // Note:  This next instruction may be in the delay slot of a branch:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2678
    __ add(length, src_pos, end_pos);  // src_pos + length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2679
    __ lduw(src, arrayOopDesc::length_offset_in_bytes(), array_length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2680
    __ cmp(end_pos, array_length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2681
    __ br(Assembler::greater, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2682
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2683
    //  if (dst_pos + length > arrayOop(dst)->length() ) FAIL;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2684
    __ delayed()->add(length, dst_pos, end_pos); // dst_pos + length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2685
    __ lduw(dst, arrayOopDesc::length_offset_in_bytes(), array_length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2686
    __ cmp(end_pos, array_length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2687
    __ br(Assembler::greater, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2688
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2689
    // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2690
    // Move with sign extension can be used since they are positive.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2691
    __ delayed()->signx(src_pos, src_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2692
    __ signx(dst_pos, dst_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2693
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2694
    BLOCK_COMMENT("arraycopy_range_checks done");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2695
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2696
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2697
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2698
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2699
  //  Generate generic array copy stubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2700
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2701
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2702
  //    O0    -  src oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2703
  //    O1    -  src_pos
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2704
  //    O2    -  dst oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2705
  //    O3    -  dst_pos
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2706
  //    O4    -  element count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2707
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2708
  //  Output:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2709
  //    O0 ==  0  -  success
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2710
  //    O0 == -1  -  need to call System.arraycopy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2711
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2712
  address generate_generic_copy(const char *name,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2713
                                address entry_jbyte_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2714
                                address entry_jshort_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2715
                                address entry_jint_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2716
                                address entry_oop_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2717
                                address entry_jlong_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2718
                                address entry_checkcast_arraycopy) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2719
    Label L_failed, L_objArray;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2720
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2721
    // Input registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2722
    const Register src      = O0;  // source array oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2723
    const Register src_pos  = O1;  // source position
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2724
    const Register dst      = O2;  // destination array oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2725
    const Register dst_pos  = O3;  // destination position
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2726
    const Register length   = O4;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2727
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2728
    // registers used as temp
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2729
    const Register G3_src_klass = G3; // source array klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2730
    const Register G4_dst_klass = G4; // destination array klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2731
    const Register G5_lh        = G5; // layout handler
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2732
    const Register O5_temp      = O5;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2733
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2734
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2735
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2736
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2737
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2738
    // bump this on entry, not on exit:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2739
    inc_counter_np(SharedRuntime::_generic_array_copy_ctr, G1, G3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2740
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2741
    // In principle, the int arguments could be dirty.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2742
    //assert_clean_int(src_pos, G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2743
    //assert_clean_int(dst_pos, G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2744
    //assert_clean_int(length, G1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2745
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2746
    //-----------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2747
    // Assembler stubs will be used for this call to arraycopy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2748
    // if the following conditions are met:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2749
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2750
    // (1) src and dst must not be null.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2751
    // (2) src_pos must not be negative.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2752
    // (3) dst_pos must not be negative.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2753
    // (4) length  must not be negative.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2754
    // (5) src klass and dst klass should be the same and not NULL.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2755
    // (6) src and dst should be arrays.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2756
    // (7) src_pos + length must not exceed length of src.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2757
    // (8) dst_pos + length must not exceed length of dst.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2758
    BLOCK_COMMENT("arraycopy initial argument checks");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2759
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2760
    //  if (src == NULL) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2761
    __ br_null(src, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2762
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2763
    //  if (src_pos < 0) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2764
    __ delayed()->tst(src_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2765
    __ br(Assembler::negative, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2766
    __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2767
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2768
    //  if (dst == NULL) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2769
    __ br_null(dst, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2770
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2771
    //  if (dst_pos < 0) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2772
    __ delayed()->tst(dst_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2773
    __ br(Assembler::negative, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2774
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2775
    //  if (length < 0) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2776
    __ delayed()->tst(length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2777
    __ br(Assembler::negative, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2778
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2779
    BLOCK_COMMENT("arraycopy argument klass checks");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2780
    //  get src->klass()
19979
ebe1dbb6e1aa 8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents: 18740
diff changeset
  2781
    if (UseCompressedClassPointers) {
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2782
      __ delayed()->nop(); // ??? not good
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2783
      __ load_klass(src, G3_src_klass);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2784
    } else {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2785
      __ delayed()->ld_ptr(src, oopDesc::klass_offset_in_bytes(), G3_src_klass);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2786
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2787
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2788
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2789
    //  assert(src->klass() != NULL);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2790
    BLOCK_COMMENT("assert klasses not null");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2791
    { Label L_a, L_b;
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  2792
      __ br_notnull_short(G3_src_klass, Assembler::pt, L_b); // it is broken if klass is NULL
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2793
      __ bind(L_a);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2794
      __ stop("broken null klass");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2795
      __ bind(L_b);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2796
      __ load_klass(dst, G4_dst_klass);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2797
      __ br_null(G4_dst_klass, false, Assembler::pn, L_a); // this would be broken also
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2798
      __ delayed()->mov(G0, G4_dst_klass);      // scribble the temp
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2799
      BLOCK_COMMENT("assert done");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2800
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2801
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2802
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2803
    // Load layout helper
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2804
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2805
    //  |array_tag|     | header_size | element_type |     |log2_element_size|
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2806
    // 32        30    24            16              8     2                 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2807
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2808
    //   array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2809
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2810
11430
718fc06da49a 7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents: 10566
diff changeset
  2811
    int lh_offset = in_bytes(Klass::layout_helper_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2812
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2813
    // Load 32-bits signed value. Use br() instruction with it to check icc.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2814
    __ lduw(G3_src_klass, lh_offset, G5_lh);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2815
19979
ebe1dbb6e1aa 8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents: 18740
diff changeset
  2816
    if (UseCompressedClassPointers) {
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2817
      __ load_klass(dst, G4_dst_klass);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2818
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2819
    // Handle objArrays completely differently...
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2820
    juint objArray_lh = Klass::array_layout_helper(T_OBJECT);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2821
    __ set(objArray_lh, O5_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2822
    __ cmp(G5_lh,       O5_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2823
    __ br(Assembler::equal, false, Assembler::pt, L_objArray);
19979
ebe1dbb6e1aa 8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents: 18740
diff changeset
  2824
    if (UseCompressedClassPointers) {
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2825
      __ delayed()->nop();
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2826
    } else {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2827
      __ delayed()->ld_ptr(dst, oopDesc::klass_offset_in_bytes(), G4_dst_klass);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2828
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2829
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2830
    //  if (src->klass() != dst->klass()) return -1;
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  2831
    __ cmp_and_brx_short(G3_src_klass, G4_dst_klass, Assembler::notEqual, Assembler::pn, L_failed);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2832
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2833
    //  if (!src->is_Array()) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2834
    __ cmp(G5_lh, Klass::_lh_neutral_value); // < 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2835
    __ br(Assembler::greaterEqual, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2836
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2837
    // At this point, it is known to be a typeArray (array_tag 0x3).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2838
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2839
    __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2840
    { Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2841
      jint lh_prim_tag_in_place = (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2842
      __ set(lh_prim_tag_in_place, O5_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2843
      __ cmp(G5_lh,                O5_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2844
      __ br(Assembler::greaterEqual, false, Assembler::pt, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2845
      __ delayed()->nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2846
      __ stop("must be a primitive array");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2847
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2848
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2849
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2850
    __ delayed();                               // match next insn to prev branch
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2851
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2852
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2853
    arraycopy_range_checks(src, src_pos, dst, dst_pos, length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2854
                           O5_temp, G4_dst_klass, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2855
13952
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2856
    // TypeArrayKlass
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2857
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2858
    // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2859
    // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2860
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2861
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2862
    const Register G4_offset = G4_dst_klass;    // array offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2863
    const Register G3_elsize = G3_src_klass;    // log2 element size
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2864
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2865
    __ srl(G5_lh, Klass::_lh_header_size_shift, G4_offset);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2866
    __ and3(G4_offset, Klass::_lh_header_size_mask, G4_offset); // array_offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2867
    __ add(src, G4_offset, src);       // src array offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2868
    __ add(dst, G4_offset, dst);       // dst array offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2869
    __ and3(G5_lh, Klass::_lh_log2_element_size_mask, G3_elsize); // log2 element size
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2870
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2871
    // next registers should be set before the jump to corresponding stub
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2872
    const Register from     = O0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2873
    const Register to       = O1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2874
    const Register count    = O2;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2875
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2876
    // 'from', 'to', 'count' registers should be set in this order
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2877
    // since they are the same as 'src', 'src_pos', 'dst'.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2878
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2879
    BLOCK_COMMENT("scale indexes to element size");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2880
    __ sll_ptr(src_pos, G3_elsize, src_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2881
    __ sll_ptr(dst_pos, G3_elsize, dst_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2882
    __ add(src, src_pos, from);       // src_addr
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2883
    __ add(dst, dst_pos, to);         // dst_addr
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2884
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2885
    BLOCK_COMMENT("choose copy loop based on element size");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2886
    __ cmp(G3_elsize, 0);
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2887
    __ br(Assembler::equal, true, Assembler::pt, entry_jbyte_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2888
    __ delayed()->signx(length, count); // length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2889
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2890
    __ cmp(G3_elsize, LogBytesPerShort);
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2891
    __ br(Assembler::equal, true, Assembler::pt, entry_jshort_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2892
    __ delayed()->signx(length, count); // length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2893
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2894
    __ cmp(G3_elsize, LogBytesPerInt);
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2895
    __ br(Assembler::equal, true, Assembler::pt, entry_jint_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2896
    __ delayed()->signx(length, count); // length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2897
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2898
    { Label L;
10252
0981ce1c3eef 7063628: Use cbcond on T4
kvn
parents: 10004
diff changeset
  2899
      __ cmp_and_br_short(G3_elsize, LogBytesPerLong, Assembler::equal, Assembler::pt, L);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2900
      __ stop("must be long copy, but elsize is wrong");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2901
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2902
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2903
#endif
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2904
    __ br(Assembler::always, false, Assembler::pt, entry_jlong_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2905
    __ delayed()->signx(length, count); // length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2906
13952
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2907
    // ObjArrayKlass
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2908
  __ BIND(L_objArray);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2909
    // live at this point:  G3_src_klass, G4_dst_klass, src[_pos], dst[_pos], length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2910
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2911
    Label L_plain_copy, L_checkcast_copy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2912
    //  test array classes for subtyping
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2913
    __ cmp(G3_src_klass, G4_dst_klass);         // usual case is exact equality
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2914
    __ brx(Assembler::notEqual, true, Assembler::pn, L_checkcast_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2915
    __ delayed()->lduw(G4_dst_klass, lh_offset, O5_temp); // hoisted from below
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2916
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2917
    // Identically typed arrays can be copied without element-wise checks.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2918
    arraycopy_range_checks(src, src_pos, dst, dst_pos, length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2919
                           O5_temp, G5_lh, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2920
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2921
    __ add(src, arrayOopDesc::base_offset_in_bytes(T_OBJECT), src); //src offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2922
    __ add(dst, arrayOopDesc::base_offset_in_bytes(T_OBJECT), dst); //dst offset
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2923
    __ sll_ptr(src_pos, LogBytesPerHeapOop, src_pos);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2924
    __ sll_ptr(dst_pos, LogBytesPerHeapOop, dst_pos);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2925
    __ add(src, src_pos, from);       // src_addr
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2926
    __ add(dst, dst_pos, to);         // dst_addr
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2927
  __ BIND(L_plain_copy);
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2928
    __ br(Assembler::always, false, Assembler::pt, entry_oop_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2929
    __ delayed()->signx(length, count); // length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2930
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2931
  __ BIND(L_checkcast_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2932
    // live at this point:  G3_src_klass, G4_dst_klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2933
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2934
      // Before looking at dst.length, make sure dst is also an objArray.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2935
      // lduw(G4_dst_klass, lh_offset, O5_temp); // hoisted to delay slot
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2936
      __ cmp(G5_lh,                    O5_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2937
      __ br(Assembler::notEqual, false, Assembler::pn, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2938
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2939
      // It is safe to examine both src.length and dst.length.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2940
      __ delayed();                             // match next insn to prev branch
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2941
      arraycopy_range_checks(src, src_pos, dst, dst_pos, length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2942
                             O5_temp, G5_lh, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2943
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2944
      // Marshal the base address arguments now, freeing registers.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2945
      __ add(src, arrayOopDesc::base_offset_in_bytes(T_OBJECT), src); //src offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2946
      __ add(dst, arrayOopDesc::base_offset_in_bytes(T_OBJECT), dst); //dst offset
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2947
      __ sll_ptr(src_pos, LogBytesPerHeapOop, src_pos);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2948
      __ sll_ptr(dst_pos, LogBytesPerHeapOop, dst_pos);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2949
      __ add(src, src_pos, from);               // src_addr
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2950
      __ add(dst, dst_pos, to);                 // dst_addr
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2951
      __ signx(length, count);                  // length (reloaded)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2952
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2953
      Register sco_temp = O3;                   // this register is free now
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2954
      assert_different_registers(from, to, count, sco_temp,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2955
                                 G4_dst_klass, G3_src_klass);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2956
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2957
      // Generate the type check.
11430
718fc06da49a 7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents: 10566
diff changeset
  2958
      int sco_offset = in_bytes(Klass::super_check_offset_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2959
      __ lduw(G4_dst_klass, sco_offset, sco_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2960
      generate_type_check(G3_src_klass, sco_temp, G4_dst_klass,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2961
                          O5_temp, L_plain_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2962
13952
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2963
      // Fetch destination element klass from the ObjArrayKlass header.
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2964
      int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2965
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2966
      // the checkcast_copy loop needs two extra arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2967
      __ ld_ptr(G4_dst_klass, ek_offset, O4);   // dest elem klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2968
      // lduw(O4, sco_offset, O3);              // sco of elem klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2969
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  2970
      __ br(Assembler::always, false, Assembler::pt, entry_checkcast_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2971
      __ delayed()->lduw(O4, sco_offset, O3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2972
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2973
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2974
  __ BIND(L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2975
    __ retl();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2976
    __ delayed()->sub(G0, 1, O0); // return -1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2977
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2978
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2979
10501
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2980
  //
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2981
  //  Generate stub for heap zeroing.
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2982
  //  "to" address is aligned to jlong (8 bytes).
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2983
  //
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2984
  // Arguments for generated stub:
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2985
  //      to:    O0
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2986
  //      count: O1 treated as signed (count of HeapWord)
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2987
  //             count could be 0
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2988
  //
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2989
  address generate_zero_aligned_words(const char* name) {
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2990
    __ align(CodeEntryAlignment);
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2991
    StubCodeMark mark(this, "StubRoutines", name);
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2992
    address start = __ pc();
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2993
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2994
    const Register to    = O0;   // source array address
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2995
    const Register count = O1;   // HeapWords count
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2996
    const Register temp  = O2;   // scratch
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2997
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2998
    Label Ldone;
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  2999
    __ sllx(count, LogHeapWordSize, count); // to bytes count
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3000
    // Use BIS for zeroing
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3001
    __ bis_zeroing(to, count, temp, Ldone);
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3002
    __ bind(Ldone);
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3003
    __ retl();
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3004
    __ delayed()->nop();
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3005
    return start;
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3006
}
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3007
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3008
  void generate_arraycopy_stubs() {
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3009
    address entry;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3010
    address entry_jbyte_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3011
    address entry_jshort_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3012
    address entry_jint_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3013
    address entry_oop_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3014
    address entry_jlong_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3015
    address entry_checkcast_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3016
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3017
    //*** jbyte
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3018
    // Always need aligned and unaligned versions
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3019
    StubRoutines::_jbyte_disjoint_arraycopy         = generate_disjoint_byte_copy(false, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3020
                                                                                  "jbyte_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3021
    StubRoutines::_jbyte_arraycopy                  = generate_conjoint_byte_copy(false, entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3022
                                                                                  &entry_jbyte_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3023
                                                                                  "jbyte_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3024
    StubRoutines::_arrayof_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3025
                                                                                  "arrayof_jbyte_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3026
    StubRoutines::_arrayof_jbyte_arraycopy          = generate_conjoint_byte_copy(true, entry, NULL,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3027
                                                                                  "arrayof_jbyte_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3028
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3029
    //*** jshort
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3030
    // Always need aligned and unaligned versions
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3031
    StubRoutines::_jshort_disjoint_arraycopy         = generate_disjoint_short_copy(false, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3032
                                                                                    "jshort_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3033
    StubRoutines::_jshort_arraycopy                  = generate_conjoint_short_copy(false, entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3034
                                                                                    &entry_jshort_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3035
                                                                                    "jshort_arraycopy");
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3036
    StubRoutines::_arrayof_jshort_disjoint_arraycopy = generate_disjoint_short_copy(true, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3037
                                                                                    "arrayof_jshort_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3038
    StubRoutines::_arrayof_jshort_arraycopy          = generate_conjoint_short_copy(true, entry, NULL,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3039
                                                                                    "arrayof_jshort_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3040
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3041
    //*** jint
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3042
    // Aligned versions
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3043
    StubRoutines::_arrayof_jint_disjoint_arraycopy = generate_disjoint_int_copy(true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3044
                                                                                "arrayof_jint_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3045
    StubRoutines::_arrayof_jint_arraycopy          = generate_conjoint_int_copy(true, entry, &entry_jint_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3046
                                                                                "arrayof_jint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3047
    // In 64 bit we need both aligned and unaligned versions of jint arraycopy.
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3048
    // entry_jint_arraycopy always points to the unaligned version (notice that we overwrite it).
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3049
    StubRoutines::_jint_disjoint_arraycopy         = generate_disjoint_int_copy(false, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3050
                                                                                "jint_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3051
    StubRoutines::_jint_arraycopy                  = generate_conjoint_int_copy(false, entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3052
                                                                                &entry_jint_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3053
                                                                                "jint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3054
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3055
    //*** jlong
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3056
    // It is always aligned
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3057
    StubRoutines::_arrayof_jlong_disjoint_arraycopy = generate_disjoint_long_copy(true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3058
                                                                                  "arrayof_jlong_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3059
    StubRoutines::_arrayof_jlong_arraycopy          = generate_conjoint_long_copy(true, entry, &entry_jlong_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3060
                                                                                  "arrayof_jlong_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3061
    StubRoutines::_jlong_disjoint_arraycopy         = StubRoutines::_arrayof_jlong_disjoint_arraycopy;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3062
    StubRoutines::_jlong_arraycopy                  = StubRoutines::_arrayof_jlong_arraycopy;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3063
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3064
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3065
    //*** oops
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3066
    // Aligned versions
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3067
    StubRoutines::_arrayof_oop_disjoint_arraycopy        = generate_disjoint_oop_copy(true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3068
                                                                                      "arrayof_oop_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3069
    StubRoutines::_arrayof_oop_arraycopy                 = generate_conjoint_oop_copy(true, entry, &entry_oop_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3070
                                                                                      "arrayof_oop_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3071
    // Aligned versions without pre-barriers
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3072
    StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = generate_disjoint_oop_copy(true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3073
                                                                                      "arrayof_oop_disjoint_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3074
                                                                                      /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3075
    StubRoutines::_arrayof_oop_arraycopy_uninit          = generate_conjoint_oop_copy(true, entry, NULL,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3076
                                                                                      "arrayof_oop_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3077
                                                                                      /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3078
    if (UseCompressedOops) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3079
      // With compressed oops we need unaligned versions, notice that we overwrite entry_oop_arraycopy.
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3080
      StubRoutines::_oop_disjoint_arraycopy            = generate_disjoint_oop_copy(false, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3081
                                                                                    "oop_disjoint_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3082
      StubRoutines::_oop_arraycopy                     = generate_conjoint_oop_copy(false, entry, &entry_oop_arraycopy,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3083
                                                                                    "oop_arraycopy");
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3084
      // Unaligned versions without pre-barriers
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3085
      StubRoutines::_oop_disjoint_arraycopy_uninit     = generate_disjoint_oop_copy(false, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3086
                                                                                    "oop_disjoint_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3087
                                                                                    /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3088
      StubRoutines::_oop_arraycopy_uninit              = generate_conjoint_oop_copy(false, entry, NULL,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3089
                                                                                    "oop_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3090
                                                                                    /*dest_uninitialized*/true);
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  3091
    } else {
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3092
      // oop arraycopy is always aligned on 32bit and 64bit without compressed oops
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3093
      StubRoutines::_oop_disjoint_arraycopy            = StubRoutines::_arrayof_oop_disjoint_arraycopy;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3094
      StubRoutines::_oop_arraycopy                     = StubRoutines::_arrayof_oop_arraycopy;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3095
      StubRoutines::_oop_disjoint_arraycopy_uninit     = StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3096
      StubRoutines::_oop_arraycopy_uninit              = StubRoutines::_arrayof_oop_arraycopy_uninit;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3097
    }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3098
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3099
    StubRoutines::_checkcast_arraycopy        = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3100
    StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3101
                                                                        /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  3102
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3103
    StubRoutines::_unsafe_arraycopy    = generate_unsafe_copy("unsafe_arraycopy",
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3104
                                                              entry_jbyte_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3105
                                                              entry_jshort_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3106
                                                              entry_jint_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3107
                                                              entry_jlong_arraycopy);
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3108
    StubRoutines::_generic_arraycopy   = generate_generic_copy("generic_arraycopy",
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3109
                                                               entry_jbyte_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3110
                                                               entry_jshort_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3111
                                                               entry_jint_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3112
                                                               entry_oop_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3113
                                                               entry_jlong_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7397
diff changeset
  3114
                                                               entry_checkcast_arraycopy);
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3115
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3116
    StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3117
    StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3118
    StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3119
    StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3120
    StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 6057
diff changeset
  3121
    StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
10501
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3122
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3123
    if (UseBlockZeroing) {
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3124
      StubRoutines::_zero_aligned_words = generate_zero_aligned_words("zero_aligned_words");
5bce84af0883 7059037: Use BIS for zeroing on T4
kvn
parents: 10252
diff changeset
  3125
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3126
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3127
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3128
  address generate_aescrypt_encryptBlock() {
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3129
    // required since we read expanded key 'int' array starting first element without alignment considerations
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3130
    assert((arrayOopDesc::base_offset_in_bytes(T_INT) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3131
           "the following code assumes that first element of an int array is aligned to 8 bytes");
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3132
    __ align(CodeEntryAlignment);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3133
    StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3134
    Label L_load_misaligned_input, L_load_expanded_key, L_doLast128bit, L_storeOutput, L_store_misaligned_output;
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3135
    address start = __ pc();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3136
    Register from = O0; // source byte array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3137
    Register to = O1;   // destination byte array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3138
    Register key = O2;  // expanded key array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3139
    const Register keylen = O4; //reg for storing expanded key array length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3140
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3141
    // read expanded key length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3142
    __ ldsw(Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)), keylen, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3143
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3144
    // Method to address arbitrary alignment for load instructions:
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3145
    // Check last 3 bits of 'from' address to see if it is aligned to 8-byte boundary
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3146
    // If zero/aligned then continue with double FP load instructions
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3147
    // If not zero/mis-aligned then alignaddr will set GSR.align with number of bytes to skip during faligndata
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3148
    // alignaddr will also convert arbitrary aligned 'from' address to nearest 8-byte aligned address
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3149
    // load 3 * 8-byte components (to read 16 bytes input) in 3 different FP regs starting at this aligned address
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3150
    // faligndata will then extract (based on GSR.align value) the appropriate 8 bytes from the 2 source regs
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3151
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3152
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3153
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3154
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_input);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3155
    __ delayed()->alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3156
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3157
    // aligned case: load input into F54-F56
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3158
    __ ldf(FloatRegisterImpl::D, from, 0, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3159
    __ ldf(FloatRegisterImpl::D, from, 8, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3160
    __ ba_short(L_load_expanded_key);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3161
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3162
    __ BIND(L_load_misaligned_input);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3163
    __ ldf(FloatRegisterImpl::D, from, 0, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3164
    __ ldf(FloatRegisterImpl::D, from, 8, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3165
    __ ldf(FloatRegisterImpl::D, from, 16, F58);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3166
    __ faligndata(F54, F56, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3167
    __ faligndata(F56, F58, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3168
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3169
    __ BIND(L_load_expanded_key);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3170
    // Since we load expanded key buffers starting first element, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3171
    for ( int i = 0;  i <= 38; i += 2 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3172
      __ ldf(FloatRegisterImpl::D, key, i*4, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3173
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3174
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3175
    // perform cipher transformation
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3176
    __ fxor(FloatRegisterImpl::D, F0, F54, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3177
    __ fxor(FloatRegisterImpl::D, F2, F56, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3178
    // rounds 1 through 8
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3179
    for ( int i = 4;  i <= 28; i += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3180
      __ aes_eround01(as_FloatRegister(i), F54, F56, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3181
      __ aes_eround23(as_FloatRegister(i+2), F54, F56, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3182
      __ aes_eround01(as_FloatRegister(i+4), F58, F60, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3183
      __ aes_eround23(as_FloatRegister(i+6), F58, F60, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3184
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3185
    __ aes_eround01(F36, F54, F56, F58); //round 9
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3186
    __ aes_eround23(F38, F54, F56, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3187
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3188
    // 128-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3189
    __ cmp_and_brx_short(keylen, 44, Assembler::equal, Assembler::pt, L_doLast128bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3190
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3191
    for ( int i = 40;  i <= 50; i += 2 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3192
      __ ldf(FloatRegisterImpl::D, key, i*4, as_FloatRegister(i) );
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3193
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3194
    __ aes_eround01(F40, F58, F60, F54); //round 10
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3195
    __ aes_eround23(F42, F58, F60, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3196
    __ aes_eround01(F44, F54, F56, F58); //round 11
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3197
    __ aes_eround23(F46, F54, F56, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3198
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3199
    // 192-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3200
    __ cmp_and_brx_short(keylen, 52, Assembler::equal, Assembler::pt, L_storeOutput);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3201
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3202
    __ ldf(FloatRegisterImpl::D, key, 208, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3203
    __ aes_eround01(F48, F58, F60, F54); //round 12
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3204
    __ aes_eround23(F50, F58, F60, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3205
    __ ldf(FloatRegisterImpl::D, key, 216, F46);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3206
    __ ldf(FloatRegisterImpl::D, key, 224, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3207
    __ ldf(FloatRegisterImpl::D, key, 232, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3208
    __ aes_eround01(F52, F54, F56, F58); //round 13
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3209
    __ aes_eround23(F46, F54, F56, F60);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3210
    __ ba_short(L_storeOutput);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3211
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3212
    __ BIND(L_doLast128bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3213
    __ ldf(FloatRegisterImpl::D, key, 160, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3214
    __ ldf(FloatRegisterImpl::D, key, 168, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3215
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3216
    __ BIND(L_storeOutput);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3217
    // perform last round of encryption common for all key sizes
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3218
    __ aes_eround01_l(F48, F58, F60, F54); //last round
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3219
    __ aes_eround23_l(F50, F58, F60, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3220
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3221
    // Method to address arbitrary alignment for store instructions:
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3222
    // Check last 3 bits of 'dest' address to see if it is aligned to 8-byte boundary
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3223
    // If zero/aligned then continue with double FP store instructions
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3224
    // If not zero/mis-aligned then edge8n will generate edge mask in result reg (O3 in below case)
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3225
    // Example: If dest address is 0x07 and nearest 8-byte aligned address is 0x00 then edge mask will be 00000001
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3226
    // Compute (8-n) where n is # of bytes skipped by partial store(stpartialf) inst from edge mask, n=7 in this case
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3227
    // We get the value of n from the andcc that checks 'dest' alignment. n is available in O5 in below case.
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3228
    // Set GSR.align to (8-n) using alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3229
    // Circular byte shift store values by n places so that the original bytes are at correct position for stpartialf
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3230
    // Set the arbitrarily aligned 'dest' address to nearest 8-byte aligned address
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3231
    // Store (partial) the original first (8-n) bytes starting at the original 'dest' address
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3232
    // Negate the edge mask so that the subsequent stpartialf can store the original (8-n-1)th through 8th bytes at appropriate address
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3233
    // We need to execute this process for both the 8-byte result values
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3234
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3235
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3236
    __ andcc(to, 7, O5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3237
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3238
    __ delayed()->edge8n(to, G0, O3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3239
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3240
    // aligned case: store output into the destination array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3241
    __ stf(FloatRegisterImpl::D, F54, to, 0);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3242
    __ retl();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3243
    __ delayed()->stf(FloatRegisterImpl::D, F56, to, 8);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3244
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3245
    __ BIND(L_store_misaligned_output);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3246
    __ add(to, 8, O4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3247
    __ mov(8, O2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3248
    __ sub(O2, O5, O2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3249
    __ alignaddr(O2, G0, O2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3250
    __ faligndata(F54, F54, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3251
    __ faligndata(F56, F56, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3252
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3253
    __ and3(O4, -8, O4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3254
    __ stpartialf(to, O3, F54, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3255
    __ stpartialf(O4, O3, F56, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3256
    __ add(to, 8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3257
    __ add(O4, 8, O4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3258
    __ orn(G0, O3, O3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3259
    __ stpartialf(to, O3, F54, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3260
    __ retl();
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3261
    __ delayed()->stpartialf(O4, O3, F56, Assembler::ASI_PST8_PRIMARY);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3262
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3263
    return start;
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3264
  }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3265
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3266
  address generate_aescrypt_decryptBlock() {
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3267
    assert((arrayOopDesc::base_offset_in_bytes(T_INT) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3268
           "the following code assumes that first element of an int array is aligned to 8 bytes");
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3269
    // required since we read original key 'byte' array as well in the decryption stubs
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3270
    assert((arrayOopDesc::base_offset_in_bytes(T_BYTE) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3271
           "the following code assumes that first element of a byte array is aligned to 8 bytes");
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3272
    __ align(CodeEntryAlignment);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3273
    StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3274
    address start = __ pc();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3275
    Label L_load_misaligned_input, L_load_original_key, L_expand192bit, L_expand256bit, L_reload_misaligned_input;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3276
    Label L_256bit_transform, L_common_transform, L_store_misaligned_output;
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3277
    Register from = O0; // source byte array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3278
    Register to = O1;   // destination byte array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3279
    Register key = O2;  // expanded key array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3280
    Register original_key = O3;  // original key array only required during decryption
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3281
    const Register keylen = O4;  // reg for storing expanded key array length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3282
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3283
    // read expanded key array length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3284
    __ ldsw(Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)), keylen, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3285
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3286
    // save 'from' since we may need to recheck alignment in case of 256-bit decryption
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3287
    __ mov(from, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3288
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3289
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3290
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3291
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_input);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3292
    __ delayed()->alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3293
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3294
    // aligned case: load input into F52-F54
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3295
    __ ldf(FloatRegisterImpl::D, from, 0, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3296
    __ ldf(FloatRegisterImpl::D, from, 8, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3297
    __ ba_short(L_load_original_key);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3298
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3299
    __ BIND(L_load_misaligned_input);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3300
    __ ldf(FloatRegisterImpl::D, from, 0, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3301
    __ ldf(FloatRegisterImpl::D, from, 8, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3302
    __ ldf(FloatRegisterImpl::D, from, 16, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3303
    __ faligndata(F52, F54, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3304
    __ faligndata(F54, F56, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3305
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3306
    __ BIND(L_load_original_key);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3307
    // load original key from SunJCE expanded decryption key
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3308
    // Since we load original key buffer starting first element, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3309
    for ( int i = 0;  i <= 3; i++ ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3310
      __ ldf(FloatRegisterImpl::S, original_key, i*4, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3311
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3312
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3313
    // 256-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3314
    __ cmp_and_brx_short(keylen, 60, Assembler::equal, Assembler::pn, L_expand256bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3315
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3316
    // 192-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3317
    __ cmp_and_brx_short(keylen, 52, Assembler::equal, Assembler::pn, L_expand192bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3318
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3319
    // 128-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3320
    // perform key expansion since SunJCE decryption-key expansion is not compatible with SPARC crypto instructions
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3321
    for ( int i = 0;  i <= 36; i += 4 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3322
      __ aes_kexpand1(as_FloatRegister(i), as_FloatRegister(i+2), i/4, as_FloatRegister(i+4));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3323
      __ aes_kexpand2(as_FloatRegister(i+2), as_FloatRegister(i+4), as_FloatRegister(i+6));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3324
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3325
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3326
    // perform 128-bit key specific inverse cipher transformation
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3327
    __ fxor(FloatRegisterImpl::D, F42, F54, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3328
    __ fxor(FloatRegisterImpl::D, F40, F52, F52);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3329
    __ ba_short(L_common_transform);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3330
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3331
    __ BIND(L_expand192bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3332
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3333
    // start loading rest of the 192-bit key
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3334
    __ ldf(FloatRegisterImpl::S, original_key, 16, F4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3335
    __ ldf(FloatRegisterImpl::S, original_key, 20, F5);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3336
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3337
    // perform key expansion since SunJCE decryption-key expansion is not compatible with SPARC crypto instructions
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3338
    for ( int i = 0;  i <= 36; i += 6 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3339
      __ aes_kexpand1(as_FloatRegister(i), as_FloatRegister(i+4), i/6, as_FloatRegister(i+6));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3340
      __ aes_kexpand2(as_FloatRegister(i+2), as_FloatRegister(i+6), as_FloatRegister(i+8));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3341
      __ aes_kexpand2(as_FloatRegister(i+4), as_FloatRegister(i+8), as_FloatRegister(i+10));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3342
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3343
    __ aes_kexpand1(F42, F46, 7, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3344
    __ aes_kexpand2(F44, F48, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3345
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3346
    // perform 192-bit key specific inverse cipher transformation
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3347
    __ fxor(FloatRegisterImpl::D, F50, F54, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3348
    __ fxor(FloatRegisterImpl::D, F48, F52, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3349
    __ aes_dround23(F46, F52, F54, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3350
    __ aes_dround01(F44, F52, F54, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3351
    __ aes_dround23(F42, F56, F58, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3352
    __ aes_dround01(F40, F56, F58, F52);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3353
    __ ba_short(L_common_transform);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3354
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3355
    __ BIND(L_expand256bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3356
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3357
    // load rest of the 256-bit key
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3358
    for ( int i = 4;  i <= 7; i++ ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3359
      __ ldf(FloatRegisterImpl::S, original_key, i*4, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3360
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3361
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3362
    // perform key expansion since SunJCE decryption-key expansion is not compatible with SPARC crypto instructions
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3363
    for ( int i = 0;  i <= 40; i += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3364
      __ aes_kexpand1(as_FloatRegister(i), as_FloatRegister(i+6), i/8, as_FloatRegister(i+8));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3365
      __ aes_kexpand2(as_FloatRegister(i+2), as_FloatRegister(i+8), as_FloatRegister(i+10));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3366
      __ aes_kexpand0(as_FloatRegister(i+4), as_FloatRegister(i+10), as_FloatRegister(i+12));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3367
      __ aes_kexpand2(as_FloatRegister(i+6), as_FloatRegister(i+12), as_FloatRegister(i+14));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3368
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3369
    __ aes_kexpand1(F48, F54, 6, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3370
    __ aes_kexpand2(F50, F56, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3371
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3372
    for ( int i = 0;  i <= 6; i += 2 ) {
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3373
      __ fsrc2(FloatRegisterImpl::D, as_FloatRegister(58-i), as_FloatRegister(i));
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3374
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3375
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3376
    // reload original 'from' address
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3377
    __ mov(G1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3378
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3379
    // re-check 8-byte alignment
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3380
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3381
    __ br(Assembler::notZero, true, Assembler::pn, L_reload_misaligned_input);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3382
    __ delayed()->alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3383
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3384
    // aligned case: load input into F52-F54
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3385
    __ ldf(FloatRegisterImpl::D, from, 0, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3386
    __ ldf(FloatRegisterImpl::D, from, 8, F54);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3387
    __ ba_short(L_256bit_transform);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3388
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3389
    __ BIND(L_reload_misaligned_input);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3390
    __ ldf(FloatRegisterImpl::D, from, 0, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3391
    __ ldf(FloatRegisterImpl::D, from, 8, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3392
    __ ldf(FloatRegisterImpl::D, from, 16, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3393
    __ faligndata(F52, F54, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3394
    __ faligndata(F54, F56, F54);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3395
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3396
    // perform 256-bit key specific inverse cipher transformation
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3397
    __ BIND(L_256bit_transform);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3398
    __ fxor(FloatRegisterImpl::D, F0, F54, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3399
    __ fxor(FloatRegisterImpl::D, F2, F52, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3400
    __ aes_dround23(F4, F52, F54, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3401
    __ aes_dround01(F6, F52, F54, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3402
    __ aes_dround23(F50, F56, F58, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3403
    __ aes_dround01(F48, F56, F58, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3404
    __ aes_dround23(F46, F52, F54, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3405
    __ aes_dround01(F44, F52, F54, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3406
    __ aes_dround23(F42, F56, F58, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3407
    __ aes_dround01(F40, F56, F58, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3408
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3409
    for ( int i = 0;  i <= 7; i++ ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3410
      __ ldf(FloatRegisterImpl::S, original_key, i*4, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3411
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3412
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3413
    // perform inverse cipher transformations common for all key sizes
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3414
    __ BIND(L_common_transform);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3415
    for ( int i = 38;  i >= 6; i -= 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3416
      __ aes_dround23(as_FloatRegister(i), F52, F54, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3417
      __ aes_dround01(as_FloatRegister(i-2), F52, F54, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3418
      if ( i != 6) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3419
        __ aes_dround23(as_FloatRegister(i-4), F56, F58, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3420
        __ aes_dround01(as_FloatRegister(i-6), F56, F58, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3421
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3422
        __ aes_dround23_l(as_FloatRegister(i-4), F56, F58, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3423
        __ aes_dround01_l(as_FloatRegister(i-6), F56, F58, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3424
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3425
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3426
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3427
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3428
    __ andcc(to, 7, O5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3429
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3430
    __ delayed()->edge8n(to, G0, O3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3431
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3432
    // aligned case: store output into the destination array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3433
    __ stf(FloatRegisterImpl::D, F52, to, 0);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3434
    __ retl();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3435
    __ delayed()->stf(FloatRegisterImpl::D, F54, to, 8);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3436
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3437
    __ BIND(L_store_misaligned_output);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3438
    __ add(to, 8, O4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3439
    __ mov(8, O2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3440
    __ sub(O2, O5, O2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3441
    __ alignaddr(O2, G0, O2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3442
    __ faligndata(F52, F52, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3443
    __ faligndata(F54, F54, F54);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3444
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3445
    __ and3(O4, -8, O4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3446
    __ stpartialf(to, O3, F52, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3447
    __ stpartialf(O4, O3, F54, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3448
    __ add(to, 8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3449
    __ add(O4, 8, O4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3450
    __ orn(G0, O3, O3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3451
    __ stpartialf(to, O3, F52, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3452
    __ retl();
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3453
    __ delayed()->stpartialf(O4, O3, F54, Assembler::ASI_PST8_PRIMARY);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3454
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3455
    return start;
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3456
  }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3457
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3458
  address generate_cipherBlockChaining_encryptAESCrypt() {
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3459
    assert((arrayOopDesc::base_offset_in_bytes(T_INT) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3460
           "the following code assumes that first element of an int array is aligned to 8 bytes");
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3461
    assert((arrayOopDesc::base_offset_in_bytes(T_BYTE) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3462
           "the following code assumes that first element of a byte array is aligned to 8 bytes");
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3463
    __ align(CodeEntryAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3464
    StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3465
    Label L_cbcenc128, L_load_misaligned_input_128bit, L_128bit_transform, L_store_misaligned_output_128bit;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3466
    Label L_check_loop_end_128bit, L_cbcenc192, L_load_misaligned_input_192bit, L_192bit_transform;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3467
    Label L_store_misaligned_output_192bit, L_check_loop_end_192bit, L_cbcenc256, L_load_misaligned_input_256bit;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3468
    Label L_256bit_transform, L_store_misaligned_output_256bit, L_check_loop_end_256bit;
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3469
    address start = __ pc();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3470
    Register from = I0; // source byte array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3471
    Register to = I1;   // destination byte array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3472
    Register key = I2;  // expanded key array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3473
    Register rvec = I3; // init vector
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3474
    const Register len_reg = I4; // cipher length
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3475
    const Register keylen = I5;  // reg for storing expanded key array length
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3476
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3477
    __ save_frame(0);
24488
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3478
    // save cipher len to return in the end
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3479
    __ mov(len_reg, L0);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3480
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3481
    // read expanded key length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3482
    __ ldsw(Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)), keylen, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3483
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3484
    // load initial vector, 8-byte alignment is guranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3485
    __ ldf(FloatRegisterImpl::D, rvec, 0, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3486
    __ ldf(FloatRegisterImpl::D, rvec, 8, F62);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3487
    // load key, 8-byte alignment is guranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3488
    __ ldx(key,0,G1);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3489
    __ ldx(key,8,G5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3490
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3491
    // start loading expanded key, 8-byte alignment is guranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3492
    for ( int i = 0, j = 16;  i <= 38; i += 2, j += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3493
      __ ldf(FloatRegisterImpl::D, key, j, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3494
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3495
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3496
    // 128-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3497
    __ cmp_and_brx_short(keylen, 44, Assembler::equal, Assembler::pt, L_cbcenc128);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3498
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3499
    for ( int i = 40, j = 176;  i <= 46; i += 2, j += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3500
      __ ldf(FloatRegisterImpl::D, key, j, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3501
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3502
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3503
    // 192-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3504
    __ cmp_and_brx_short(keylen, 52, Assembler::equal, Assembler::pt, L_cbcenc192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3506
    for ( int i = 48, j = 208;  i <= 54; i += 2, j += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3507
      __ ldf(FloatRegisterImpl::D, key, j, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3508
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3509
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3510
    // 256-bit original key size
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3511
    __ ba_short(L_cbcenc256);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3512
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3513
    __ align(OptoLoopAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3514
    __ BIND(L_cbcenc128);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3515
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3516
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3517
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_input_128bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3518
    __ delayed()->mov(from, L1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3519
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3520
    // aligned case: load input into G3 and G4
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3521
    __ ldx(from,0,G3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3522
    __ ldx(from,8,G4);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3523
    __ ba_short(L_128bit_transform);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3524
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3525
    __ BIND(L_load_misaligned_input_128bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3526
    // can clobber F48, F50 and F52 as they are not used in 128 and 192-bit key encryption
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3527
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3528
    __ ldf(FloatRegisterImpl::D, from, 0, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3529
    __ ldf(FloatRegisterImpl::D, from, 8, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3530
    __ ldf(FloatRegisterImpl::D, from, 16, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3531
    __ faligndata(F48, F50, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3532
    __ faligndata(F50, F52, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3533
    __ movdtox(F48, G3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3534
    __ movdtox(F50, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3535
    __ mov(L1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3536
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3537
    __ BIND(L_128bit_transform);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3538
    __ xor3(G1,G3,G3);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3539
    __ xor3(G5,G4,G4);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3540
    __ movxtod(G3,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3541
    __ movxtod(G4,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3542
    __ fxor(FloatRegisterImpl::D, F60, F56, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3543
    __ fxor(FloatRegisterImpl::D, F62, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3544
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3545
    // TEN_EROUNDS
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3546
    for ( int i = 0;  i <= 32; i += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3547
      __ aes_eround01(as_FloatRegister(i), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3548
      __ aes_eround23(as_FloatRegister(i+2), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3549
      if (i != 32 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3550
        __ aes_eround01(as_FloatRegister(i+4), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3551
        __ aes_eround23(as_FloatRegister(i+6), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3552
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3553
        __ aes_eround01_l(as_FloatRegister(i+4), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3554
        __ aes_eround23_l(as_FloatRegister(i+6), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3555
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3556
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3557
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3558
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3559
    __ andcc(to, 7, L1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3560
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_128bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3561
    __ delayed()->edge8n(to, G0, L2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3562
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3563
    // aligned case: store output into the destination array
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3564
    __ stf(FloatRegisterImpl::D, F60, to, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3565
    __ stf(FloatRegisterImpl::D, F62, to, 8);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3566
    __ ba_short(L_check_loop_end_128bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3567
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3568
    __ BIND(L_store_misaligned_output_128bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3569
    __ add(to, 8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3570
    __ mov(8, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3571
    __ sub(L4, L1, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3572
    __ alignaddr(L4, G0, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3573
    // save cipher text before circular right shift
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3574
    // as it needs to be stored as iv for next block (see code before next retl)
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3575
    __ movdtox(F60, L6);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3576
    __ movdtox(F62, L7);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3577
    __ faligndata(F60, F60, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3578
    __ faligndata(F62, F62, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3579
    __ mov(to, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3580
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3581
    __ and3(L3, -8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3582
    __ stpartialf(to, L2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3583
    __ stpartialf(L3, L2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3584
    __ add(to, 8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3585
    __ add(L3, 8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3586
    __ orn(G0, L2, L2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3587
    __ stpartialf(to, L2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3588
    __ stpartialf(L3, L2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3589
    __ mov(L5, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3590
    __ movxtod(L6, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3591
    __ movxtod(L7, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3592
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3593
    __ BIND(L_check_loop_end_128bit);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3594
    __ add(from, 16, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3595
    __ add(to, 16, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3596
    __ subcc(len_reg, 16, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3597
    __ br(Assembler::notEqual, false, Assembler::pt, L_cbcenc128);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3598
    __ delayed()->nop();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3599
    // re-init intial vector for next block, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3600
    __ stf(FloatRegisterImpl::D, F60, rvec, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3601
    __ stf(FloatRegisterImpl::D, F62, rvec, 8);
24488
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3602
    __ mov(L0, I0);
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3603
    __ ret();
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3604
    __ delayed()->restore();
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3605
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3606
    __ align(OptoLoopAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3607
    __ BIND(L_cbcenc192);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3608
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3609
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3610
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_input_192bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3611
    __ delayed()->mov(from, L1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3612
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3613
    // aligned case: load input into G3 and G4
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3614
    __ ldx(from,0,G3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3615
    __ ldx(from,8,G4);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3616
    __ ba_short(L_192bit_transform);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3617
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3618
    __ BIND(L_load_misaligned_input_192bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3619
    // can clobber F48, F50 and F52 as they are not used in 128 and 192-bit key encryption
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3620
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3621
    __ ldf(FloatRegisterImpl::D, from, 0, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3622
    __ ldf(FloatRegisterImpl::D, from, 8, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3623
    __ ldf(FloatRegisterImpl::D, from, 16, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3624
    __ faligndata(F48, F50, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3625
    __ faligndata(F50, F52, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3626
    __ movdtox(F48, G3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3627
    __ movdtox(F50, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3628
    __ mov(L1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3629
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3630
    __ BIND(L_192bit_transform);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3631
    __ xor3(G1,G3,G3);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3632
    __ xor3(G5,G4,G4);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3633
    __ movxtod(G3,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3634
    __ movxtod(G4,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3635
    __ fxor(FloatRegisterImpl::D, F60, F56, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3636
    __ fxor(FloatRegisterImpl::D, F62, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3637
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3638
    // TWELEVE_EROUNDS
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3639
    for ( int i = 0;  i <= 40; i += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3640
      __ aes_eround01(as_FloatRegister(i), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3641
      __ aes_eround23(as_FloatRegister(i+2), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3642
      if (i != 40 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3643
        __ aes_eround01(as_FloatRegister(i+4), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3644
        __ aes_eround23(as_FloatRegister(i+6), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3645
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3646
        __ aes_eround01_l(as_FloatRegister(i+4), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3647
        __ aes_eround23_l(as_FloatRegister(i+6), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3648
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3649
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3650
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3651
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3652
    __ andcc(to, 7, L1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3653
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_192bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3654
    __ delayed()->edge8n(to, G0, L2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3655
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3656
    // aligned case: store output into the destination array
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3657
    __ stf(FloatRegisterImpl::D, F60, to, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3658
    __ stf(FloatRegisterImpl::D, F62, to, 8);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3659
    __ ba_short(L_check_loop_end_192bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3660
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3661
    __ BIND(L_store_misaligned_output_192bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3662
    __ add(to, 8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3663
    __ mov(8, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3664
    __ sub(L4, L1, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3665
    __ alignaddr(L4, G0, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3666
    __ movdtox(F60, L6);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3667
    __ movdtox(F62, L7);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3668
    __ faligndata(F60, F60, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3669
    __ faligndata(F62, F62, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3670
    __ mov(to, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3671
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3672
    __ and3(L3, -8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3673
    __ stpartialf(to, L2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3674
    __ stpartialf(L3, L2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3675
    __ add(to, 8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3676
    __ add(L3, 8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3677
    __ orn(G0, L2, L2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3678
    __ stpartialf(to, L2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3679
    __ stpartialf(L3, L2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3680
    __ mov(L5, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3681
    __ movxtod(L6, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3682
    __ movxtod(L7, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3683
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3684
    __ BIND(L_check_loop_end_192bit);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3685
    __ add(from, 16, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3686
    __ subcc(len_reg, 16, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3687
    __ add(to, 16, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3688
    __ br(Assembler::notEqual, false, Assembler::pt, L_cbcenc192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3689
    __ delayed()->nop();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3690
    // re-init intial vector for next block, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3691
    __ stf(FloatRegisterImpl::D, F60, rvec, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3692
    __ stf(FloatRegisterImpl::D, F62, rvec, 8);
24488
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3693
    __ mov(L0, I0);
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3694
    __ ret();
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3695
    __ delayed()->restore();
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3696
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3697
    __ align(OptoLoopAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3698
    __ BIND(L_cbcenc256);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3699
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3700
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3701
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_input_256bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3702
    __ delayed()->mov(from, L1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3703
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3704
    // aligned case: load input into G3 and G4
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3705
    __ ldx(from,0,G3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3706
    __ ldx(from,8,G4);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3707
    __ ba_short(L_256bit_transform);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3708
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3709
    __ BIND(L_load_misaligned_input_256bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3710
    // cannot clobber F48, F50 and F52. F56, F58 can be used though
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3711
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3712
    __ movdtox(F60, L2); // save F60 before overwriting
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3713
    __ ldf(FloatRegisterImpl::D, from, 0, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3714
    __ ldf(FloatRegisterImpl::D, from, 8, F58);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3715
    __ ldf(FloatRegisterImpl::D, from, 16, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3716
    __ faligndata(F56, F58, F56);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3717
    __ faligndata(F58, F60, F58);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3718
    __ movdtox(F56, G3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3719
    __ movdtox(F58, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3720
    __ mov(L1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3721
    __ movxtod(L2, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3722
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3723
    __ BIND(L_256bit_transform);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3724
    __ xor3(G1,G3,G3);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3725
    __ xor3(G5,G4,G4);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3726
    __ movxtod(G3,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3727
    __ movxtod(G4,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3728
    __ fxor(FloatRegisterImpl::D, F60, F56, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3729
    __ fxor(FloatRegisterImpl::D, F62, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3730
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3731
    // FOURTEEN_EROUNDS
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3732
    for ( int i = 0;  i <= 48; i += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3733
      __ aes_eround01(as_FloatRegister(i), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3734
      __ aes_eround23(as_FloatRegister(i+2), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3735
      if (i != 48 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3736
        __ aes_eround01(as_FloatRegister(i+4), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3737
        __ aes_eround23(as_FloatRegister(i+6), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3738
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3739
        __ aes_eround01_l(as_FloatRegister(i+4), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3740
        __ aes_eround23_l(as_FloatRegister(i+6), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3741
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3742
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3743
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3744
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3745
    __ andcc(to, 7, L1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3746
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_256bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3747
    __ delayed()->edge8n(to, G0, L2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3748
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3749
    // aligned case: store output into the destination array
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3750
    __ stf(FloatRegisterImpl::D, F60, to, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3751
    __ stf(FloatRegisterImpl::D, F62, to, 8);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3752
    __ ba_short(L_check_loop_end_256bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3753
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3754
    __ BIND(L_store_misaligned_output_256bit);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3755
    __ add(to, 8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3756
    __ mov(8, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3757
    __ sub(L4, L1, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3758
    __ alignaddr(L4, G0, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3759
    __ movdtox(F60, L6);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3760
    __ movdtox(F62, L7);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3761
    __ faligndata(F60, F60, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3762
    __ faligndata(F62, F62, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3763
    __ mov(to, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3764
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3765
    __ and3(L3, -8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3766
    __ stpartialf(to, L2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3767
    __ stpartialf(L3, L2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3768
    __ add(to, 8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3769
    __ add(L3, 8, L3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3770
    __ orn(G0, L2, L2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3771
    __ stpartialf(to, L2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3772
    __ stpartialf(L3, L2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3773
    __ mov(L5, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3774
    __ movxtod(L6, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3775
    __ movxtod(L7, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3776
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3777
    __ BIND(L_check_loop_end_256bit);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3778
    __ add(from, 16, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3779
    __ subcc(len_reg, 16, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3780
    __ add(to, 16, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3781
    __ br(Assembler::notEqual, false, Assembler::pt, L_cbcenc256);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3782
    __ delayed()->nop();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3783
    // re-init intial vector for next block, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3784
    __ stf(FloatRegisterImpl::D, F60, rvec, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3785
    __ stf(FloatRegisterImpl::D, F62, rvec, 8);
24488
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3786
    __ mov(L0, I0);
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3787
    __ ret();
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3788
    __ delayed()->restore();
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3789
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3790
    return start;
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3791
  }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3792
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3793
  address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3794
    assert((arrayOopDesc::base_offset_in_bytes(T_INT) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3795
           "the following code assumes that first element of an int array is aligned to 8 bytes");
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3796
    assert((arrayOopDesc::base_offset_in_bytes(T_BYTE) & 7) == 0,
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3797
           "the following code assumes that first element of a byte array is aligned to 8 bytes");
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3798
    __ align(CodeEntryAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3799
    StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3800
    Label L_cbcdec_end, L_expand192bit, L_expand256bit, L_dec_first_block_start;
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3801
    Label L_dec_first_block128, L_dec_first_block192, L_dec_next2_blocks128, L_dec_next2_blocks192, L_dec_next2_blocks256;
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3802
    Label L_load_misaligned_input_first_block, L_transform_first_block, L_load_misaligned_next2_blocks128, L_transform_next2_blocks128;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3803
    Label L_load_misaligned_next2_blocks192, L_transform_next2_blocks192, L_load_misaligned_next2_blocks256, L_transform_next2_blocks256;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3804
    Label L_store_misaligned_output_first_block, L_check_decrypt_end, L_store_misaligned_output_next2_blocks128;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3805
    Label L_check_decrypt_loop_end128, L_store_misaligned_output_next2_blocks192, L_check_decrypt_loop_end192;
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3806
    Label L_store_misaligned_output_next2_blocks256, L_check_decrypt_loop_end256;
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3807
    address start = __ pc();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3808
    Register from = I0; // source byte array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3809
    Register to = I1;   // destination byte array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3810
    Register key = I2;  // expanded key array
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3811
    Register rvec = I3; // init vector
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3812
    const Register len_reg = I4; // cipher length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3813
    const Register original_key = I5;  // original key array only required during decryption
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3814
    const Register keylen = L6;  // reg for storing expanded key array length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3815
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3816
    __ save_frame(0); //args are read from I* registers since we save the frame in the beginning
24488
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3817
    // save cipher len to return in the end
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  3818
    __ mov(len_reg, L7);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3819
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3820
    // load original key from SunJCE expanded decryption key
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3821
    // Since we load original key buffer starting first element, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3822
    for ( int i = 0;  i <= 3; i++ ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3823
      __ ldf(FloatRegisterImpl::S, original_key, i*4, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3824
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3825
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3826
    // load initial vector, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3827
    __ ldx(rvec,0,L0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3828
    __ ldx(rvec,8,L1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3829
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3830
    // read expanded key array length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3831
    __ ldsw(Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)), keylen, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3832
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3833
    // 256-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3834
    __ cmp_and_brx_short(keylen, 60, Assembler::equal, Assembler::pn, L_expand256bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3835
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3836
    // 192-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3837
    __ cmp_and_brx_short(keylen, 52, Assembler::equal, Assembler::pn, L_expand192bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3838
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3839
    // 128-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3840
    // perform key expansion since SunJCE decryption-key expansion is not compatible with SPARC crypto instructions
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3841
    for ( int i = 0;  i <= 36; i += 4 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3842
      __ aes_kexpand1(as_FloatRegister(i), as_FloatRegister(i+2), i/4, as_FloatRegister(i+4));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3843
      __ aes_kexpand2(as_FloatRegister(i+2), as_FloatRegister(i+4), as_FloatRegister(i+6));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3844
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3845
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3846
    // load expanded key[last-1] and key[last] elements
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3847
    __ movdtox(F40,L2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3848
    __ movdtox(F42,L3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3849
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3850
    __ and3(len_reg, 16, L4);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3851
    __ br_null_short(L4, Assembler::pt, L_dec_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3852
    __ nop();
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3853
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3854
    __ ba_short(L_dec_first_block_start);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3855
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3856
    __ BIND(L_expand192bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3857
    // load rest of the 192-bit key
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3858
    __ ldf(FloatRegisterImpl::S, original_key, 16, F4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3859
    __ ldf(FloatRegisterImpl::S, original_key, 20, F5);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3860
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3861
    // perform key expansion since SunJCE decryption-key expansion is not compatible with SPARC crypto instructions
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3862
    for ( int i = 0;  i <= 36; i += 6 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3863
      __ aes_kexpand1(as_FloatRegister(i), as_FloatRegister(i+4), i/6, as_FloatRegister(i+6));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3864
      __ aes_kexpand2(as_FloatRegister(i+2), as_FloatRegister(i+6), as_FloatRegister(i+8));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3865
      __ aes_kexpand2(as_FloatRegister(i+4), as_FloatRegister(i+8), as_FloatRegister(i+10));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3866
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3867
    __ aes_kexpand1(F42, F46, 7, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3868
    __ aes_kexpand2(F44, F48, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3869
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3870
    // load expanded key[last-1] and key[last] elements
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3871
    __ movdtox(F48,L2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3872
    __ movdtox(F50,L3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3873
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3874
    __ and3(len_reg, 16, L4);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3875
    __ br_null_short(L4, Assembler::pt, L_dec_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3876
    __ nop();
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3877
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3878
    __ ba_short(L_dec_first_block_start);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3879
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3880
    __ BIND(L_expand256bit);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3881
    // load rest of the 256-bit key
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3882
    for ( int i = 4;  i <= 7; i++ ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3883
      __ ldf(FloatRegisterImpl::S, original_key, i*4, as_FloatRegister(i));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3884
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3885
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3886
    // perform key expansion since SunJCE decryption-key expansion is not compatible with SPARC crypto instructions
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3887
    for ( int i = 0;  i <= 40; i += 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3888
      __ aes_kexpand1(as_FloatRegister(i), as_FloatRegister(i+6), i/8, as_FloatRegister(i+8));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3889
      __ aes_kexpand2(as_FloatRegister(i+2), as_FloatRegister(i+8), as_FloatRegister(i+10));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3890
      __ aes_kexpand0(as_FloatRegister(i+4), as_FloatRegister(i+10), as_FloatRegister(i+12));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3891
      __ aes_kexpand2(as_FloatRegister(i+6), as_FloatRegister(i+12), as_FloatRegister(i+14));
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3892
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3893
    __ aes_kexpand1(F48, F54, 6, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3894
    __ aes_kexpand2(F50, F56, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3895
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3896
    // load expanded key[last-1] and key[last] elements
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3897
    __ movdtox(F56,L2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3898
    __ movdtox(F58,L3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3899
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3900
    __ and3(len_reg, 16, L4);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3901
    __ br_null_short(L4, Assembler::pt, L_dec_next2_blocks256);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3902
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3903
    __ BIND(L_dec_first_block_start);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3904
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3905
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3906
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_input_first_block);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3907
    __ delayed()->mov(from, G1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3908
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3909
    // aligned case: load input into L4 and L5
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3910
    __ ldx(from,0,L4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3911
    __ ldx(from,8,L5);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3912
    __ ba_short(L_transform_first_block);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3913
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3914
    __ BIND(L_load_misaligned_input_first_block);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3915
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3916
    // F58, F60, F62 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3917
    __ ldf(FloatRegisterImpl::D, from, 0, F58);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3918
    __ ldf(FloatRegisterImpl::D, from, 8, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3919
    __ ldf(FloatRegisterImpl::D, from, 16, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3920
    __ faligndata(F58, F60, F58);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3921
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3922
    __ movdtox(F58, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3923
    __ movdtox(F60, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3924
    __ mov(G1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3925
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3926
    __ BIND(L_transform_first_block);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3927
    __ xor3(L2,L4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3928
    __ movxtod(G1,F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3929
    __ xor3(L3,L5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3930
    __ movxtod(G1,F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3931
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3932
    // 128-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3933
    __ cmp_and_brx_short(keylen, 44, Assembler::equal, Assembler::pn, L_dec_first_block128);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3934
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3935
    // 192-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3936
    __ cmp_and_brx_short(keylen, 52, Assembler::equal, Assembler::pn, L_dec_first_block192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3937
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3938
    __ aes_dround23(F54, F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3939
    __ aes_dround01(F52, F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3940
    __ aes_dround23(F50, F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3941
    __ aes_dround01(F48, F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3942
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3943
    __ BIND(L_dec_first_block192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3944
    __ aes_dround23(F46, F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3945
    __ aes_dround01(F44, F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3946
    __ aes_dround23(F42, F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3947
    __ aes_dround01(F40, F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3948
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3949
    __ BIND(L_dec_first_block128);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3950
    for ( int i = 38;  i >= 6; i -= 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3951
      __ aes_dround23(as_FloatRegister(i), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3952
      __ aes_dround01(as_FloatRegister(i-2), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3953
      if ( i != 6) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3954
        __ aes_dround23(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3955
        __ aes_dround01(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3956
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3957
        __ aes_dround23_l(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3958
        __ aes_dround01_l(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3959
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3960
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3961
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3962
    __ movxtod(L0,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3963
    __ movxtod(L1,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3964
    __ mov(L4,L0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3965
    __ mov(L5,L1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3966
    __ fxor(FloatRegisterImpl::D, F56, F60, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3967
    __ fxor(FloatRegisterImpl::D, F58, F62, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3968
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3969
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3970
    __ andcc(to, 7, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3971
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_first_block);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3972
    __ delayed()->edge8n(to, G0, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3973
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3974
    // aligned case: store output into the destination array
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3975
    __ stf(FloatRegisterImpl::D, F60, to, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3976
    __ stf(FloatRegisterImpl::D, F62, to, 8);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3977
    __ ba_short(L_check_decrypt_end);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3978
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3979
    __ BIND(L_store_misaligned_output_first_block);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3980
    __ add(to, 8, G3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3981
    __ mov(8, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3982
    __ sub(G4, G1, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3983
    __ alignaddr(G4, G0, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3984
    __ faligndata(F60, F60, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3985
    __ faligndata(F62, F62, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3986
    __ mov(to, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3987
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3988
    __ and3(G3, -8, G3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3989
    __ stpartialf(to, G2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3990
    __ stpartialf(G3, G2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3991
    __ add(to, 8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3992
    __ add(G3, 8, G3);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3993
    __ orn(G0, G2, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3994
    __ stpartialf(to, G2, F60, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3995
    __ stpartialf(G3, G2, F62, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3996
    __ mov(G1, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3997
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  3998
    __ BIND(L_check_decrypt_end);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  3999
    __ add(from, 16, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4000
    __ add(to, 16, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4001
    __ subcc(len_reg, 16, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4002
    __ br(Assembler::equal, false, Assembler::pt, L_cbcdec_end);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4003
    __ delayed()->nop();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4004
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4005
    // 256-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4006
    __ cmp_and_brx_short(keylen, 60, Assembler::equal, Assembler::pn, L_dec_next2_blocks256);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4007
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4008
    // 192-bit original key size
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4009
    __ cmp_and_brx_short(keylen, 52, Assembler::equal, Assembler::pn, L_dec_next2_blocks192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4010
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4011
    __ align(OptoLoopAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4012
    __ BIND(L_dec_next2_blocks128);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4013
    __ nop();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4014
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4015
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4016
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4017
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4018
    __ delayed()->mov(from, G1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4019
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4020
    // aligned case: load input into G4, G5, L4 and L5
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4021
    __ ldx(from,0,G4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4022
    __ ldx(from,8,G5);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4023
    __ ldx(from,16,L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4024
    __ ldx(from,24,L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4025
    __ ba_short(L_transform_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4026
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4027
    __ BIND(L_load_misaligned_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4028
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4029
    // F40, F42, F58, F60, F62 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4030
    __ ldf(FloatRegisterImpl::D, from, 0, F40);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4031
    __ ldf(FloatRegisterImpl::D, from, 8, F42);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4032
    __ ldf(FloatRegisterImpl::D, from, 16, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4033
    __ ldf(FloatRegisterImpl::D, from, 24, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4034
    __ ldf(FloatRegisterImpl::D, from, 32, F58);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4035
    __ faligndata(F40, F42, F40);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4036
    __ faligndata(F42, F60, F42);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4037
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4038
    __ faligndata(F62, F58, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4039
    __ movdtox(F40, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4040
    __ movdtox(F42, G5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4041
    __ movdtox(F60, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4042
    __ movdtox(F62, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4043
    __ mov(G1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4044
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4045
    __ BIND(L_transform_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4046
    // F40:F42 used for first 16-bytes
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4047
    __ xor3(L2,G4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4048
    __ movxtod(G1,F40);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4049
    __ xor3(L3,G5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4050
    __ movxtod(G1,F42);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4051
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4052
    // F60:F62 used for next 16-bytes
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4053
    __ xor3(L2,L4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4054
    __ movxtod(G1,F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4055
    __ xor3(L3,L5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4056
    __ movxtod(G1,F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4057
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4058
    for ( int i = 38;  i >= 6; i -= 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4059
      __ aes_dround23(as_FloatRegister(i), F40, F42, F44);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4060
      __ aes_dround01(as_FloatRegister(i-2), F40, F42, F46);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4061
      __ aes_dround23(as_FloatRegister(i), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4062
      __ aes_dround01(as_FloatRegister(i-2), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4063
      if (i != 6 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4064
        __ aes_dround23(as_FloatRegister(i-4), F46, F44, F42);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4065
        __ aes_dround01(as_FloatRegister(i-6), F46, F44, F40);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4066
        __ aes_dround23(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4067
        __ aes_dround01(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4068
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4069
        __ aes_dround23_l(as_FloatRegister(i-4), F46, F44, F42);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4070
        __ aes_dround01_l(as_FloatRegister(i-6), F46, F44, F40);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4071
        __ aes_dround23_l(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4072
        __ aes_dround01_l(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4073
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4074
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4075
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4076
    __ movxtod(L0,F46);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4077
    __ movxtod(L1,F44);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4078
    __ fxor(FloatRegisterImpl::D, F46, F40, F40);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4079
    __ fxor(FloatRegisterImpl::D, F44, F42, F42);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4080
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4081
    __ movxtod(G4,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4082
    __ movxtod(G5,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4083
    __ mov(L4,L0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4084
    __ mov(L5,L1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4085
    __ fxor(FloatRegisterImpl::D, F56, F60, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4086
    __ fxor(FloatRegisterImpl::D, F58, F62, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4087
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4088
    // For mis-aligned store of 32 bytes of result we can do:
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4089
    // Circular right-shift all 4 FP registers so that 'head' and 'tail'
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4090
    // parts that need to be stored starting at mis-aligned address are in a FP reg
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4091
    // the other 3 FP regs can thus be stored using regular store
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4092
    // we then use the edge + partial-store mechanism to store the 'head' and 'tail' parts
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4093
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4094
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4095
    __ andcc(to, 7, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4096
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4097
    __ delayed()->edge8n(to, G0, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4098
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4099
    // aligned case: store output into the destination array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4100
    __ stf(FloatRegisterImpl::D, F40, to, 0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4101
    __ stf(FloatRegisterImpl::D, F42, to, 8);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4102
    __ stf(FloatRegisterImpl::D, F60, to, 16);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4103
    __ stf(FloatRegisterImpl::D, F62, to, 24);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4104
    __ ba_short(L_check_decrypt_loop_end128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4105
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4106
    __ BIND(L_store_misaligned_output_next2_blocks128);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4107
    __ mov(8, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4108
    __ sub(G4, G1, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4109
    __ alignaddr(G4, G0, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4110
    __ faligndata(F40, F42, F56); // F56 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4111
    __ faligndata(F42, F60, F42);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4112
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4113
    __ faligndata(F62, F40, F40);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4114
    __ mov(to, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4115
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4116
    __ stpartialf(to, G2, F40, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4117
    __ stf(FloatRegisterImpl::D, F56, to, 8);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4118
    __ stf(FloatRegisterImpl::D, F42, to, 16);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4119
    __ stf(FloatRegisterImpl::D, F60, to, 24);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4120
    __ add(to, 32, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4121
    __ orn(G0, G2, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4122
    __ stpartialf(to, G2, F40, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4123
    __ mov(G1, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4124
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4125
    __ BIND(L_check_decrypt_loop_end128);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4126
    __ add(from, 32, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4127
    __ add(to, 32, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4128
    __ subcc(len_reg, 32, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4129
    __ br(Assembler::notEqual, false, Assembler::pt, L_dec_next2_blocks128);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4130
    __ delayed()->nop();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4131
    __ ba_short(L_cbcdec_end);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4132
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4133
    __ align(OptoLoopAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4134
    __ BIND(L_dec_next2_blocks192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4135
    __ nop();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4136
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4137
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4138
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4139
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4140
    __ delayed()->mov(from, G1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4141
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4142
    // aligned case: load input into G4, G5, L4 and L5
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4143
    __ ldx(from,0,G4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4144
    __ ldx(from,8,G5);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4145
    __ ldx(from,16,L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4146
    __ ldx(from,24,L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4147
    __ ba_short(L_transform_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4148
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4149
    __ BIND(L_load_misaligned_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4150
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4151
    // F48, F50, F52, F60, F62 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4152
    __ ldf(FloatRegisterImpl::D, from, 0, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4153
    __ ldf(FloatRegisterImpl::D, from, 8, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4154
    __ ldf(FloatRegisterImpl::D, from, 16, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4155
    __ ldf(FloatRegisterImpl::D, from, 24, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4156
    __ ldf(FloatRegisterImpl::D, from, 32, F52);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4157
    __ faligndata(F48, F50, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4158
    __ faligndata(F50, F60, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4159
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4160
    __ faligndata(F62, F52, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4161
    __ movdtox(F48, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4162
    __ movdtox(F50, G5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4163
    __ movdtox(F60, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4164
    __ movdtox(F62, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4165
    __ mov(G1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4166
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4167
    __ BIND(L_transform_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4168
    // F48:F50 used for first 16-bytes
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4169
    __ xor3(L2,G4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4170
    __ movxtod(G1,F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4171
    __ xor3(L3,G5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4172
    __ movxtod(G1,F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4173
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4174
    // F60:F62 used for next 16-bytes
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4175
    __ xor3(L2,L4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4176
    __ movxtod(G1,F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4177
    __ xor3(L3,L5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4178
    __ movxtod(G1,F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4179
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4180
    for ( int i = 46;  i >= 6; i -= 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4181
      __ aes_dround23(as_FloatRegister(i), F48, F50, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4182
      __ aes_dround01(as_FloatRegister(i-2), F48, F50, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4183
      __ aes_dround23(as_FloatRegister(i), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4184
      __ aes_dround01(as_FloatRegister(i-2), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4185
      if (i != 6 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4186
        __ aes_dround23(as_FloatRegister(i-4), F54, F52, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4187
        __ aes_dround01(as_FloatRegister(i-6), F54, F52, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4188
        __ aes_dround23(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4189
        __ aes_dround01(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4190
      } else {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4191
        __ aes_dround23_l(as_FloatRegister(i-4), F54, F52, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4192
        __ aes_dround01_l(as_FloatRegister(i-6), F54, F52, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4193
        __ aes_dround23_l(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4194
        __ aes_dround01_l(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4195
      }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4196
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4197
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4198
    __ movxtod(L0,F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4199
    __ movxtod(L1,F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4200
    __ fxor(FloatRegisterImpl::D, F54, F48, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4201
    __ fxor(FloatRegisterImpl::D, F52, F50, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4202
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4203
    __ movxtod(G4,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4204
    __ movxtod(G5,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4205
    __ mov(L4,L0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4206
    __ mov(L5,L1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4207
    __ fxor(FloatRegisterImpl::D, F56, F60, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4208
    __ fxor(FloatRegisterImpl::D, F58, F62, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4209
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4210
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4211
    __ andcc(to, 7, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4212
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4213
    __ delayed()->edge8n(to, G0, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4214
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4215
    // aligned case: store output into the destination array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4216
    __ stf(FloatRegisterImpl::D, F48, to, 0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4217
    __ stf(FloatRegisterImpl::D, F50, to, 8);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4218
    __ stf(FloatRegisterImpl::D, F60, to, 16);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4219
    __ stf(FloatRegisterImpl::D, F62, to, 24);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4220
    __ ba_short(L_check_decrypt_loop_end192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4221
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4222
    __ BIND(L_store_misaligned_output_next2_blocks192);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4223
    __ mov(8, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4224
    __ sub(G4, G1, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4225
    __ alignaddr(G4, G0, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4226
    __ faligndata(F48, F50, F56); // F56 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4227
    __ faligndata(F50, F60, F50);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4228
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4229
    __ faligndata(F62, F48, F48);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4230
    __ mov(to, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4231
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4232
    __ stpartialf(to, G2, F48, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4233
    __ stf(FloatRegisterImpl::D, F56, to, 8);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4234
    __ stf(FloatRegisterImpl::D, F50, to, 16);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4235
    __ stf(FloatRegisterImpl::D, F60, to, 24);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4236
    __ add(to, 32, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4237
    __ orn(G0, G2, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4238
    __ stpartialf(to, G2, F48, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4239
    __ mov(G1, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4240
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4241
    __ BIND(L_check_decrypt_loop_end192);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4242
    __ add(from, 32, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4243
    __ add(to, 32, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4244
    __ subcc(len_reg, 32, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4245
    __ br(Assembler::notEqual, false, Assembler::pt, L_dec_next2_blocks192);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4246
    __ delayed()->nop();
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4247
    __ ba_short(L_cbcdec_end);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4248
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4249
    __ align(OptoLoopAlignment);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4250
    __ BIND(L_dec_next2_blocks256);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4251
    __ nop();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4252
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4253
    // check for 8-byte alignment since source byte array may have an arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4254
    __ andcc(from, 7, G0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4255
    __ br(Assembler::notZero, true, Assembler::pn, L_load_misaligned_next2_blocks256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4256
    __ delayed()->mov(from, G1); // save original 'from' address before alignaddr
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4257
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4258
    // aligned case: load input into G4, G5, L4 and L5
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4259
    __ ldx(from,0,G4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4260
    __ ldx(from,8,G5);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4261
    __ ldx(from,16,L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4262
    __ ldx(from,24,L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4263
    __ ba_short(L_transform_next2_blocks256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4264
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4265
    __ BIND(L_load_misaligned_next2_blocks256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4266
    __ alignaddr(from, G0, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4267
    // F0, F2, F4, F60, F62 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4268
    __ ldf(FloatRegisterImpl::D, from, 0, F0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4269
    __ ldf(FloatRegisterImpl::D, from, 8, F2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4270
    __ ldf(FloatRegisterImpl::D, from, 16, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4271
    __ ldf(FloatRegisterImpl::D, from, 24, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4272
    __ ldf(FloatRegisterImpl::D, from, 32, F4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4273
    __ faligndata(F0, F2, F0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4274
    __ faligndata(F2, F60, F2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4275
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4276
    __ faligndata(F62, F4, F62);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4277
    __ movdtox(F0, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4278
    __ movdtox(F2, G5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4279
    __ movdtox(F60, L4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4280
    __ movdtox(F62, L5);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4281
    __ mov(G1, from);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4282
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4283
    __ BIND(L_transform_next2_blocks256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4284
    // F0:F2 used for first 16-bytes
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4285
    __ xor3(L2,G4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4286
    __ movxtod(G1,F0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4287
    __ xor3(L3,G5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4288
    __ movxtod(G1,F2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4289
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4290
    // F60:F62 used for next 16-bytes
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4291
    __ xor3(L2,L4,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4292
    __ movxtod(G1,F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4293
    __ xor3(L3,L5,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4294
    __ movxtod(G1,F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4295
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4296
    __ aes_dround23(F54, F0, F2, F4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4297
    __ aes_dround01(F52, F0, F2, F6);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4298
    __ aes_dround23(F54, F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4299
    __ aes_dround01(F52, F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4300
    __ aes_dround23(F50, F6, F4, F2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4301
    __ aes_dround01(F48, F6, F4, F0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4302
    __ aes_dround23(F50, F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4303
    __ aes_dround01(F48, F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4304
    // save F48:F54 in temp registers
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4305
    __ movdtox(F54,G2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4306
    __ movdtox(F52,G3);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4307
    __ movdtox(F50,G6);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4308
    __ movdtox(F48,G1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4309
    for ( int i = 46;  i >= 14; i -= 8 ) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4310
      __ aes_dround23(as_FloatRegister(i), F0, F2, F4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4311
      __ aes_dround01(as_FloatRegister(i-2), F0, F2, F6);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4312
      __ aes_dround23(as_FloatRegister(i), F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4313
      __ aes_dround01(as_FloatRegister(i-2), F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4314
      __ aes_dround23(as_FloatRegister(i-4), F6, F4, F2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4315
      __ aes_dround01(as_FloatRegister(i-6), F6, F4, F0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4316
      __ aes_dround23(as_FloatRegister(i-4), F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4317
      __ aes_dround01(as_FloatRegister(i-6), F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4318
    }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4319
    // init F48:F54 with F0:F6 values (original key)
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4320
    __ ldf(FloatRegisterImpl::D, original_key, 0, F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4321
    __ ldf(FloatRegisterImpl::D, original_key, 8, F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4322
    __ ldf(FloatRegisterImpl::D, original_key, 16, F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4323
    __ ldf(FloatRegisterImpl::D, original_key, 24, F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4324
    __ aes_dround23(F54, F0, F2, F4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4325
    __ aes_dround01(F52, F0, F2, F6);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4326
    __ aes_dround23(F54, F60, F62, F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4327
    __ aes_dround01(F52, F60, F62, F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4328
    __ aes_dround23_l(F50, F6, F4, F2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4329
    __ aes_dround01_l(F48, F6, F4, F0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4330
    __ aes_dround23_l(F50, F56, F58, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4331
    __ aes_dround01_l(F48, F56, F58, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4332
    // re-init F48:F54 with their original values
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4333
    __ movxtod(G2,F54);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4334
    __ movxtod(G3,F52);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4335
    __ movxtod(G6,F50);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4336
    __ movxtod(G1,F48);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4337
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4338
    __ movxtod(L0,F6);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4339
    __ movxtod(L1,F4);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4340
    __ fxor(FloatRegisterImpl::D, F6, F0, F0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4341
    __ fxor(FloatRegisterImpl::D, F4, F2, F2);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4342
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4343
    __ movxtod(G4,F56);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4344
    __ movxtod(G5,F58);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4345
    __ mov(L4,L0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4346
    __ mov(L5,L1);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4347
    __ fxor(FloatRegisterImpl::D, F56, F60, F60);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4348
    __ fxor(FloatRegisterImpl::D, F58, F62, F62);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4349
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4350
    // check for 8-byte alignment since dest byte array may have arbitrary alignment if offset mod 8 is non-zero
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4351
    __ andcc(to, 7, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4352
    __ br(Assembler::notZero, true, Assembler::pn, L_store_misaligned_output_next2_blocks256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4353
    __ delayed()->edge8n(to, G0, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4354
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4355
    // aligned case: store output into the destination array
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4356
    __ stf(FloatRegisterImpl::D, F0, to, 0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4357
    __ stf(FloatRegisterImpl::D, F2, to, 8);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4358
    __ stf(FloatRegisterImpl::D, F60, to, 16);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4359
    __ stf(FloatRegisterImpl::D, F62, to, 24);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4360
    __ ba_short(L_check_decrypt_loop_end256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4361
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4362
    __ BIND(L_store_misaligned_output_next2_blocks256);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4363
    __ mov(8, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4364
    __ sub(G4, G1, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4365
    __ alignaddr(G4, G0, G4);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4366
    __ faligndata(F0, F2, F56); // F56 can be clobbered
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4367
    __ faligndata(F2, F60, F2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4368
    __ faligndata(F60, F62, F60);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4369
    __ faligndata(F62, F0, F0);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4370
    __ mov(to, G1);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4371
    __ and3(to, -8, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4372
    __ stpartialf(to, G2, F0, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4373
    __ stf(FloatRegisterImpl::D, F56, to, 8);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4374
    __ stf(FloatRegisterImpl::D, F2, to, 16);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4375
    __ stf(FloatRegisterImpl::D, F60, to, 24);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4376
    __ add(to, 32, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4377
    __ orn(G0, G2, G2);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4378
    __ stpartialf(to, G2, F0, Assembler::ASI_PST8_PRIMARY);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4379
    __ mov(G1, to);
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4380
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4381
    __ BIND(L_check_decrypt_loop_end256);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4382
    __ add(from, 32, from);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4383
    __ add(to, 32, to);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4384
    __ subcc(len_reg, 32, len_reg);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4385
    __ br(Assembler::notEqual, false, Assembler::pt, L_dec_next2_blocks256);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4386
    __ delayed()->nop();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4387
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4388
    __ BIND(L_cbcdec_end);
24328
bddefb356fba 8035936: SIGBUS in StubRoutines::aesencryptBlock, solaris-sparc
kvn
parents: 24326
diff changeset
  4389
    // re-init intial vector for next block, 8-byte alignment is guaranteed
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4390
    __ stx(L0, rvec, 0);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4391
    __ stx(L1, rvec, 8);
24488
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  4392
    __ mov(L7, I0);
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  4393
    __ ret();
6872367f4335 8043274: Test compiler/7184394/TestAESMain.java gets NPE on solaris
kvn
parents: 24328
diff changeset
  4394
    __ delayed()->restore();
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4395
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4396
    return start;
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4397
  }
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  4398
24953
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4399
  address generate_sha1_implCompress(bool multi_block, const char *name) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4400
    __ align(CodeEntryAlignment);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4401
    StubCodeMark mark(this, "StubRoutines", name);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4402
    address start = __ pc();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4403
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4404
    Label L_sha1_loop, L_sha1_unaligned_input, L_sha1_unaligned_input_loop;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4405
    int i;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4406
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4407
    Register buf   = O0; // byte[] source+offset
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4408
    Register state = O1; // int[]  SHA.state
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4409
    Register ofs   = O2; // int    offset
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4410
    Register limit = O3; // int    limit
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4411
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4412
    // load state into F0-F4
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4413
    for (i = 0; i < 5; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4414
      __ ldf(FloatRegisterImpl::S, state, i*4, as_FloatRegister(i));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4415
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4416
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4417
    __ andcc(buf, 7, G0);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4418
    __ br(Assembler::notZero, false, Assembler::pn, L_sha1_unaligned_input);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4419
    __ delayed()->nop();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4420
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4421
    __ BIND(L_sha1_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4422
    // load buf into F8-F22
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4423
    for (i = 0; i < 8; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4424
      __ ldf(FloatRegisterImpl::D, buf, i*8, as_FloatRegister(i*2 + 8));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4425
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4426
    __ sha1();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4427
    if (multi_block) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4428
      __ add(ofs, 64, ofs);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4429
      __ add(buf, 64, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4430
      __ cmp_and_brx_short(ofs, limit, Assembler::lessEqual, Assembler::pt, L_sha1_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4431
      __ mov(ofs, O0); // to be returned
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4432
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4433
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4434
    // store F0-F4 into state and return
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4435
    for (i = 0; i < 4; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4436
      __ stf(FloatRegisterImpl::S, as_FloatRegister(i), state, i*4);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4437
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4438
    __ retl();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4439
    __ delayed()->stf(FloatRegisterImpl::S, F4, state, 0x10);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4440
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4441
    __ BIND(L_sha1_unaligned_input);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4442
    __ alignaddr(buf, G0, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4443
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4444
    __ BIND(L_sha1_unaligned_input_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4445
    // load buf into F8-F22
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4446
    for (i = 0; i < 9; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4447
      __ ldf(FloatRegisterImpl::D, buf, i*8, as_FloatRegister(i*2 + 8));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4448
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4449
    for (i = 0; i < 8; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4450
      __ faligndata(as_FloatRegister(i*2 + 8), as_FloatRegister(i*2 + 10), as_FloatRegister(i*2 + 8));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4451
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4452
    __ sha1();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4453
    if (multi_block) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4454
      __ add(ofs, 64, ofs);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4455
      __ add(buf, 64, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4456
      __ cmp_and_brx_short(ofs, limit, Assembler::lessEqual, Assembler::pt, L_sha1_unaligned_input_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4457
      __ mov(ofs, O0); // to be returned
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4458
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4459
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4460
    // store F0-F4 into state and return
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4461
    for (i = 0; i < 4; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4462
      __ stf(FloatRegisterImpl::S, as_FloatRegister(i), state, i*4);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4463
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4464
    __ retl();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4465
    __ delayed()->stf(FloatRegisterImpl::S, F4, state, 0x10);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4466
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4467
    return start;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4468
  }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4469
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4470
  address generate_sha256_implCompress(bool multi_block, const char *name) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4471
    __ align(CodeEntryAlignment);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4472
    StubCodeMark mark(this, "StubRoutines", name);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4473
    address start = __ pc();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4474
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4475
    Label L_sha256_loop, L_sha256_unaligned_input, L_sha256_unaligned_input_loop;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4476
    int i;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4477
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4478
    Register buf   = O0; // byte[] source+offset
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4479
    Register state = O1; // int[]  SHA2.state
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4480
    Register ofs   = O2; // int    offset
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4481
    Register limit = O3; // int    limit
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4482
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4483
    // load state into F0-F7
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4484
    for (i = 0; i < 8; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4485
      __ ldf(FloatRegisterImpl::S, state, i*4, as_FloatRegister(i));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4486
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4487
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4488
    __ andcc(buf, 7, G0);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4489
    __ br(Assembler::notZero, false, Assembler::pn, L_sha256_unaligned_input);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4490
    __ delayed()->nop();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4491
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4492
    __ BIND(L_sha256_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4493
    // load buf into F8-F22
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4494
    for (i = 0; i < 8; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4495
      __ ldf(FloatRegisterImpl::D, buf, i*8, as_FloatRegister(i*2 + 8));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4496
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4497
    __ sha256();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4498
    if (multi_block) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4499
      __ add(ofs, 64, ofs);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4500
      __ add(buf, 64, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4501
      __ cmp_and_brx_short(ofs, limit, Assembler::lessEqual, Assembler::pt, L_sha256_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4502
      __ mov(ofs, O0); // to be returned
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4503
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4504
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4505
    // store F0-F7 into state and return
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4506
    for (i = 0; i < 7; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4507
      __ stf(FloatRegisterImpl::S, as_FloatRegister(i), state, i*4);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4508
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4509
    __ retl();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4510
    __ delayed()->stf(FloatRegisterImpl::S, F7, state, 0x1c);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4511
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4512
    __ BIND(L_sha256_unaligned_input);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4513
    __ alignaddr(buf, G0, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4514
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4515
    __ BIND(L_sha256_unaligned_input_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4516
    // load buf into F8-F22
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4517
    for (i = 0; i < 9; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4518
      __ ldf(FloatRegisterImpl::D, buf, i*8, as_FloatRegister(i*2 + 8));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4519
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4520
    for (i = 0; i < 8; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4521
      __ faligndata(as_FloatRegister(i*2 + 8), as_FloatRegister(i*2 + 10), as_FloatRegister(i*2 + 8));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4522
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4523
    __ sha256();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4524
    if (multi_block) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4525
      __ add(ofs, 64, ofs);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4526
      __ add(buf, 64, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4527
      __ cmp_and_brx_short(ofs, limit, Assembler::lessEqual, Assembler::pt, L_sha256_unaligned_input_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4528
      __ mov(ofs, O0); // to be returned
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4529
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4530
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4531
    // store F0-F7 into state and return
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4532
    for (i = 0; i < 7; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4533
      __ stf(FloatRegisterImpl::S, as_FloatRegister(i), state, i*4);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4534
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4535
    __ retl();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4536
    __ delayed()->stf(FloatRegisterImpl::S, F7, state, 0x1c);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4537
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4538
    return start;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4539
  }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4540
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4541
  address generate_sha512_implCompress(bool multi_block, const char *name) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4542
    __ align(CodeEntryAlignment);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4543
    StubCodeMark mark(this, "StubRoutines", name);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4544
    address start = __ pc();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4545
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4546
    Label L_sha512_loop, L_sha512_unaligned_input, L_sha512_unaligned_input_loop;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4547
    int i;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4548
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4549
    Register buf   = O0; // byte[] source+offset
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4550
    Register state = O1; // long[] SHA5.state
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4551
    Register ofs   = O2; // int    offset
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4552
    Register limit = O3; // int    limit
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4553
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4554
    // load state into F0-F14
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4555
    for (i = 0; i < 8; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4556
      __ ldf(FloatRegisterImpl::D, state, i*8, as_FloatRegister(i*2));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4557
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4558
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4559
    __ andcc(buf, 7, G0);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4560
    __ br(Assembler::notZero, false, Assembler::pn, L_sha512_unaligned_input);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4561
    __ delayed()->nop();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4562
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4563
    __ BIND(L_sha512_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4564
    // load buf into F16-F46
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4565
    for (i = 0; i < 16; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4566
      __ ldf(FloatRegisterImpl::D, buf, i*8, as_FloatRegister(i*2 + 16));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4567
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4568
    __ sha512();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4569
    if (multi_block) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4570
      __ add(ofs, 128, ofs);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4571
      __ add(buf, 128, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4572
      __ cmp_and_brx_short(ofs, limit, Assembler::lessEqual, Assembler::pt, L_sha512_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4573
      __ mov(ofs, O0); // to be returned
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4574
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4575
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4576
    // store F0-F14 into state and return
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4577
    for (i = 0; i < 7; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4578
      __ stf(FloatRegisterImpl::D, as_FloatRegister(i*2), state, i*8);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4579
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4580
    __ retl();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4581
    __ delayed()->stf(FloatRegisterImpl::D, F14, state, 0x38);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4582
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4583
    __ BIND(L_sha512_unaligned_input);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4584
    __ alignaddr(buf, G0, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4585
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4586
    __ BIND(L_sha512_unaligned_input_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4587
    // load buf into F16-F46
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4588
    for (i = 0; i < 17; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4589
      __ ldf(FloatRegisterImpl::D, buf, i*8, as_FloatRegister(i*2 + 16));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4590
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4591
    for (i = 0; i < 16; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4592
      __ faligndata(as_FloatRegister(i*2 + 16), as_FloatRegister(i*2 + 18), as_FloatRegister(i*2 + 16));
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4593
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4594
    __ sha512();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4595
    if (multi_block) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4596
      __ add(ofs, 128, ofs);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4597
      __ add(buf, 128, buf);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4598
      __ cmp_and_brx_short(ofs, limit, Assembler::lessEqual, Assembler::pt, L_sha512_unaligned_input_loop);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4599
      __ mov(ofs, O0); // to be returned
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4600
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4601
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4602
    // store F0-F14 into state and return
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4603
    for (i = 0; i < 7; i++) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4604
      __ stf(FloatRegisterImpl::D, as_FloatRegister(i*2), state, i*8);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4605
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4606
    __ retl();
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4607
    __ delayed()->stf(FloatRegisterImpl::D, F14, state, 0x38);
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4608
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4609
    return start;
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4610
  }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  4611
31404
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4612
  /* Single and multi-block ghash operations */
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4613
  address generate_ghash_processBlocks() {
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4614
      __ align(CodeEntryAlignment);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4615
      Label L_ghash_loop, L_aligned, L_main;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4616
      StubCodeMark mark(this, "StubRoutines", "ghash_processBlocks");
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4617
      address start = __ pc();
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4618
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4619
      Register state = I0;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4620
      Register subkeyH = I1;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4621
      Register data = I2;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4622
      Register len = I3;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4623
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4624
      __ save_frame(0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4625
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4626
      __ ldx(state, 0, O0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4627
      __ ldx(state, 8, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4628
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4629
      // Loop label for multiblock operations
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4630
      __ BIND(L_ghash_loop);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4631
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4632
      // Check if 'data' is unaligned
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4633
      __ andcc(data, 7, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4634
      __ br(Assembler::zero, false, Assembler::pt, L_aligned);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4635
      __ delayed()->nop();
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4636
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4637
      Register left_shift = L1;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4638
      Register right_shift = L2;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4639
      Register data_ptr = L3;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4640
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4641
      // Get left and right shift values in bits
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4642
      __ sll(G1, LogBitsPerByte, left_shift);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4643
      __ mov(64, right_shift);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4644
      __ sub(right_shift, left_shift, right_shift);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4645
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4646
      // Align to read 'data'
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4647
      __ sub(data, G1, data_ptr);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4648
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4649
      // Load first 8 bytes of 'data'
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4650
      __ ldx(data_ptr, 0, O4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4651
      __ sllx(O4, left_shift, O4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4652
      __ ldx(data_ptr, 8, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4653
      __ srlx(O5, right_shift, G4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4654
      __ bset(G4, O4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4655
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4656
      // Load second 8 bytes of 'data'
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4657
      __ sllx(O5, left_shift, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4658
      __ ldx(data_ptr, 16, G4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4659
      __ srlx(G4, right_shift, G4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4660
      __ ba(L_main);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4661
      __ delayed()->bset(G4, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4662
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4663
      // If 'data' is aligned, load normally
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4664
      __ BIND(L_aligned);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4665
      __ ldx(data, 0, O4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4666
      __ ldx(data, 8, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4667
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4668
      __ BIND(L_main);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4669
      __ ldx(subkeyH, 0, O2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4670
      __ ldx(subkeyH, 8, O3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4671
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4672
      __ xor3(O0, O4, O0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4673
      __ xor3(O1, O5, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4674
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4675
      __ xmulxhi(O0, O3, G3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4676
      __ xmulx(O0, O2, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4677
      __ xmulxhi(O1, O2, G4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4678
      __ xmulxhi(O1, O3, G5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4679
      __ xmulx(O0, O3, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4680
      __ xmulx(O1, O3, G2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4681
      __ xmulx(O1, O2, O3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4682
      __ xmulxhi(O0, O2, O4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4683
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4684
      __ mov(0xE1, O0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4685
      __ sllx(O0, 56, O0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4686
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4687
      __ xor3(O5, G3, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4688
      __ xor3(O5, G4, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4689
      __ xor3(G5, G1, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4690
      __ xor3(G1, O3, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4691
      __ srlx(G2, 63, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4692
      __ srlx(G1, 63, G3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4693
      __ sllx(G2, 63, O3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4694
      __ sllx(G2, 58, O2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4695
      __ xor3(O3, O2, O2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4696
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4697
      __ sllx(G1, 1, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4698
      __ or3(G1, O1, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4699
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4700
      __ xor3(G1, O2, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4701
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4702
      __ sllx(G2, 1, G2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4703
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4704
      __ xmulxhi(G1, O0, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4705
      __ xmulx(G1, O0, O2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4706
      __ xmulxhi(G2, O0, O3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4707
      __ xmulx(G2, O0, G1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4708
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4709
      __ xor3(O4, O1, O4);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4710
      __ xor3(O5, O2, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4711
      __ xor3(O5, O3, O5);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4712
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4713
      __ sllx(O4, 1, O2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4714
      __ srlx(O5, 63, O3);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4715
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4716
      __ or3(O2, O3, O0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4717
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4718
      __ sllx(O5, 1, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4719
      __ srlx(G1, 63, O2);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4720
      __ or3(O1, O2, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4721
      __ xor3(O1, G3, O1);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4722
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4723
      __ deccc(len);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4724
      __ br(Assembler::notZero, true, Assembler::pt, L_ghash_loop);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4725
      __ delayed()->add(data, 16, data);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4726
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4727
      __ stx(O0, I0, 0);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4728
      __ stx(O1, I0, 8);
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4729
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4730
      __ ret();
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4731
      __ delayed()->restore();
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4732
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4733
      return start;
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4734
  }
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  4735
31515
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4736
  /**
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4737
   *  Arguments:
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4738
   *
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4739
   * Inputs:
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4740
   *   O0   - int   crc
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4741
   *   O1   - byte* buf
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4742
   *   O2   - int   len
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4743
   *   O3   - int*  table
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4744
   *
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4745
   * Output:
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4746
   *   O0   - int crc result
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4747
   */
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4748
  address generate_updateBytesCRC32C() {
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4749
    assert(UseCRC32CIntrinsics, "need CRC32C instruction");
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4750
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4751
    __ align(CodeEntryAlignment);
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4752
    StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32C");
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4753
    address start = __ pc();
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4754
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4755
    const Register crc   = O0;  // crc
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4756
    const Register buf   = O1;  // source java byte array address
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4757
    const Register len   = O2;  // number of bytes
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4758
    const Register table = O3;  // byteTable
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4759
38237
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  4760
    __ kernel_crc32c(crc, buf, len, table);
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  4761
31515
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4762
    __ retl();
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4763
    __ delayed()->nop();
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4764
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4765
    return start;
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4766
  }
6aed85dadbe6 8073583: C2 support for CRC32C on SPARC
kvn
parents: 31404
diff changeset
  4767
32581
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4768
#define ADLER32_NUM_TEMPS 16
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4769
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4770
  /**
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4771
   *  Arguments:
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4772
   *
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4773
   * Inputs:
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4774
   *   O0   - int   adler
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4775
   *   O1   - byte* buff
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4776
   *   O2   - int   len
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4777
   *
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4778
   * Output:
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4779
   *   O0   - int adler result
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4780
   */
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4781
  address generate_updateBytesAdler32() {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4782
    __ align(CodeEntryAlignment);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4783
    StubCodeMark mark(this, "StubRoutines", "updateBytesAdler32");
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4784
    address start = __ pc();
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4785
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4786
    Label L_cleanup_loop, L_cleanup_loop_check;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4787
    Label L_main_loop_check, L_main_loop, L_inner_loop, L_inner_loop_check;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4788
    Label L_nmax_check_done;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4789
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4790
    // Aliases
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4791
    Register s1     = O0;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4792
    Register s2     = O3;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4793
    Register buff   = O1;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4794
    Register len    = O2;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4795
    Register temp[ADLER32_NUM_TEMPS] = {L0, L1, L2, L3, L4, L5, L6, L7, I0, I1, I2, I3, I4, I5, G3, I7};
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4796
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4797
    // Max number of bytes we can process before having to take the mod
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4798
    // 0x15B0 is 5552 in decimal, the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4799
    unsigned long NMAX = 0x15B0;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4800
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4801
    // Zero-out the upper bits of len
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4802
    __ clruwu(len);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4803
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4804
    // Create the mask 0xFFFF
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4805
    __ set64(0x00FFFF, O4, O5); // O5 is the temp register
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4806
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4807
    // s1 is initialized to the lower 16 bits of adler
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4808
    // s2 is initialized to the upper 16 bits of adler
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4809
    __ srlx(O0, 16, O5); // adler >> 16
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4810
    __ and3(O0, O4, s1); // s1  = (adler & 0xFFFF)
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4811
    __ and3(O5, O4, s2); // s2  = ((adler >> 16) & 0xFFFF)
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4812
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4813
    // The pipelined loop needs at least 16 elements for 1 iteration
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4814
    // It does check this, but it is more effective to skip to the cleanup loop
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4815
    // Setup the constant for cutoff checking
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4816
    __ mov(15, O4);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4817
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4818
    // Check if we are above the cutoff, if not go to the cleanup loop immediately
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4819
    __ cmp_and_br_short(len, O4, Assembler::lessEqualUnsigned, Assembler::pt, L_cleanup_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4820
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4821
    // Free up some registers for our use
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4822
    for (int i = 0; i < ADLER32_NUM_TEMPS; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4823
      __ movxtod(temp[i], as_FloatRegister(2*i));
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4824
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4825
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4826
    // Loop maintenance stuff is done at the end of the loop, so skip to there
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4827
    __ ba_short(L_main_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4828
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4829
    __ BIND(L_main_loop);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4830
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4831
    // Prologue for inner loop
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4832
    __ ldub(buff, 0, L0);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4833
    __ dec(O5);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4834
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4835
    for (int i = 1; i < 8; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4836
      __ ldub(buff, i, temp[i]);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4837
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4838
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4839
    __ inc(buff, 8);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4840
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4841
    // Inner loop processes 16 elements at a time, might never execute if only 16 elements
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4842
    // to be processed by the outter loop
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4843
    __ ba_short(L_inner_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4844
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4845
    __ BIND(L_inner_loop);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4846
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4847
    for (int i = 0; i < 8; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4848
      __ ldub(buff, (2*i), temp[(8+(2*i)) % ADLER32_NUM_TEMPS]);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4849
      __ add(s1, temp[i], s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4850
      __ ldub(buff, (2*i)+1, temp[(8+(2*i)+1) % ADLER32_NUM_TEMPS]);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4851
      __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4852
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4853
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4854
    // Original temp 0-7 used and new loads to temp 0-7 issued
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4855
    // temp 8-15 ready to be consumed
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4856
    __ add(s1, I0, s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4857
    __ dec(O5);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4858
    __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4859
    __ add(s1, I1, s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4860
    __ inc(buff, 16);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4861
    __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4862
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4863
    for (int i = 0; i < 6; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4864
      __ add(s1, temp[10+i], s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4865
      __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4866
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4867
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4868
    __ BIND(L_inner_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4869
    __ nop();
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4870
    __ cmp_and_br_short(O5, 0, Assembler::notEqual, Assembler::pt, L_inner_loop);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4871
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4872
    // Epilogue
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4873
    for (int i = 0; i < 4; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4874
      __ ldub(buff, (2*i), temp[8+(2*i)]);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4875
      __ add(s1, temp[i], s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4876
      __ ldub(buff, (2*i)+1, temp[8+(2*i)+1]);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4877
      __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4878
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4879
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4880
    __ add(s1, temp[4], s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4881
    __ inc(buff, 8);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4882
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4883
    for (int i = 0; i < 11; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4884
      __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4885
      __ add(s1, temp[5+i], s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4886
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4887
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4888
    __ add(s2, s1, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4889
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4890
    // Take the mod for s1 and s2
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4891
    __ set64(0xFFF1, L0, L1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4892
    __ udivx(s1, L0, L1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4893
    __ udivx(s2, L0, L2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4894
    __ mulx(L0, L1, L1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4895
    __ mulx(L0, L2, L2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4896
    __ sub(s1, L1, s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4897
    __ sub(s2, L2, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4898
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4899
    // Make sure there is something left to process
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4900
    __ BIND(L_main_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4901
    __ set64(NMAX, L0, L1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4902
    // k = len < NMAX ? len : NMAX
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4903
    __ cmp_and_br_short(len, L0, Assembler::greaterEqualUnsigned, Assembler::pt, L_nmax_check_done);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4904
    __ andn(len, 0x0F, L0); // only loop a multiple of 16 times
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4905
    __ BIND(L_nmax_check_done);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4906
    __ mov(L0, O5);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4907
    __ sub(len, L0, len); // len -= k
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4908
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4909
    __ srlx(O5, 4, O5); // multiplies of 16
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4910
    __ cmp_and_br_short(O5, 0, Assembler::notEqual, Assembler::pt, L_main_loop);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4911
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4912
    // Restore anything we used, take the mod one last time, combine and return
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4913
    // Restore any registers we saved
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4914
    for (int i = 0; i < ADLER32_NUM_TEMPS; i++) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4915
      __ movdtox(as_FloatRegister(2*i), temp[i]);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4916
    }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4917
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4918
    // There might be nothing left to process
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4919
    __ ba_short(L_cleanup_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4920
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4921
    __ BIND(L_cleanup_loop);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4922
    __ ldub(buff, 0, O4); // load single byte form buffer
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4923
    __ inc(buff); // buff++
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4924
    __ add(s1, O4, s1); // s1 += *buff++;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4925
    __ dec(len); // len--
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4926
    __ add(s1, s2, s2); // s2 += s1;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4927
    __ BIND(L_cleanup_loop_check);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4928
    __ nop();
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4929
    __ cmp_and_br_short(len, 0, Assembler::notEqual, Assembler::pt, L_cleanup_loop);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4930
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4931
    // Take the mod one last time
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4932
    __ set64(0xFFF1, O1, O2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4933
    __ udivx(s1, O1, O2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4934
    __ udivx(s2, O1, O5);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4935
    __ mulx(O1, O2, O2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4936
    __ mulx(O1, O5, O5);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4937
    __ sub(s1, O2, s1);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4938
    __ sub(s2, O5, s2);
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4939
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4940
    // Combine lower bits and higher bits
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4941
    __ sllx(s2, 16, s2); // s2 = s2 << 16
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4942
    __ or3(s1, s2, s1);  // adler = s2 | s1
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4943
    // Final return value is in O0
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4944
    __ retl();
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4945
    __ delayed()->nop();
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4946
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4947
    return start;
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4948
  }
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  4949
47561
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4950
  /**
34205
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4951
   *  Arguments:
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4952
   *
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4953
   * Inputs:
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4954
   *   O0   - int   crc
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4955
   *   O1   - byte* buf
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4956
   *   O2   - int   len
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4957
   *   O3   - int*  table
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4958
   *
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4959
   * Output:
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4960
   *   O0   - int crc result
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4961
   */
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4962
  address generate_updateBytesCRC32() {
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4963
    assert(UseCRC32Intrinsics, "need VIS3 instructions");
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4964
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4965
    __ align(CodeEntryAlignment);
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4966
    StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4967
    address start = __ pc();
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4968
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4969
    const Register crc   = O0; // crc
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4970
    const Register buf   = O1; // source java byte array address
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4971
    const Register len   = O2; // length
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4972
    const Register table = O3; // crc_table address (reuse register)
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4973
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4974
    __ kernel_crc32(crc, buf, len, table);
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4975
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4976
    __ retl();
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4977
    __ delayed()->nop();
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4978
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4979
    return start;
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4980
  }
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  4981
47561
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4982
  /**
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4983
   * Arguments:
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4984
   *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4985
   * Inputs:
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4986
   *   I0   - int* x-addr
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4987
   *   I1   - int  x-len
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4988
   *   I2   - int* y-addr
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4989
   *   I3   - int  y-len
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4990
   *   I4   - int* z-addr   (output vector)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4991
   *   I5   - int  z-len
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4992
   */
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4993
  address generate_multiplyToLen() {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4994
    assert(UseMultiplyToLenIntrinsic, "need VIS3 instructions");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4995
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4996
    __ align(CodeEntryAlignment);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4997
    StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4998
    address start = __ pc();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  4999
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5000
    __ save_frame(0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5001
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5002
    const Register xptr = I0; // input address
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5003
    const Register xlen = I1; // ...and length in 32b-words
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5004
    const Register yptr = I2; //
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5005
    const Register ylen = I3; //
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5006
    const Register zptr = I4; // output address
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5007
    const Register zlen = I5; // ...and length in 32b-words
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5008
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5009
    /* The minimal "limb" representation suggest that odd length vectors are as
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5010
     * likely as even length dittos. This in turn suggests that we need to cope
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5011
     * with odd/even length arrays and data not aligned properly for 64-bit read
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5012
     * and write operations. We thus use a number of different kernels:
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5013
     *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5014
     *   if (is_even(x.len) && is_even(y.len))
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5015
     *      if (is_align64(x) && is_align64(y) && is_align64(z))
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5016
     *         if (x.len == y.len && 16 <= x.len && x.len <= 64)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5017
     *            memv_mult_mpmul(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5018
     *         else
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5019
     *            memv_mult_64x64(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5020
     *      else
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5021
     *         memv_mult_64x64u(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5022
     *   else
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5023
     *      memv_mult_32x32(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5024
     *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5025
     * Here we assume VIS3 support (for 'umulxhi', 'addxc' and 'addxccc').
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5026
     * In case CBCOND instructions are supported, we will use 'cxbX'. If the
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5027
     * MPMUL instruction is supported, we will generate a kernel using 'mpmul'
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5028
     * (for vectors with proper characteristics).
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5029
     */
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5030
    const Register tmp0 = L0;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5031
    const Register tmp1 = L1;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5032
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5033
    Label L_mult_32x32;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5034
    Label L_mult_64x64u;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5035
    Label L_mult_64x64;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5036
    Label L_exit;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5037
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5038
    if_both_even(xlen, ylen, tmp0, false, L_mult_32x32);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5039
    if_all3_aligned(xptr, yptr, zptr, tmp1, 64, false, L_mult_64x64u);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5040
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5041
    if (UseMPMUL) {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5042
      if_eq(xlen, ylen, false, L_mult_64x64);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5043
      if_in_rng(xlen, 16, 64, tmp0, tmp1, false, L_mult_64x64);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5044
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5045
      // 1. Multiply naturally aligned 64b-datums using a generic 'mpmul' kernel,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5046
      //    operating on equal length vectors of size [16..64].
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5047
      gen_mult_mpmul(xlen, xptr, yptr, zptr, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5048
    }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5049
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5050
    // 2. Multiply naturally aligned 64-bit datums (64x64).
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5051
    __ bind(L_mult_64x64);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5052
    gen_mult_64x64(xptr, xlen, yptr, ylen, zptr, zlen, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5053
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5054
    // 3. Multiply unaligned 64-bit datums (64x64).
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5055
    __ bind(L_mult_64x64u);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5056
    gen_mult_64x64_unaligned(xptr, xlen, yptr, ylen, zptr, zlen, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5057
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5058
    // 4. Multiply naturally aligned 32-bit datums (32x32).
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5059
    __ bind(L_mult_32x32);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5060
    gen_mult_32x32(xptr, xlen, yptr, ylen, zptr, zlen, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5061
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5062
    __ bind(L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5063
    __ ret();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5064
    __ delayed()->restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5065
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5066
    return start;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5067
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5068
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5069
  // Additional help functions used by multiplyToLen generation.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5070
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5071
  void if_both_even(Register r1, Register r2, Register tmp, bool iseven, Label &L)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5072
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5073
    __ or3(r1, r2, tmp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5074
    __ andcc(tmp, 0x1, tmp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5075
    __ br_icc_zero(iseven, Assembler::pn, L);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5076
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5077
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5078
  void if_all3_aligned(Register r1, Register r2, Register r3,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5079
                       Register tmp, uint align, bool isalign, Label &L)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5080
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5081
    __ or3(r1, r2, tmp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5082
    __ or3(r3, tmp, tmp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5083
    __ andcc(tmp, (align - 1), tmp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5084
    __ br_icc_zero(isalign, Assembler::pn, L);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5085
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5086
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5087
  void if_eq(Register x, Register y, bool iseq, Label &L)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5088
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5089
    Assembler::Condition cf = (iseq ? Assembler::equal : Assembler::notEqual);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5090
    __ cmp_and_br_short(x, y, cf, Assembler::pt, L);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5091
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5092
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5093
  void if_in_rng(Register x, int lb, int ub, Register t1, Register t2, bool inrng, Label &L)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5094
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5095
    assert(Assembler::is_simm13(lb), "Small ints only!");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5096
    assert(Assembler::is_simm13(ub), "Small ints only!");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5097
    // Compute (x - lb) * (ub - x) >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5098
    // NOTE: With the local use of this routine, we rely on small integers to
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5099
    //       guarantee that we do not overflow in the multiplication.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5100
    __ add(G0, ub, t2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5101
    __ sub(x, lb, t1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5102
    __ sub(t2, x, t2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5103
    __ mulx(t1, t2, t1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5104
    Assembler::Condition cf = (inrng ? Assembler::greaterEqual : Assembler::less);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5105
    __ cmp_and_br_short(t1, G0, cf, Assembler::pt, L);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5106
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5107
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5108
  void ldd_entry(Register base, Register offs, FloatRegister dest)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5109
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5110
    __ ldd(base, offs, dest);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5111
    __ inc(offs, 8);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5112
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5113
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5114
  void ldx_entry(Register base, Register offs, Register dest)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5115
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5116
    __ ldx(base, offs, dest);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5117
    __ inc(offs, 8);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5118
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5119
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5120
  void mpmul_entry(int m, Label &next)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5121
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5122
    __ mpmul(m);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5123
    __ cbcond(Assembler::equal, Assembler::icc, G0, G0, next);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5124
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5125
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5126
  void stx_entry(Label &L, Register r1, Register r2, Register base, Register offs)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5127
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5128
    __ bind(L);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5129
    __ stx(r1, base, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5130
    __ inc(offs, 8);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5131
    __ stx(r2, base, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5132
    __ inc(offs, 8);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5133
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5134
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5135
  void offs_entry(Label &Lbl0, Label &Lbl1)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5136
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5137
    assert(Lbl0.is_bound(), "must be");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5138
    assert(Lbl1.is_bound(), "must be");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5139
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5140
    int offset = Lbl0.loc_pos() - Lbl1.loc_pos();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5141
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5142
    __ emit_data(offset);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5143
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5144
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5145
  /* Generate the actual multiplication kernels for BigInteger vectors:
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5146
   *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5147
   *   1. gen_mult_mpmul(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5148
   *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5149
   *   2. gen_mult_64x64(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5150
   *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5151
   *   3. gen_mult_64x64_unaligned(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5152
   *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5153
   *   4. gen_mult_32x32(...)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5154
   */
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5155
  void gen_mult_mpmul(Register len, Register xptr, Register yptr, Register zptr,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5156
                      Label &L_exit)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5157
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5158
    const Register zero = G0;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5159
    const Register gxp  = G1;   // Need to use global registers across RWs.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5160
    const Register gyp  = G2;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5161
    const Register gzp  = G3;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5162
    const Register offs = G4;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5163
    const Register disp = G5;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5164
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5165
    __ mov(xptr, gxp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5166
    __ mov(yptr, gyp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5167
    __ mov(zptr, gzp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5168
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5169
    /* Compute jump vector entry:
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5170
     *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5171
     *   1. mpmul input size (0..31) x 64b
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5172
     *   2. vector input size in 32b limbs (even number)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5173
     *   3. branch entries in reverse order (31..0), using two
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5174
     *      instructions per entry (2 * 4 bytes).
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5175
     *
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5176
     *   displacement = byte_offset(bra_offset(len))
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5177
     *                = byte_offset((64 - len)/2)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5178
     *                = 8 * (64 - len)/2
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5179
     *                = 4 * (64 - len)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5180
     */
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5181
    Register temp = I5;         // Alright to use input regs. in first batch.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5182
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5183
    __ sub(zero, len, temp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5184
    __ add(temp, 64, temp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5185
    __ sllx(temp, 2, disp);     // disp := (64 - len) << 2
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5186
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5187
    // Dispatch relative current PC, into instruction table below.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5188
    __ rdpc(temp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5189
    __ add(temp, 16, temp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5190
    __ jmp(temp, disp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5191
    __ delayed()->clr(offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5192
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5193
    ldd_entry(gxp, offs, F22);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5194
    ldd_entry(gxp, offs, F20);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5195
    ldd_entry(gxp, offs, F18);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5196
    ldd_entry(gxp, offs, F16);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5197
    ldd_entry(gxp, offs, F14);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5198
    ldd_entry(gxp, offs, F12);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5199
    ldd_entry(gxp, offs, F10);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5200
    ldd_entry(gxp, offs, F8);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5201
    ldd_entry(gxp, offs, F6);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5202
    ldd_entry(gxp, offs, F4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5203
    ldx_entry(gxp, offs, I5);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5204
    ldx_entry(gxp, offs, I4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5205
    ldx_entry(gxp, offs, I3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5206
    ldx_entry(gxp, offs, I2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5207
    ldx_entry(gxp, offs, I1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5208
    ldx_entry(gxp, offs, I0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5209
    ldx_entry(gxp, offs, L7);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5210
    ldx_entry(gxp, offs, L6);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5211
    ldx_entry(gxp, offs, L5);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5212
    ldx_entry(gxp, offs, L4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5213
    ldx_entry(gxp, offs, L3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5214
    ldx_entry(gxp, offs, L2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5215
    ldx_entry(gxp, offs, L1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5216
    ldx_entry(gxp, offs, L0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5217
    ldd_entry(gxp, offs, F2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5218
    ldd_entry(gxp, offs, F0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5219
    ldx_entry(gxp, offs, O5);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5220
    ldx_entry(gxp, offs, O4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5221
    ldx_entry(gxp, offs, O3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5222
    ldx_entry(gxp, offs, O2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5223
    ldx_entry(gxp, offs, O1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5224
    ldx_entry(gxp, offs, O0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5225
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5226
    __ save(SP, -176, SP);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5227
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5228
    const Register addr = gxp;  // Alright to reuse 'gxp'.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5229
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5230
    // Dispatch relative current PC, into instruction table below.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5231
    __ rdpc(addr);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5232
    __ add(addr, 16, addr);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5233
    __ jmp(addr, disp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5234
    __ delayed()->clr(offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5235
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5236
    ldd_entry(gyp, offs, F58);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5237
    ldd_entry(gyp, offs, F56);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5238
    ldd_entry(gyp, offs, F54);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5239
    ldd_entry(gyp, offs, F52);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5240
    ldd_entry(gyp, offs, F50);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5241
    ldd_entry(gyp, offs, F48);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5242
    ldd_entry(gyp, offs, F46);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5243
    ldd_entry(gyp, offs, F44);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5244
    ldd_entry(gyp, offs, F42);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5245
    ldd_entry(gyp, offs, F40);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5246
    ldd_entry(gyp, offs, F38);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5247
    ldd_entry(gyp, offs, F36);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5248
    ldd_entry(gyp, offs, F34);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5249
    ldd_entry(gyp, offs, F32);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5250
    ldd_entry(gyp, offs, F30);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5251
    ldd_entry(gyp, offs, F28);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5252
    ldd_entry(gyp, offs, F26);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5253
    ldd_entry(gyp, offs, F24);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5254
    ldx_entry(gyp, offs, O5);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5255
    ldx_entry(gyp, offs, O4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5256
    ldx_entry(gyp, offs, O3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5257
    ldx_entry(gyp, offs, O2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5258
    ldx_entry(gyp, offs, O1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5259
    ldx_entry(gyp, offs, O0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5260
    ldx_entry(gyp, offs, L7);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5261
    ldx_entry(gyp, offs, L6);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5262
    ldx_entry(gyp, offs, L5);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5263
    ldx_entry(gyp, offs, L4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5264
    ldx_entry(gyp, offs, L3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5265
    ldx_entry(gyp, offs, L2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5266
    ldx_entry(gyp, offs, L1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5267
    ldx_entry(gyp, offs, L0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5268
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5269
    __ save(SP, -176, SP);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5270
    __ save(SP, -176, SP);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5271
    __ save(SP, -176, SP);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5272
    __ save(SP, -176, SP);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5273
    __ save(SP, -176, SP);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5274
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5275
    Label L_mpmul_restore_4, L_mpmul_restore_3, L_mpmul_restore_2;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5276
    Label L_mpmul_restore_1, L_mpmul_restore_0;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5277
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5278
    // Dispatch relative current PC, into instruction table below.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5279
    __ rdpc(addr);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5280
    __ add(addr, 16, addr);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5281
    __ jmp(addr, disp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5282
    __ delayed()->clr(offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5283
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5284
    mpmul_entry(31, L_mpmul_restore_0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5285
    mpmul_entry(30, L_mpmul_restore_0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5286
    mpmul_entry(29, L_mpmul_restore_0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5287
    mpmul_entry(28, L_mpmul_restore_0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5288
    mpmul_entry(27, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5289
    mpmul_entry(26, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5290
    mpmul_entry(25, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5291
    mpmul_entry(24, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5292
    mpmul_entry(23, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5293
    mpmul_entry(22, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5294
    mpmul_entry(21, L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5295
    mpmul_entry(20, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5296
    mpmul_entry(19, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5297
    mpmul_entry(18, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5298
    mpmul_entry(17, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5299
    mpmul_entry(16, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5300
    mpmul_entry(15, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5301
    mpmul_entry(14, L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5302
    mpmul_entry(13, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5303
    mpmul_entry(12, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5304
    mpmul_entry(11, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5305
    mpmul_entry(10, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5306
    mpmul_entry( 9, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5307
    mpmul_entry( 8, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5308
    mpmul_entry( 7, L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5309
    mpmul_entry( 6, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5310
    mpmul_entry( 5, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5311
    mpmul_entry( 4, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5312
    mpmul_entry( 3, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5313
    mpmul_entry( 2, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5314
    mpmul_entry( 1, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5315
    mpmul_entry( 0, L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5316
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5317
    Label L_z31, L_z30, L_z29, L_z28, L_z27, L_z26, L_z25, L_z24;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5318
    Label L_z23, L_z22, L_z21, L_z20, L_z19, L_z18, L_z17, L_z16;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5319
    Label L_z15, L_z14, L_z13, L_z12, L_z11, L_z10, L_z09, L_z08;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5320
    Label L_z07, L_z06, L_z05, L_z04, L_z03, L_z02, L_z01, L_z00;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5321
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5322
    Label L_zst_base;    // Store sequence base address.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5323
    __ bind(L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5324
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5325
    stx_entry(L_z31, L7, L6, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5326
    stx_entry(L_z30, L5, L4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5327
    stx_entry(L_z29, L3, L2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5328
    stx_entry(L_z28, L1, L0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5329
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5330
    stx_entry(L_z27, O5, O4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5331
    stx_entry(L_z26, O3, O2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5332
    stx_entry(L_z25, O1, O0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5333
    stx_entry(L_z24, L7, L6, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5334
    stx_entry(L_z23, L5, L4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5335
    stx_entry(L_z22, L3, L2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5336
    stx_entry(L_z21, L1, L0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5337
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5338
    stx_entry(L_z20, O5, O4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5339
    stx_entry(L_z19, O3, O2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5340
    stx_entry(L_z18, O1, O0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5341
    stx_entry(L_z17, L7, L6, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5342
    stx_entry(L_z16, L5, L4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5343
    stx_entry(L_z15, L3, L2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5344
    stx_entry(L_z14, L1, L0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5345
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5346
    stx_entry(L_z13, O5, O4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5347
    stx_entry(L_z12, O3, O2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5348
    stx_entry(L_z11, O1, O0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5349
    stx_entry(L_z10, L7, L6, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5350
    stx_entry(L_z09, L5, L4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5351
    stx_entry(L_z08, L3, L2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5352
    stx_entry(L_z07, L1, L0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5353
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5354
    stx_entry(L_z06, O5, O4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5355
    stx_entry(L_z05, O3, O2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5356
    stx_entry(L_z04, O1, O0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5357
    stx_entry(L_z03, L7, L6, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5358
    stx_entry(L_z02, L5, L4, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5359
    stx_entry(L_z01, L3, L2, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5360
    stx_entry(L_z00, L1, L0, gzp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5361
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5362
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5363
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5364
    // Exit out of 'mpmul' routine, back to multiplyToLen.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5365
    __ ba_short(L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5366
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5367
    Label L_zst_offs;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5368
    __ bind(L_zst_offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5369
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5370
    offs_entry(L_z31, L_zst_base);  // index 31: 2048x2048
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5371
    offs_entry(L_z30, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5372
    offs_entry(L_z29, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5373
    offs_entry(L_z28, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5374
    offs_entry(L_z27, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5375
    offs_entry(L_z26, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5376
    offs_entry(L_z25, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5377
    offs_entry(L_z24, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5378
    offs_entry(L_z23, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5379
    offs_entry(L_z22, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5380
    offs_entry(L_z21, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5381
    offs_entry(L_z20, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5382
    offs_entry(L_z19, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5383
    offs_entry(L_z18, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5384
    offs_entry(L_z17, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5385
    offs_entry(L_z16, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5386
    offs_entry(L_z15, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5387
    offs_entry(L_z14, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5388
    offs_entry(L_z13, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5389
    offs_entry(L_z12, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5390
    offs_entry(L_z11, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5391
    offs_entry(L_z10, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5392
    offs_entry(L_z09, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5393
    offs_entry(L_z08, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5394
    offs_entry(L_z07, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5395
    offs_entry(L_z06, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5396
    offs_entry(L_z05, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5397
    offs_entry(L_z04, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5398
    offs_entry(L_z03, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5399
    offs_entry(L_z02, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5400
    offs_entry(L_z01, L_zst_base);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5401
    offs_entry(L_z00, L_zst_base);  // index  0:   64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5402
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5403
    __ bind(L_mpmul_restore_4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5404
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5405
    __ bind(L_mpmul_restore_3);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5406
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5407
    __ bind(L_mpmul_restore_2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5408
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5409
    __ bind(L_mpmul_restore_1);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5410
    __ restore();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5411
    __ bind(L_mpmul_restore_0);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5412
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5413
    // Dispatch via offset vector entry, into z-store sequence.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5414
    Label L_zst_rdpc;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5415
    __ bind(L_zst_rdpc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5416
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5417
    assert(L_zst_base.is_bound(), "must be");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5418
    assert(L_zst_offs.is_bound(), "must be");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5419
    assert(L_zst_rdpc.is_bound(), "must be");
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5420
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5421
    int dbase = L_zst_rdpc.loc_pos() - L_zst_base.loc_pos();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5422
    int doffs = L_zst_rdpc.loc_pos() - L_zst_offs.loc_pos();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5423
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5424
    temp = gyp;   // Alright to reuse 'gyp'.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5425
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5426
    __ rdpc(addr);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5427
    __ sub(addr, doffs, temp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5428
    __ srlx(disp, 1, disp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5429
    __ lduw(temp, disp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5430
    __ sub(addr, dbase, temp);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5431
    __ jmp(temp, offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5432
    __ delayed()->clr(offs);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5433
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5434
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5435
  void gen_mult_64x64(Register xp, Register xn,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5436
                      Register yp, Register yn,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5437
                      Register zp, Register zn, Label &L_exit)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5438
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5439
    // Assuming that a stack frame has already been created, i.e. local and
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5440
    // output registers are available for immediate use.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5441
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5442
    const Register ri = L0;     // Outer loop index, xv[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5443
    const Register rj = L1;     // Inner loop index, yv[j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5444
    const Register rk = L2;     // Output loop index, zv[k]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5445
    const Register rx = L4;     // x-vector datum [i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5446
    const Register ry = L5;     // y-vector datum [j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5447
    const Register rz = L6;     // z-vector datum [k]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5448
    const Register rc = L7;     // carry over (to z-vector datum [k-1])
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5449
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5450
    const Register lop = O0;    // lo-64b product
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5451
    const Register hip = O1;    // hi-64b product
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5452
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5453
    const Register zero = G0;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5454
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5455
    Label L_loop_i,  L_exit_loop_i;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5456
    Label L_loop_j;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5457
    Label L_loop_i2, L_exit_loop_i2;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5458
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5459
    __ srlx(xn, 1, xn);         // index for u32 to u64 ditto
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5460
    __ srlx(yn, 1, yn);         // index for u32 to u64 ditto
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5461
    __ srlx(zn, 1, zn);         // index for u32 to u64 ditto
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5462
    __ dec(xn);                 // Adjust [0..(N/2)-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5463
    __ dec(yn);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5464
    __ dec(zn);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5465
    __ clr(rc);                 // u64 c = 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5466
    __ sllx(xn, 3, ri);         // int i = xn (byte offset i = 8*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5467
    __ sllx(yn, 3, rj);         // int j = yn (byte offset i = 8*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5468
    __ sllx(zn, 3, rk);         // int k = zn (byte offset k = 8*zn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5469
    __ ldx(yp, rj, ry);         // u64 y = yp[yn]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5470
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5471
    // for (int i = xn; i >= 0; i--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5472
    __ bind(L_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5473
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5474
    __ cmp_and_br_short(ri, 0,  // i >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5475
                        Assembler::less, Assembler::pn, L_exit_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5476
    __ ldx(xp, ri, rx);         // x = xp[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5477
    __ mulx(rx, ry, lop);       // lo-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5478
    __ umulxhi(rx, ry, hip);    // hi-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5479
    __ addcc(rc, lop, lop);     // Accumulate lower order bits (producing carry)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5480
    __ addxc(hip, zero, rc);    // carry over to next datum [k-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5481
    __ stx(lop, zp, rk);        // z[k] = lop
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5482
    __ dec(rk, 8);              // k--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5483
    __ dec(ri, 8);              // i--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5484
    __ ba_short(L_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5485
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5486
    __ bind(L_exit_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5487
    __ stx(rc, zp, rk);         // z[k] = c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5488
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5489
    // for (int j = yn - 1; j >= 0; j--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5490
    __ sllx(yn, 3, rj);         // int j = yn - 1 (byte offset j = 8*yn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5491
    __ dec(rj, 8);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5492
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5493
    __ bind(L_loop_j);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5494
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5495
    __ cmp_and_br_short(rj, 0,  // j >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5496
                        Assembler::less, Assembler::pn, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5497
    __ clr(rc);                 // u64 c = 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5498
    __ ldx(yp, rj, ry);         // u64 y = yp[j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5499
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5500
    // for (int i = xn, k = --zn; i >= 0; i--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5501
    __ dec(zn);                 // --zn
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5502
    __ sllx(xn, 3, ri);         // int i = xn (byte offset i = 8*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5503
    __ sllx(zn, 3, rk);         // int k = zn (byte offset k = 8*zn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5504
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5505
    __ bind(L_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5506
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5507
    __ cmp_and_br_short(ri, 0,  // i >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5508
                        Assembler::less, Assembler::pn, L_exit_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5509
    __ ldx(xp, ri, rx);         // x = xp[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5510
    __ ldx(zp, rk, rz);         // z = zp[k], accumulator
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5511
    __ mulx(rx, ry, lop);       // lo-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5512
    __ umulxhi(rx, ry, hip);    // hi-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5513
    __ addcc(rz, rc, rz);       // Accumulate lower order bits,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5514
    __ addxc(hip, zero, rc);    // Accumulate higher order bits to carry
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5515
    __ addcc(rz, lop, rz);      //    z += lo(p) + c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5516
    __ addxc(rc, zero, rc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5517
    __ stx(rz, zp, rk);         // zp[k] = z
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5518
    __ dec(rk, 8);              // k--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5519
    __ dec(ri, 8);              // i--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5520
    __ ba_short(L_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5521
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5522
    __ bind(L_exit_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5523
    __ stx(rc, zp, rk);         // z[k] = c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5524
    __ dec(rj, 8);              // j--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5525
    __ ba_short(L_loop_j);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5526
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5527
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5528
  void gen_mult_64x64_unaligned(Register xp, Register xn,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5529
                                Register yp, Register yn,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5530
                                Register zp, Register zn, Label &L_exit)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5531
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5532
    // Assuming that a stack frame has already been created, i.e. local and
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5533
    // output registers are available for use.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5534
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5535
    const Register xpc = L0;    // Outer loop cursor, xp[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5536
    const Register ypc = L1;    // Inner loop cursor, yp[j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5537
    const Register zpc = L2;    // Output loop cursor, zp[k]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5538
    const Register rx  = L4;    // x-vector datum [i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5539
    const Register ry  = L5;    // y-vector datum [j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5540
    const Register rz  = L6;    // z-vector datum [k]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5541
    const Register rc  = L7;    // carry over (to z-vector datum [k-1])
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5542
    const Register rt  = O2;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5543
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5544
    const Register lop = O0;    // lo-64b product
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5545
    const Register hip = O1;    // hi-64b product
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5546
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5547
    const Register zero = G0;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5548
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5549
    Label L_loop_i,  L_exit_loop_i;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5550
    Label L_loop_j;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5551
    Label L_loop_i2, L_exit_loop_i2;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5552
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5553
    __ srlx(xn, 1, xn);         // index for u32 to u64 ditto
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5554
    __ srlx(yn, 1, yn);         // index for u32 to u64 ditto
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5555
    __ srlx(zn, 1, zn);         // index for u32 to u64 ditto
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5556
    __ dec(xn);                 // Adjust [0..(N/2)-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5557
    __ dec(yn);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5558
    __ dec(zn);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5559
    __ clr(rc);                 // u64 c = 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5560
    __ sllx(xn, 3, xpc);        // u32* xpc = &xp[xn] (byte offset 8*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5561
    __ add(xp, xpc, xpc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5562
    __ sllx(yn, 3, ypc);        // u32* ypc = &yp[yn] (byte offset 8*yn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5563
    __ add(yp, ypc, ypc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5564
    __ sllx(zn, 3, zpc);        // u32* zpc = &zp[zn] (byte offset 8*zn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5565
    __ add(zp, zpc, zpc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5566
    __ lduw(ypc, 0, rt);        // u64 y = yp[yn]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5567
    __ lduw(ypc, 4, ry);        //   ...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5568
    __ sllx(rt, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5569
    __ or3(rt, ry, ry);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5570
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5571
    // for (int i = xn; i >= 0; i--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5572
    __ bind(L_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5573
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5574
    __ cmp_and_br_short(xpc, xp,// i >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5575
                        Assembler::less, Assembler::pn, L_exit_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5576
    __ lduw(xpc, 0, rt);        // u64 x = xp[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5577
    __ lduw(xpc, 4, rx);        //   ...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5578
    __ sllx(rt, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5579
    __ or3(rt, rx, rx);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5580
    __ mulx(rx, ry, lop);       // lo-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5581
    __ umulxhi(rx, ry, hip);    // hi-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5582
    __ addcc(rc, lop, lop);     // Accumulate lower order bits (producing carry)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5583
    __ addxc(hip, zero, rc);    // carry over to next datum [k-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5584
    __ srlx(lop, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5585
    __ stw(rt, zpc, 0);         // z[k] = lop
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5586
    __ stw(lop, zpc, 4);        //   ...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5587
    __ dec(zpc, 8);             // k-- (zpc--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5588
    __ dec(xpc, 8);             // i-- (xpc--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5589
    __ ba_short(L_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5590
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5591
    __ bind(L_exit_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5592
    __ srlx(rc, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5593
    __ stw(rt, zpc, 0);         // z[k] = c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5594
    __ stw(rc, zpc, 4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5595
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5596
    // for (int j = yn - 1; j >= 0; j--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5597
    __ sllx(yn, 3, ypc);        // u32* ypc = &yp[yn] (byte offset 8*yn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5598
    __ add(yp, ypc, ypc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5599
    __ dec(ypc, 8);             // yn - 1 (ypc--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5600
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5601
    __ bind(L_loop_j);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5602
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5603
    __ cmp_and_br_short(ypc, yp,// j >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5604
                        Assembler::less, Assembler::pn, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5605
    __ clr(rc);                 // u64 c = 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5606
    __ lduw(ypc, 0, rt);        // u64 y = yp[j] (= *ypc)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5607
    __ lduw(ypc, 4, ry);        //   ...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5608
    __ sllx(rt, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5609
    __ or3(rt, ry, ry);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5610
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5611
    // for (int i = xn, k = --zn; i >= 0; i--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5612
    __ sllx(xn, 3, xpc);        // u32* xpc = &xp[xn] (byte offset 8*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5613
    __ add(xp, xpc, xpc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5614
    __ dec(zn);                 // --zn
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5615
    __ sllx(zn, 3, zpc);        // u32* zpc = &zp[zn] (byte offset 8*zn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5616
    __ add(zp, zpc, zpc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5617
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5618
    __ bind(L_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5619
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5620
    __ cmp_and_br_short(xpc, xp,// i >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5621
                        Assembler::less, Assembler::pn, L_exit_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5622
    __ lduw(xpc, 0, rt);        // u64 x = xp[i] (= *xpc)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5623
    __ lduw(xpc, 4, rx);        //   ...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5624
    __ sllx(rt, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5625
    __ or3(rt, rx, rx);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5626
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5627
    __ lduw(zpc, 0, rt);        // u64 z = zp[k] (= *zpc)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5628
    __ lduw(zpc, 4, rz);        //   ...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5629
    __ sllx(rt, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5630
    __ or3(rt, rz, rz);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5631
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5632
    __ mulx(rx, ry, lop);       // lo-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5633
    __ umulxhi(rx, ry, hip);    // hi-64b-part of result 64x64
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5634
    __ addcc(rz, rc, rz);       // Accumulate lower order bits...
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5635
    __ addxc(hip, zero, rc);    // Accumulate higher order bits to carry
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5636
    __ addcc(rz, lop, rz);      // ... z += lo(p) + c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5637
    __ addxccc(rc, zero, rc);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5638
    __ srlx(rz, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5639
    __ stw(rt, zpc, 0);         // zp[k] = z    (*zpc = z)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5640
    __ stw(rz, zpc, 4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5641
    __ dec(zpc, 8);             // k-- (zpc--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5642
    __ dec(xpc, 8);             // i-- (xpc--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5643
    __ ba_short(L_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5644
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5645
    __ bind(L_exit_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5646
    __ srlx(rc, 32, rt);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5647
    __ stw(rt, zpc, 0);         // z[k] = c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5648
    __ stw(rc, zpc, 4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5649
    __ dec(ypc, 8);             // j-- (ypc--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5650
    __ ba_short(L_loop_j);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5651
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5652
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5653
  void gen_mult_32x32(Register xp, Register xn,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5654
                      Register yp, Register yn,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5655
                      Register zp, Register zn, Label &L_exit)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5656
  {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5657
    // Assuming that a stack frame has already been created, i.e. local and
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5658
    // output registers are available for use.
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5659
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5660
    const Register ri = L0;     // Outer loop index, xv[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5661
    const Register rj = L1;     // Inner loop index, yv[j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5662
    const Register rk = L2;     // Output loop index, zv[k]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5663
    const Register rx = L4;     // x-vector datum [i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5664
    const Register ry = L5;     // y-vector datum [j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5665
    const Register rz = L6;     // z-vector datum [k]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5666
    const Register rc = L7;     // carry over (to z-vector datum [k-1])
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5667
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5668
    const Register p64 = O0;    // 64b product
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5669
    const Register z65 = O1;    // carry+64b accumulator
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5670
    const Register c65 = O2;    // carry at bit 65
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5671
    const Register c33 = O2;    // carry at bit 33 (after shift)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5672
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5673
    const Register zero = G0;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5674
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5675
    Label L_loop_i,  L_exit_loop_i;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5676
    Label L_loop_j;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5677
    Label L_loop_i2, L_exit_loop_i2;
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5678
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5679
    __ dec(xn);                 // Adjust [0..N-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5680
    __ dec(yn);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5681
    __ dec(zn);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5682
    __ clr(rc);                 // u32 c = 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5683
    __ sllx(xn, 2, ri);         // int i = xn (byte offset i = 4*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5684
    __ sllx(yn, 2, rj);         // int j = yn (byte offset i = 4*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5685
    __ sllx(zn, 2, rk);         // int k = zn (byte offset k = 4*zn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5686
    __ lduw(yp, rj, ry);        // u32 y = yp[yn]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5687
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5688
    // for (int i = xn; i >= 0; i--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5689
    __ bind(L_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5690
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5691
    __ cmp_and_br_short(ri, 0,  // i >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5692
                        Assembler::less, Assembler::pn, L_exit_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5693
    __ lduw(xp, ri, rx);        // x = xp[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5694
    __ mulx(rx, ry, p64);       // 64b result of 32x32
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5695
    __ addcc(rc, p64, z65);     // Accumulate to 65 bits (producing carry)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5696
    __ addxc(zero, zero, c65);  // Materialise carry (in bit 65) into lsb,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5697
    __ sllx(c65, 32, c33);      // and shift into bit 33
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5698
    __ srlx(z65, 32, rc);       // carry = c33 | hi(z65) >> 32
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5699
    __ add(c33, rc, rc);        // carry over to next datum [k-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5700
    __ stw(z65, zp, rk);        // z[k] = lo(z65)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5701
    __ dec(rk, 4);              // k--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5702
    __ dec(ri, 4);              // i--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5703
    __ ba_short(L_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5704
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5705
    __ bind(L_exit_loop_i);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5706
    __ stw(rc, zp, rk);         // z[k] = c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5707
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5708
    // for (int j = yn - 1; j >= 0; j--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5709
    __ sllx(yn, 2, rj);         // int j = yn - 1 (byte offset j = 4*yn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5710
    __ dec(rj, 4);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5711
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5712
    __ bind(L_loop_j);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5713
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5714
    __ cmp_and_br_short(rj, 0,  // j >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5715
                        Assembler::less, Assembler::pn, L_exit);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5716
    __ clr(rc);                 // u32 c = 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5717
    __ lduw(yp, rj, ry);        // u32 y = yp[j]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5718
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5719
    // for (int i = xn, k = --zn; i >= 0; i--)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5720
    __ dec(zn);                 // --zn
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5721
    __ sllx(xn, 2, ri);         // int i = xn (byte offset i = 4*xn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5722
    __ sllx(zn, 2, rk);         // int k = zn (byte offset k = 4*zn)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5723
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5724
    __ bind(L_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5725
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5726
    __ cmp_and_br_short(ri, 0,  // i >= 0
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5727
                        Assembler::less, Assembler::pn, L_exit_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5728
    __ lduw(xp, ri, rx);        // x = xp[i]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5729
    __ lduw(zp, rk, rz);        // z = zp[k], accumulator
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5730
    __ mulx(rx, ry, p64);       // 64b result of 32x32
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5731
    __ add(rz, rc, rz);         // Accumulate lower order bits,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5732
    __ addcc(rz, p64, z65);     //   z += lo(p64) + c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5733
    __ addxc(zero, zero, c65);  // Materialise carry (in bit 65) into lsb,
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5734
    __ sllx(c65, 32, c33);      // and shift into bit 33
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5735
    __ srlx(z65, 32, rc);       // carry = c33 | hi(z65) >> 32
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5736
    __ add(c33, rc, rc);        // carry over to next datum [k-1]
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5737
    __ stw(z65, zp, rk);        // zp[k] = lo(z65)
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5738
    __ dec(rk, 4);              // k--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5739
    __ dec(ri, 4);              // i--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5740
    __ ba_short(L_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5741
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5742
    __ bind(L_exit_loop_i2);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5743
    __ stw(rc, zp, rk);         // z[k] = c
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5744
    __ dec(rj, 4);              // j--
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5745
    __ ba_short(L_loop_j);
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5746
  }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5747
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5748
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5749
  void generate_initial() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5750
    // Generates all stubs and initializes the entry points
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5751
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5752
    //------------------------------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5753
    // entry points that exist in all platforms
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5754
    // Note: This is code that could be shared among different platforms - however the benefit seems to be smaller than
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5755
    //       the disadvantage of having a much more complicated generator structure. See also comment in stubRoutines.hpp.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5756
    StubRoutines::_forward_exception_entry                 = generate_forward_exception();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5757
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5758
    StubRoutines::_call_stub_entry                         = generate_call_stub(StubRoutines::_call_stub_return_address);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5759
    StubRoutines::_catch_exception_entry                   = generate_catch_exception();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5760
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5761
    //------------------------------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5762
    // entry points that are platform specific
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5763
    StubRoutines::Sparc::_test_stop_entry                  = generate_test_stop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5764
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5765
    StubRoutines::Sparc::_stop_subroutine_entry            = generate_stop_subroutine();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5766
    StubRoutines::Sparc::_flush_callers_register_windows_entry = generate_flush_callers_register_windows();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5767
11411
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 10566
diff changeset
  5768
    // Build this early so it's available for the interpreter.
35071
a0910b1d3e0d 8046936: JEP 270: Reserved Stack Areas for Critical Sections
fparain
parents: 34205
diff changeset
  5769
    StubRoutines::_throw_StackOverflowError_entry =
a0910b1d3e0d 8046936: JEP 270: Reserved Stack Areas for Critical Sections
fparain
parents: 34205
diff changeset
  5770
            generate_throw_exception("StackOverflowError throw_exception",
a0910b1d3e0d 8046936: JEP 270: Reserved Stack Areas for Critical Sections
fparain
parents: 34205
diff changeset
  5771
            CAST_FROM_FN_PTR(address, SharedRuntime::throw_StackOverflowError));
a0910b1d3e0d 8046936: JEP 270: Reserved Stack Areas for Critical Sections
fparain
parents: 34205
diff changeset
  5772
    StubRoutines::_throw_delayed_StackOverflowError_entry =
a0910b1d3e0d 8046936: JEP 270: Reserved Stack Areas for Critical Sections
fparain
parents: 34205
diff changeset
  5773
            generate_throw_exception("delayed StackOverflowError throw_exception",
a0910b1d3e0d 8046936: JEP 270: Reserved Stack Areas for Critical Sections
fparain
parents: 34205
diff changeset
  5774
            CAST_FROM_FN_PTR(address, SharedRuntime::throw_delayed_StackOverflowError));
34205
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  5775
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  5776
    if (UseCRC32Intrinsics) {
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  5777
      // set table address before stub generation which use it
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  5778
      StubRoutines::_crc_table_adr = (address)StubRoutines::Sparc::_crc_table;
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  5779
      StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
9ec51d30a11e 8143012: CRC32 Intrinsics support on SPARC
kvn
parents: 32627
diff changeset
  5780
    }
38237
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  5781
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  5782
    if (UseCRC32CIntrinsics) {
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  5783
      // set table address before stub generation which use it
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  5784
      StubRoutines::_crc32c_table_addr = (address)StubRoutines::Sparc::_crc32c_table;
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  5785
      StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C();
d972e3a2df53 8155162: java.util.zip.CRC32C Interpreter/C1 intrinsics support on SPARC
kvn
parents: 37466
diff changeset
  5786
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5787
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5788
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5789
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5790
  void generate_all() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5791
    // Generates all stubs and initializes the entry points
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5792
2254
f13dda645a4b 6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents: 1502
diff changeset
  5793
    // Generate partial_subtype_check first here since its code depends on
f13dda645a4b 6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents: 1502
diff changeset
  5794
    // UseZeroBaseCompressedOops which is defined after heap initialization.
f13dda645a4b 6791178: Specialize for zero as the compressed oop vm heap base
kvn
parents: 1502
diff changeset
  5795
    StubRoutines::Sparc::_partial_subtype_check                = generate_partial_subtype_check();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5796
    // These entry points require SharedInfo::stack0 to be set up in non-core builds
10545
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10512
diff changeset
  5797
    StubRoutines::_throw_AbstractMethodError_entry         = generate_throw_exception("AbstractMethodError throw_exception",          CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError));
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10512
diff changeset
  5798
    StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError));
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10512
diff changeset
  5799
    StubRoutines::_throw_NullPointerException_at_call_entry= generate_throw_exception("NullPointerException at call throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_NullPointerException_at_call));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5800
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5801
    // support for verify_oop (must happen after universe_init)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5802
    StubRoutines::_verify_oop_subroutine_entry     = generate_verify_oop_subroutine();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5803
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5804
    // arraycopy stubs used by compilers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5805
    generate_arraycopy_stubs();
4645
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 2571
diff changeset
  5806
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 2571
diff changeset
  5807
    // Don't initialize the platform math functions since sparc
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 2571
diff changeset
  5808
    // doesn't have intrinsics for these operations.
18740
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5809
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5810
    // Safefetch stubs.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5811
    generate_safefetch("SafeFetch32", sizeof(int),     &StubRoutines::_safefetch32_entry,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5812
                                                       &StubRoutines::_safefetch32_fault_pc,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5813
                                                       &StubRoutines::_safefetch32_continuation_pc);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5814
    generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5815
                                                       &StubRoutines::_safefetchN_fault_pc,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18097
diff changeset
  5816
                                                       &StubRoutines::_safefetchN_continuation_pc);
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5817
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5818
    // generate AES intrinsics code
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5819
    if (UseAESIntrinsics) {
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5820
      StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5821
      StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5822
      StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5823
      StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 22234
diff changeset
  5824
    }
31404
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  5825
    // generate GHASH intrinsics code
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  5826
    if (UseGHASHIntrinsics) {
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  5827
      StubRoutines::_ghash_processBlocks = generate_ghash_processBlocks();
63e8fcd70bfc 8073108: Use x86 and SPARC CPU instructions for GHASH acceleration
ascarpino
parents: 29695
diff changeset
  5828
    }
24953
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5829
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5830
    // generate SHA1/SHA256/SHA512 intrinsics code
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5831
    if (UseSHA1Intrinsics) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5832
      StubRoutines::_sha1_implCompress     = generate_sha1_implCompress(false,   "sha1_implCompress");
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5833
      StubRoutines::_sha1_implCompressMB   = generate_sha1_implCompress(true,    "sha1_implCompressMB");
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5834
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5835
    if (UseSHA256Intrinsics) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5836
      StubRoutines::_sha256_implCompress   = generate_sha256_implCompress(false, "sha256_implCompress");
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5837
      StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(true,  "sha256_implCompressMB");
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5838
    }
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5839
    if (UseSHA512Intrinsics) {
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5840
      StubRoutines::_sha512_implCompress   = generate_sha512_implCompress(false, "sha512_implCompress");
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5841
      StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(true,  "sha512_implCompressMB");
9680119572be 8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
kvn
parents: 24488
diff changeset
  5842
    }
32581
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  5843
    // generate Adler32 intrinsics code
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  5844
    if (UseAdler32Intrinsics) {
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  5845
      StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
632402f18fe6 8132081: C2 support for Adler32 on SPARC
kvn
parents: 31515
diff changeset
  5846
    }
47561
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5847
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5848
#ifdef COMPILER2
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5849
    // Intrinsics supported by C2 only:
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5850
    if (UseMultiplyToLenIntrinsic) {
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5851
      StubRoutines::_multiplyToLen = generate_multiplyToLen();
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5852
    }
f59f0e51ef8a 8167199: Add C2 SPARC intrinsic for BigInteger::multiplyToLen() method.
neliasso
parents: 47216
diff changeset
  5853
#endif // COMPILER2
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5854
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5855
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5856
 public:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5857
  StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5858
    // replace the standard masm with a special one:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5859
    _masm = new MacroAssembler(code);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5860
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5861
    _stub_count = !all ? 0x100 : 0x200;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5862
    if (all) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5863
      generate_all();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5864
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5865
      generate_initial();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5866
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5867
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5868
    // make sure this stub is available for all local calls
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5869
    if (_atomic_add_stub.is_unbound()) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5870
      // generate a second time, if necessary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5871
      (void) generate_atomic_add();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5872
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5873
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5874
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5875
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5876
 private:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5877
  int _stub_count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5878
  void stub_prolog(StubCodeDesc* cdesc) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5879
    # ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5880
      // put extra information in the stub code, to make it more readable
46462
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  5881
      // Write the high part of the address
f92a713126b1 8179903: Clean up SPARC 32-bit support
gtriantafill
parents: 46422
diff changeset
  5882
      // [RGV] Check if there is a dependency on the size of this prolog
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5883
      __ emit_data((intptr_t)cdesc >> 32,    relocInfo::none);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5884
      __ emit_data((intptr_t)cdesc,    relocInfo::none);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5885
      __ emit_data(++_stub_count, relocInfo::none);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5886
    # endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5887
    align(true);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5888
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5889
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5890
  void align(bool at_header = false) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5891
    // %%%%% move this constant somewhere else
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5892
    // UltraSPARC cache line size is 8 instructions:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5893
    const unsigned int icache_line_size = 32;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5894
    const unsigned int icache_half_line_size = 16;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5895
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5896
    if (at_header) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5897
      while ((intptr_t)(__ pc()) % icache_line_size != 0) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5898
        __ emit_data(0, relocInfo::none);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5899
      }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5900
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5901
      while ((intptr_t)(__ pc()) % icache_half_line_size != 0) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5902
        __ nop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5903
      }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5904
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5905
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5906
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5907
}; // end class declaration
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5908
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5909
void StubGenerator_generate(CodeBuffer* code, bool all) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5910
  StubGenerator g(code, all);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  5911
}