hotspot/src/cpu/x86/vm/stubGenerator_x86_64.cpp
author duke
Wed, 05 Jul 2017 20:10:48 +0200
changeset 27963 88d7c7c376e9
parent 27691 733f189ad1f7
child 29325 0e86e64c66e5
permissions -rw-r--r--
Merge
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
     1
/*
18507
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
     2
 * Copyright (c) 2003, 2013, 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: 6433
diff changeset
    25
#include "precompiled.hpp"
14626
0cf4eccf130f 8003240: x86: move MacroAssembler into separate file
twisti
parents: 14132
diff changeset
    26
#include "asm/macroAssembler.hpp"
0cf4eccf130f 8003240: x86: move MacroAssembler into separate file
twisti
parents: 14132
diff changeset
    27
#include "asm/macroAssembler.inline.hpp"
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    28
#include "interpreter/interpreter.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    29
#include "nativeInst_x86.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    30
#include "oops/instanceOop.hpp"
13728
882756847a04 6964458: Reimplement class meta-data storage to use native memory
coleenp
parents: 13391
diff changeset
    31
#include "oops/method.hpp"
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    32
#include "oops/objArrayKlass.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    33
#include "oops/oop.inline.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    34
#include "prims/methodHandles.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    35
#include "runtime/frame.inline.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    36
#include "runtime/handles.inline.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    37
#include "runtime/sharedRuntime.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    38
#include "runtime/stubCodeGenerator.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    39
#include "runtime/stubRoutines.hpp"
14583
d70ee55535f4 8003935: Simplify the needed includes for using Thread::current()
stefank
parents: 14132
diff changeset
    40
#include "runtime/thread.inline.hpp"
7397
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    41
#include "utilities/top.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    42
#ifdef COMPILER2
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    43
#include "opto/runtime.hpp"
5b173b4ca846 6989984: Use standard include model for Hospot
stefank
parents: 6433
diff changeset
    44
#endif
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    45
489c9b5090e2 Initial load
duke
parents:
diff changeset
    46
// Declaration and definition of StubGenerator (no .hpp file).
489c9b5090e2 Initial load
duke
parents:
diff changeset
    47
// For a more detailed description of the stub routine structure
489c9b5090e2 Initial load
duke
parents:
diff changeset
    48
// see the comment in stubRoutines.hpp
489c9b5090e2 Initial load
duke
parents:
diff changeset
    49
489c9b5090e2 Initial load
duke
parents:
diff changeset
    50
#define __ _masm->
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
    51
#define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
    52
#define a__ ((Assembler*)_masm)->
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    53
489c9b5090e2 Initial load
duke
parents:
diff changeset
    54
#ifdef PRODUCT
489c9b5090e2 Initial load
duke
parents:
diff changeset
    55
#define BLOCK_COMMENT(str) /* nothing */
489c9b5090e2 Initial load
duke
parents:
diff changeset
    56
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
    57
#define BLOCK_COMMENT(str) __ block_comment(str)
489c9b5090e2 Initial load
duke
parents:
diff changeset
    58
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
    59
489c9b5090e2 Initial load
duke
parents:
diff changeset
    60
#define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
489c9b5090e2 Initial load
duke
parents:
diff changeset
    61
const int MXCSR_MASK = 0xFFC0;  // Mask out any pending exceptions
489c9b5090e2 Initial load
duke
parents:
diff changeset
    62
489c9b5090e2 Initial load
duke
parents:
diff changeset
    63
// Stub Code definitions
489c9b5090e2 Initial load
duke
parents:
diff changeset
    64
489c9b5090e2 Initial load
duke
parents:
diff changeset
    65
static address handle_unsafe_access() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
    66
  JavaThread* thread = JavaThread::current();
489c9b5090e2 Initial load
duke
parents:
diff changeset
    67
  address pc = thread->saved_exception_pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
    68
  // pc is the instruction which we must emulate
489c9b5090e2 Initial load
duke
parents:
diff changeset
    69
  // doing a no-op is fine:  return garbage from the load
489c9b5090e2 Initial load
duke
parents:
diff changeset
    70
  // therefore, compute npc
489c9b5090e2 Initial load
duke
parents:
diff changeset
    71
  address npc = Assembler::locate_next_instruction(pc);
489c9b5090e2 Initial load
duke
parents:
diff changeset
    72
489c9b5090e2 Initial load
duke
parents:
diff changeset
    73
  // request an async exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
    74
  thread->set_pending_unsafe_access_error();
489c9b5090e2 Initial load
duke
parents:
diff changeset
    75
489c9b5090e2 Initial load
duke
parents:
diff changeset
    76
  // return address of next instruction to execute
489c9b5090e2 Initial load
duke
parents:
diff changeset
    77
  return npc;
489c9b5090e2 Initial load
duke
parents:
diff changeset
    78
}
489c9b5090e2 Initial load
duke
parents:
diff changeset
    79
489c9b5090e2 Initial load
duke
parents:
diff changeset
    80
class StubGenerator: public StubCodeGenerator {
489c9b5090e2 Initial load
duke
parents:
diff changeset
    81
 private:
489c9b5090e2 Initial load
duke
parents:
diff changeset
    82
489c9b5090e2 Initial load
duke
parents:
diff changeset
    83
#ifdef PRODUCT
18073
f02460441ddc 8014431: cleanup warnings indicated by the -Wunused-value compiler option on linux
ccheung
parents: 17622
diff changeset
    84
#define inc_counter_np(counter) ((void)0)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    85
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
    86
  void inc_counter_np_(int& counter) {
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
    87
    // This can destroy rscratch1 if counter is far from the code cache
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
    88
    __ incrementl(ExternalAddress((address)&counter));
489c9b5090e2 Initial load
duke
parents:
diff changeset
    89
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
    90
#define inc_counter_np(counter) \
489c9b5090e2 Initial load
duke
parents:
diff changeset
    91
  BLOCK_COMMENT("inc_counter " #counter); \
489c9b5090e2 Initial load
duke
parents:
diff changeset
    92
  inc_counter_np_(counter);
489c9b5090e2 Initial load
duke
parents:
diff changeset
    93
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
    94
489c9b5090e2 Initial load
duke
parents:
diff changeset
    95
  // Call stubs are used to call Java from C
489c9b5090e2 Initial load
duke
parents:
diff changeset
    96
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
    97
  // Linux Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
    98
  //    c_rarg0:   call wrapper address                   address
489c9b5090e2 Initial load
duke
parents:
diff changeset
    99
  //    c_rarg1:   result                                 address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   100
  //    c_rarg2:   result type                            BasicType
13728
882756847a04 6964458: Reimplement class meta-data storage to use native memory
coleenp
parents: 13391
diff changeset
   101
  //    c_rarg3:   method                                 Method*
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   102
  //    c_rarg4:   (interpreter) entry point              address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   103
  //    c_rarg5:   parameters                             intptr_t*
489c9b5090e2 Initial load
duke
parents:
diff changeset
   104
  //    16(rbp): parameter size (in words)              int
489c9b5090e2 Initial load
duke
parents:
diff changeset
   105
  //    24(rbp): thread                                 Thread*
489c9b5090e2 Initial load
duke
parents:
diff changeset
   106
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   107
  //     [ return_from_Java     ] <--- rsp
489c9b5090e2 Initial load
duke
parents:
diff changeset
   108
  //     [ argument word n      ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   109
  //      ...
489c9b5090e2 Initial load
duke
parents:
diff changeset
   110
  // -12 [ argument word 1      ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   111
  // -11 [ saved r15            ] <--- rsp_after_call
489c9b5090e2 Initial load
duke
parents:
diff changeset
   112
  // -10 [ saved r14            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   113
  //  -9 [ saved r13            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   114
  //  -8 [ saved r12            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   115
  //  -7 [ saved rbx            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   116
  //  -6 [ call wrapper         ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   117
  //  -5 [ result               ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   118
  //  -4 [ result type          ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   119
  //  -3 [ method               ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   120
  //  -2 [ entry point          ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   121
  //  -1 [ parameters           ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   122
  //   0 [ saved rbp            ] <--- rbp
489c9b5090e2 Initial load
duke
parents:
diff changeset
   123
  //   1 [ return address       ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   124
  //   2 [ parameter size       ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   125
  //   3 [ thread               ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   126
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   127
  // Windows Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   128
  //    c_rarg0:   call wrapper address                   address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   129
  //    c_rarg1:   result                                 address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   130
  //    c_rarg2:   result type                            BasicType
13728
882756847a04 6964458: Reimplement class meta-data storage to use native memory
coleenp
parents: 13391
diff changeset
   131
  //    c_rarg3:   method                                 Method*
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   132
  //    48(rbp): (interpreter) entry point              address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   133
  //    56(rbp): parameters                             intptr_t*
489c9b5090e2 Initial load
duke
parents:
diff changeset
   134
  //    64(rbp): parameter size (in words)              int
489c9b5090e2 Initial load
duke
parents:
diff changeset
   135
  //    72(rbp): thread                                 Thread*
489c9b5090e2 Initial load
duke
parents:
diff changeset
   136
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   137
  //     [ return_from_Java     ] <--- rsp
489c9b5090e2 Initial load
duke
parents:
diff changeset
   138
  //     [ argument word n      ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   139
  //      ...
8874
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   140
  // -28 [ argument word 1      ]
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   141
  // -27 [ saved xmm15          ] <--- rsp_after_call
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   142
  //     [ saved xmm7-xmm14     ]
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   143
  //  -9 [ saved xmm6           ] (each xmm register takes 2 slots)
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   144
  //  -7 [ saved r15            ]
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   145
  //  -6 [ saved r14            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   146
  //  -5 [ saved r13            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   147
  //  -4 [ saved r12            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   148
  //  -3 [ saved rdi            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   149
  //  -2 [ saved rsi            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   150
  //  -1 [ saved rbx            ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   151
  //   0 [ saved rbp            ] <--- rbp
489c9b5090e2 Initial load
duke
parents:
diff changeset
   152
  //   1 [ return address       ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   153
  //   2 [ call wrapper         ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   154
  //   3 [ result               ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   155
  //   4 [ result type          ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   156
  //   5 [ method               ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   157
  //   6 [ entry point          ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   158
  //   7 [ parameters           ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   159
  //   8 [ parameter size       ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   160
  //   9 [ thread               ]
489c9b5090e2 Initial load
duke
parents:
diff changeset
   161
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   162
  //    Windows reserves the callers stack space for arguments 1-4.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   163
  //    We spill c_rarg0-c_rarg3 to this space.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   164
489c9b5090e2 Initial load
duke
parents:
diff changeset
   165
  // Call stub stack layout word offsets from rbp
489c9b5090e2 Initial load
duke
parents:
diff changeset
   166
  enum call_stub_layout {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   167
#ifdef _WIN64
8874
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   168
    xmm_save_first     = 6,  // save from xmm6
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   169
    xmm_save_last      = 15, // to xmm15
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   170
    xmm_save_base      = -9,
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   171
    rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   172
    r15_off            = -7,
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   173
    r14_off            = -6,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   174
    r13_off            = -5,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   175
    r12_off            = -4,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   176
    rdi_off            = -3,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   177
    rsi_off            = -2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   178
    rbx_off            = -1,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   179
    rbp_off            =  0,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   180
    retaddr_off        =  1,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   181
    call_wrapper_off   =  2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   182
    result_off         =  3,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   183
    result_type_off    =  4,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   184
    method_off         =  5,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   185
    entry_point_off    =  6,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   186
    parameters_off     =  7,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   187
    parameter_size_off =  8,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   188
    thread_off         =  9
489c9b5090e2 Initial load
duke
parents:
diff changeset
   189
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   190
    rsp_after_call_off = -12,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   191
    mxcsr_off          = rsp_after_call_off,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   192
    r15_off            = -11,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   193
    r14_off            = -10,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   194
    r13_off            = -9,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   195
    r12_off            = -8,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   196
    rbx_off            = -7,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   197
    call_wrapper_off   = -6,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   198
    result_off         = -5,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   199
    result_type_off    = -4,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   200
    method_off         = -3,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   201
    entry_point_off    = -2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   202
    parameters_off     = -1,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   203
    rbp_off            =  0,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   204
    retaddr_off        =  1,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   205
    parameter_size_off =  2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   206
    thread_off         =  3
489c9b5090e2 Initial load
duke
parents:
diff changeset
   207
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   208
  };
489c9b5090e2 Initial load
duke
parents:
diff changeset
   209
8874
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   210
#ifdef _WIN64
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   211
  Address xmm_save(int reg) {
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   212
    assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   213
    return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   214
  }
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   215
#endif
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   216
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   217
  address generate_call_stub(address& return_address) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   218
    assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
489c9b5090e2 Initial load
duke
parents:
diff changeset
   219
           (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   220
           "adjust this code");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   221
    StubCodeMark mark(this, "StubRoutines", "call_stub");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   222
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   223
489c9b5090e2 Initial load
duke
parents:
diff changeset
   224
    // same as in generate_catch_exception()!
489c9b5090e2 Initial load
duke
parents:
diff changeset
   225
    const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   226
489c9b5090e2 Initial load
duke
parents:
diff changeset
   227
    const Address call_wrapper  (rbp, call_wrapper_off   * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   228
    const Address result        (rbp, result_off         * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   229
    const Address result_type   (rbp, result_type_off    * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   230
    const Address method        (rbp, method_off         * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   231
    const Address entry_point   (rbp, entry_point_off    * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   232
    const Address parameters    (rbp, parameters_off     * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   233
    const Address parameter_size(rbp, parameter_size_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   234
489c9b5090e2 Initial load
duke
parents:
diff changeset
   235
    // same as in generate_catch_exception()!
489c9b5090e2 Initial load
duke
parents:
diff changeset
   236
    const Address thread        (rbp, thread_off         * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   237
489c9b5090e2 Initial load
duke
parents:
diff changeset
   238
    const Address r15_save(rbp, r15_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   239
    const Address r14_save(rbp, r14_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   240
    const Address r13_save(rbp, r13_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   241
    const Address r12_save(rbp, r12_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   242
    const Address rbx_save(rbp, rbx_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   243
489c9b5090e2 Initial load
duke
parents:
diff changeset
   244
    // stub code
489c9b5090e2 Initial load
duke
parents:
diff changeset
   245
    __ enter();
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   246
    __ subptr(rsp, -rsp_after_call_off * wordSize);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   247
489c9b5090e2 Initial load
duke
parents:
diff changeset
   248
    // save register parameters
489c9b5090e2 Initial load
duke
parents:
diff changeset
   249
#ifndef _WIN64
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   250
    __ movptr(parameters,   c_rarg5); // parameters
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   251
    __ movptr(entry_point,  c_rarg4); // entry_point
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   252
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   253
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   254
    __ movptr(method,       c_rarg3); // method
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   255
    __ movl(result_type,  c_rarg2);   // result type
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   256
    __ movptr(result,       c_rarg1); // result
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   257
    __ movptr(call_wrapper, c_rarg0); // call wrapper
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   258
489c9b5090e2 Initial load
duke
parents:
diff changeset
   259
    // save regs belonging to calling function
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   260
    __ movptr(rbx_save, rbx);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   261
    __ movptr(r12_save, r12);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   262
    __ movptr(r13_save, r13);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   263
    __ movptr(r14_save, r14);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   264
    __ movptr(r15_save, r15);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   265
#ifdef _WIN64
8874
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   266
    for (int i = 6; i <= 15; i++) {
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   267
      __ movdqu(xmm_save(i), as_XMMRegister(i));
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   268
    }
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   269
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   270
    const Address rdi_save(rbp, rdi_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   271
    const Address rsi_save(rbp, rsi_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   272
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   273
    __ movptr(rsi_save, rsi);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   274
    __ movptr(rdi_save, rdi);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   275
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   276
    const Address mxcsr_save(rbp, mxcsr_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   277
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   278
      Label skip_ldmx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   279
      __ stmxcsr(mxcsr_save);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   280
      __ movl(rax, mxcsr_save);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   281
      __ andl(rax, MXCSR_MASK);    // Only check control and mask bits
18958
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
   282
      ExternalAddress mxcsr_std(StubRoutines::addr_mxcsr_std());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   283
      __ cmp32(rax, mxcsr_std);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   284
      __ jcc(Assembler::equal, skip_ldmx);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   285
      __ ldmxcsr(mxcsr_std);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   286
      __ bind(skip_ldmx);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   287
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   288
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   289
489c9b5090e2 Initial load
duke
parents:
diff changeset
   290
    // Load up thread register
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   291
    __ movptr(r15_thread, thread);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
   292
    __ reinit_heapbase();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   293
489c9b5090e2 Initial load
duke
parents:
diff changeset
   294
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   295
    // make sure we have no pending exceptions
489c9b5090e2 Initial load
duke
parents:
diff changeset
   296
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   297
      Label L;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   298
      __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   299
      __ jcc(Assembler::equal, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   300
      __ stop("StubRoutines::call_stub: entered with pending exception");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   301
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   302
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   303
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   304
489c9b5090e2 Initial load
duke
parents:
diff changeset
   305
    // pass parameters if any
489c9b5090e2 Initial load
duke
parents:
diff changeset
   306
    BLOCK_COMMENT("pass parameters if any");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   307
    Label parameters_done;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   308
    __ movl(c_rarg3, parameter_size);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   309
    __ testl(c_rarg3, c_rarg3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   310
    __ jcc(Assembler::zero, parameters_done);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   311
489c9b5090e2 Initial load
duke
parents:
diff changeset
   312
    Label loop;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   313
    __ movptr(c_rarg2, parameters);       // parameter pointer
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   314
    __ movl(c_rarg1, c_rarg3);            // parameter counter is in c_rarg1
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   315
    __ BIND(loop);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   316
    __ movptr(rax, Address(c_rarg2, 0));// get parameter
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   317
    __ addptr(c_rarg2, wordSize);       // advance to next parameter
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   318
    __ decrementl(c_rarg1);             // decrement counter
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   319
    __ push(rax);                       // pass parameter
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   320
    __ jcc(Assembler::notZero, loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   321
489c9b5090e2 Initial load
duke
parents:
diff changeset
   322
    // call Java function
489c9b5090e2 Initial load
duke
parents:
diff changeset
   323
    __ BIND(parameters_done);
13728
882756847a04 6964458: Reimplement class meta-data storage to use native memory
coleenp
parents: 13391
diff changeset
   324
    __ movptr(rbx, method);             // get Method*
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   325
    __ movptr(c_rarg1, entry_point);    // get entry_point
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   326
    __ mov(r13, rsp);                   // set sender sp
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   327
    BLOCK_COMMENT("call Java function");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   328
    __ call(c_rarg1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   329
489c9b5090e2 Initial load
duke
parents:
diff changeset
   330
    BLOCK_COMMENT("call_stub_return_address:");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   331
    return_address = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   332
489c9b5090e2 Initial load
duke
parents:
diff changeset
   333
    // store result depending on type (everything that is not
489c9b5090e2 Initial load
duke
parents:
diff changeset
   334
    // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   335
    __ movptr(c_rarg0, result);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   336
    Label is_long, is_float, is_double, exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   337
    __ movl(c_rarg1, result_type);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   338
    __ cmpl(c_rarg1, T_OBJECT);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   339
    __ jcc(Assembler::equal, is_long);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   340
    __ cmpl(c_rarg1, T_LONG);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   341
    __ jcc(Assembler::equal, is_long);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   342
    __ cmpl(c_rarg1, T_FLOAT);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   343
    __ jcc(Assembler::equal, is_float);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   344
    __ cmpl(c_rarg1, T_DOUBLE);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   345
    __ jcc(Assembler::equal, is_double);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   346
489c9b5090e2 Initial load
duke
parents:
diff changeset
   347
    // handle T_INT case
489c9b5090e2 Initial load
duke
parents:
diff changeset
   348
    __ movl(Address(c_rarg0, 0), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   349
489c9b5090e2 Initial load
duke
parents:
diff changeset
   350
    __ BIND(exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   351
489c9b5090e2 Initial load
duke
parents:
diff changeset
   352
    // pop parameters
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   353
    __ lea(rsp, rsp_after_call);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   354
489c9b5090e2 Initial load
duke
parents:
diff changeset
   355
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   356
    // verify that threads correspond
489c9b5090e2 Initial load
duke
parents:
diff changeset
   357
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   358
      Label L, S;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   359
      __ cmpptr(r15_thread, thread);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   360
      __ jcc(Assembler::notEqual, S);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   361
      __ get_thread(rbx);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   362
      __ cmpptr(r15_thread, rbx);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   363
      __ jcc(Assembler::equal, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   364
      __ bind(S);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   365
      __ jcc(Assembler::equal, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   366
      __ stop("StubRoutines::call_stub: threads must correspond");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   367
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   368
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   369
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   370
489c9b5090e2 Initial load
duke
parents:
diff changeset
   371
    // restore regs belonging to calling function
8874
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   372
#ifdef _WIN64
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   373
    for (int i = 15; i >= 6; i--) {
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   374
      __ movdqu(as_XMMRegister(i), xmm_save(i));
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   375
    }
b2030880129c 6741940: Nonvolatile XMM registers not preserved across JNI calls
iveresov
parents: 8498
diff changeset
   376
#endif
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   377
    __ movptr(r15, r15_save);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   378
    __ movptr(r14, r14_save);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   379
    __ movptr(r13, r13_save);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   380
    __ movptr(r12, r12_save);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   381
    __ movptr(rbx, rbx_save);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   382
489c9b5090e2 Initial load
duke
parents:
diff changeset
   383
#ifdef _WIN64
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   384
    __ movptr(rdi, rdi_save);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   385
    __ movptr(rsi, rsi_save);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   386
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   387
    __ ldmxcsr(mxcsr_save);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   388
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   389
489c9b5090e2 Initial load
duke
parents:
diff changeset
   390
    // restore rsp
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   391
    __ addptr(rsp, -rsp_after_call_off * wordSize);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   392
489c9b5090e2 Initial load
duke
parents:
diff changeset
   393
    // return
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   394
    __ pop(rbp);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   395
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   396
489c9b5090e2 Initial load
duke
parents:
diff changeset
   397
    // handle return types different from T_INT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   398
    __ BIND(is_long);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   399
    __ movq(Address(c_rarg0, 0), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   400
    __ jmp(exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   401
489c9b5090e2 Initial load
duke
parents:
diff changeset
   402
    __ BIND(is_float);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   403
    __ movflt(Address(c_rarg0, 0), xmm0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   404
    __ jmp(exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   405
489c9b5090e2 Initial load
duke
parents:
diff changeset
   406
    __ BIND(is_double);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   407
    __ movdbl(Address(c_rarg0, 0), xmm0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   408
    __ jmp(exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   409
489c9b5090e2 Initial load
duke
parents:
diff changeset
   410
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   411
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   412
489c9b5090e2 Initial load
duke
parents:
diff changeset
   413
  // Return point for a Java call if there's an exception thrown in
489c9b5090e2 Initial load
duke
parents:
diff changeset
   414
  // Java code.  The exception is caught and transformed into a
489c9b5090e2 Initial load
duke
parents:
diff changeset
   415
  // pending exception stored in JavaThread that can be tested from
489c9b5090e2 Initial load
duke
parents:
diff changeset
   416
  // within the VM.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   417
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   418
  // Note: Usually the parameters are removed by the callee. In case
489c9b5090e2 Initial load
duke
parents:
diff changeset
   419
  // of an exception crossing an activation frame boundary, that is
489c9b5090e2 Initial load
duke
parents:
diff changeset
   420
  // not the case if the callee is compiled code => need to setup the
489c9b5090e2 Initial load
duke
parents:
diff changeset
   421
  // rsp.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   422
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   423
  // rax: exception oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
   424
489c9b5090e2 Initial load
duke
parents:
diff changeset
   425
  address generate_catch_exception() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   426
    StubCodeMark mark(this, "StubRoutines", "catch_exception");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   427
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   428
489c9b5090e2 Initial load
duke
parents:
diff changeset
   429
    // same as in generate_call_stub():
489c9b5090e2 Initial load
duke
parents:
diff changeset
   430
    const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   431
    const Address thread        (rbp, thread_off         * wordSize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   432
489c9b5090e2 Initial load
duke
parents:
diff changeset
   433
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   434
    // verify that threads correspond
489c9b5090e2 Initial load
duke
parents:
diff changeset
   435
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   436
      Label L, S;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   437
      __ cmpptr(r15_thread, thread);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   438
      __ jcc(Assembler::notEqual, S);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   439
      __ get_thread(rbx);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   440
      __ cmpptr(r15_thread, rbx);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   441
      __ jcc(Assembler::equal, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   442
      __ bind(S);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   443
      __ stop("StubRoutines::catch_exception: threads must correspond");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   444
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   445
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   446
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   447
489c9b5090e2 Initial load
duke
parents:
diff changeset
   448
    // set pending exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   449
    __ verify_oop(rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   450
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   451
    __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   452
    __ lea(rscratch1, ExternalAddress((address)__FILE__));
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   453
    __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   454
    __ movl(Address(r15_thread, Thread::exception_line_offset()), (int)  __LINE__);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   455
489c9b5090e2 Initial load
duke
parents:
diff changeset
   456
    // complete return to VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
   457
    assert(StubRoutines::_call_stub_return_address != NULL,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   458
           "_call_stub_return_address must have been generated before");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   459
    __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   460
489c9b5090e2 Initial load
duke
parents:
diff changeset
   461
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   462
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   463
489c9b5090e2 Initial load
duke
parents:
diff changeset
   464
  // Continuation point for runtime calls returning with a pending
489c9b5090e2 Initial load
duke
parents:
diff changeset
   465
  // exception.  The pending exception check happened in the runtime
489c9b5090e2 Initial load
duke
parents:
diff changeset
   466
  // or native call stub.  The pending exception in Thread is
489c9b5090e2 Initial load
duke
parents:
diff changeset
   467
  // converted into a Java-level exception.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   468
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   469
  // Contract with Java-level exception handlers:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   470
  // rax: exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   471
  // rdx: throwing pc
489c9b5090e2 Initial load
duke
parents:
diff changeset
   472
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   473
  // NOTE: At entry of this stub, exception-pc must be on stack !!
489c9b5090e2 Initial load
duke
parents:
diff changeset
   474
489c9b5090e2 Initial load
duke
parents:
diff changeset
   475
  address generate_forward_exception() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   476
    StubCodeMark mark(this, "StubRoutines", "forward exception");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   477
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   478
489c9b5090e2 Initial load
duke
parents:
diff changeset
   479
    // Upon entry, the sp points to the return address returning into
489c9b5090e2 Initial load
duke
parents:
diff changeset
   480
    // Java (interpreted or compiled) code; i.e., the return address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   481
    // becomes the throwing pc.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   482
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   483
    // Arguments pushed before the runtime call are still on the stack
489c9b5090e2 Initial load
duke
parents:
diff changeset
   484
    // but the exception handler will reset the stack pointer ->
489c9b5090e2 Initial load
duke
parents:
diff changeset
   485
    // ignore them.  A potential result in registers can be ignored as
489c9b5090e2 Initial load
duke
parents:
diff changeset
   486
    // well.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   487
489c9b5090e2 Initial load
duke
parents:
diff changeset
   488
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   489
    // make sure this code is only executed if there is a pending exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   490
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   491
      Label L;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   492
      __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t) NULL);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   493
      __ jcc(Assembler::notEqual, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   494
      __ stop("StubRoutines::forward exception: no pending exception (1)");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   495
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   496
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   497
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   498
489c9b5090e2 Initial load
duke
parents:
diff changeset
   499
    // compute exception handler into rbx
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   500
    __ movptr(c_rarg0, Address(rsp, 0));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   501
    BLOCK_COMMENT("call exception_handler_for_return_address");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   502
    __ call_VM_leaf(CAST_FROM_FN_PTR(address,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   503
                         SharedRuntime::exception_handler_for_return_address),
5046
27e801a857cb 6919934: JSR 292 needs to support x86 C1
twisti
parents: 4740
diff changeset
   504
                    r15_thread, c_rarg0);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   505
    __ mov(rbx, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   506
489c9b5090e2 Initial load
duke
parents:
diff changeset
   507
    // setup rax & rdx, remove return address & clear pending exception
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   508
    __ pop(rdx);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   509
    __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
1888
bbf498fb4354 6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents: 1437
diff changeset
   510
    __ movptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   511
489c9b5090e2 Initial load
duke
parents:
diff changeset
   512
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
   513
    // make sure exception is set
489c9b5090e2 Initial load
duke
parents:
diff changeset
   514
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   515
      Label L;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   516
      __ testptr(rax, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   517
      __ jcc(Assembler::notEqual, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   518
      __ stop("StubRoutines::forward exception: no pending exception (2)");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   519
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   520
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   521
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
   522
489c9b5090e2 Initial load
duke
parents:
diff changeset
   523
    // continue at exception handler (return address removed)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   524
    // rax: exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
   525
    // rbx: exception handler
489c9b5090e2 Initial load
duke
parents:
diff changeset
   526
    // rdx: throwing pc
489c9b5090e2 Initial load
duke
parents:
diff changeset
   527
    __ verify_oop(rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   528
    __ jmp(rbx);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   529
489c9b5090e2 Initial load
duke
parents:
diff changeset
   530
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   531
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   532
489c9b5090e2 Initial load
duke
parents:
diff changeset
   533
  // Support for jint atomic::xchg(jint exchange_value, volatile jint* dest)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   534
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   535
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   536
  //    c_rarg0: exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   537
  //    c_rarg0: dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   538
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   539
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   540
  //    *dest <- ex, return (orig *dest)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   541
  address generate_atomic_xchg() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   542
    StubCodeMark mark(this, "StubRoutines", "atomic_xchg");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   543
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   544
489c9b5090e2 Initial load
duke
parents:
diff changeset
   545
    __ movl(rax, c_rarg0); // Copy to eax we need a return value anyhow
489c9b5090e2 Initial load
duke
parents:
diff changeset
   546
    __ xchgl(rax, Address(c_rarg1, 0)); // automatic LOCK
489c9b5090e2 Initial load
duke
parents:
diff changeset
   547
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   548
489c9b5090e2 Initial load
duke
parents:
diff changeset
   549
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   550
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   551
489c9b5090e2 Initial load
duke
parents:
diff changeset
   552
  // Support for intptr_t atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   553
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   554
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   555
  //    c_rarg0: exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   556
  //    c_rarg1: dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   557
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   558
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   559
  //    *dest <- ex, return (orig *dest)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   560
  address generate_atomic_xchg_ptr() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   561
    StubCodeMark mark(this, "StubRoutines", "atomic_xchg_ptr");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   562
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   563
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   564
    __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   565
    __ xchgptr(rax, Address(c_rarg1, 0)); // automatic LOCK
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   566
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   567
489c9b5090e2 Initial load
duke
parents:
diff changeset
   568
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   569
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   570
489c9b5090e2 Initial load
duke
parents:
diff changeset
   571
  // Support for jint atomic::atomic_cmpxchg(jint exchange_value, volatile jint* dest,
489c9b5090e2 Initial load
duke
parents:
diff changeset
   572
  //                                         jint compare_value)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   573
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   574
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   575
  //    c_rarg0: exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   576
  //    c_rarg1: dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   577
  //    c_rarg2: compare_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   578
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   579
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   580
  //    if ( compare_value == *dest ) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   581
  //       *dest = exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   582
  //       return compare_value;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   583
  //    else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   584
  //       return *dest;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   585
  address generate_atomic_cmpxchg() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   586
    StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   587
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   588
489c9b5090e2 Initial load
duke
parents:
diff changeset
   589
    __ movl(rax, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   590
   if ( os::is_MP() ) __ lock();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   591
    __ cmpxchgl(c_rarg0, Address(c_rarg1, 0));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   592
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   593
489c9b5090e2 Initial load
duke
parents:
diff changeset
   594
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   595
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   596
27691
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   597
  // Support for jbyte atomic::atomic_cmpxchg(jbyte exchange_value, volatile jbyte* dest,
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   598
  //                                          jbyte compare_value)
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   599
  //
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   600
  // Arguments :
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   601
  //    c_rarg0: exchange_value
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   602
  //    c_rarg1: dest
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   603
  //    c_rarg2: compare_value
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   604
  //
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   605
  // Result:
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   606
  //    if ( compare_value == *dest ) {
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   607
  //       *dest = exchange_value
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   608
  //       return compare_value;
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   609
  //    else
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   610
  //       return *dest;
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   611
  address generate_atomic_cmpxchg_byte() {
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   612
    StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_byte");
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   613
    address start = __ pc();
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   614
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   615
    __ movsbq(rax, c_rarg2);
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   616
   if ( os::is_MP() ) __ lock();
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   617
    __ cmpxchgb(c_rarg0, Address(c_rarg1, 0));
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   618
    __ ret(0);
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   619
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   620
    return start;
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   621
  }
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   622
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   623
  // Support for jlong atomic::atomic_cmpxchg(jlong exchange_value,
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   624
  //                                          volatile jlong* dest,
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
   625
  //                                          jlong compare_value)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   626
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   627
  //    c_rarg0: exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   628
  //    c_rarg1: dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   629
  //    c_rarg2: compare_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   630
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   631
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   632
  //    if ( compare_value == *dest ) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   633
  //       *dest = exchange_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   634
  //       return compare_value;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   635
  //    else
489c9b5090e2 Initial load
duke
parents:
diff changeset
   636
  //       return *dest;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   637
  address generate_atomic_cmpxchg_long() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   638
    StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   639
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   640
489c9b5090e2 Initial load
duke
parents:
diff changeset
   641
    __ movq(rax, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   642
   if ( os::is_MP() ) __ lock();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   643
    __ cmpxchgq(c_rarg0, Address(c_rarg1, 0));
489c9b5090e2 Initial load
duke
parents:
diff changeset
   644
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   645
489c9b5090e2 Initial load
duke
parents:
diff changeset
   646
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   647
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   648
489c9b5090e2 Initial load
duke
parents:
diff changeset
   649
  // Support for jint atomic::add(jint add_value, volatile jint* dest)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   650
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   651
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   652
  //    c_rarg0: add_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   653
  //    c_rarg1: dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   654
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   655
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   656
  //    *dest += add_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   657
  //    return *dest;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   658
  address generate_atomic_add() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   659
    StubCodeMark mark(this, "StubRoutines", "atomic_add");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   660
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   661
489c9b5090e2 Initial load
duke
parents:
diff changeset
   662
    __ movl(rax, c_rarg0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   663
   if ( os::is_MP() ) __ lock();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   664
    __ xaddl(Address(c_rarg1, 0), c_rarg0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   665
    __ addl(rax, c_rarg0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   666
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   667
489c9b5090e2 Initial load
duke
parents:
diff changeset
   668
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   669
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   670
489c9b5090e2 Initial load
duke
parents:
diff changeset
   671
  // Support for intptr_t atomic::add_ptr(intptr_t add_value, volatile intptr_t* dest)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   672
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   673
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   674
  //    c_rarg0: add_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   675
  //    c_rarg1: dest
489c9b5090e2 Initial load
duke
parents:
diff changeset
   676
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   677
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   678
  //    *dest += add_value
489c9b5090e2 Initial load
duke
parents:
diff changeset
   679
  //    return *dest;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   680
  address generate_atomic_add_ptr() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   681
    StubCodeMark mark(this, "StubRoutines", "atomic_add_ptr");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   682
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   683
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   684
    __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   685
   if ( os::is_MP() ) __ lock();
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   686
    __ xaddptr(Address(c_rarg1, 0), c_rarg0);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   687
    __ addptr(rax, c_rarg0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   688
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   689
489c9b5090e2 Initial load
duke
parents:
diff changeset
   690
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   691
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   692
489c9b5090e2 Initial load
duke
parents:
diff changeset
   693
  // Support for intptr_t OrderAccess::fence()
489c9b5090e2 Initial load
duke
parents:
diff changeset
   694
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   695
  // Arguments :
489c9b5090e2 Initial load
duke
parents:
diff changeset
   696
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   697
  // Result:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   698
  address generate_orderaccess_fence() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   699
    StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   700
    address start = __ pc();
2338
a8660a1b709b 6822204: volatile fences should prefer lock:addl to actual mfence instructions
never
parents: 2259
diff changeset
   701
    __ membar(Assembler::StoreLoad);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   702
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   703
489c9b5090e2 Initial load
duke
parents:
diff changeset
   704
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   705
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   706
489c9b5090e2 Initial load
duke
parents:
diff changeset
   707
  // Support for intptr_t get_previous_fp()
489c9b5090e2 Initial load
duke
parents:
diff changeset
   708
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   709
  // This routine is used to find the previous frame pointer for the
489c9b5090e2 Initial load
duke
parents:
diff changeset
   710
  // caller (current_frame_guess). This is used as part of debugging
489c9b5090e2 Initial load
duke
parents:
diff changeset
   711
  // ps() is seemingly lost trying to find frames.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   712
  // This code assumes that caller current_frame_guess) has a frame.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   713
  address generate_get_previous_fp() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   714
    StubCodeMark mark(this, "StubRoutines", "get_previous_fp");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   715
    const Address old_fp(rbp, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   716
    const Address older_fp(rax, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   717
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   718
489c9b5090e2 Initial load
duke
parents:
diff changeset
   719
    __ enter();
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   720
    __ movptr(rax, old_fp); // callers fp
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   721
    __ movptr(rax, older_fp); // the frame for ps()
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   722
    __ pop(rbp);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   723
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   724
489c9b5090e2 Initial load
duke
parents:
diff changeset
   725
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   726
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   727
11961
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   728
  // Support for intptr_t get_previous_sp()
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   729
  //
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   730
  // This routine is used to find the previous stack pointer for the
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   731
  // caller.
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   732
  address generate_get_previous_sp() {
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   733
    StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   734
    address start = __ pc();
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   735
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   736
    __ movptr(rax, rsp);
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   737
    __ addptr(rax, 8); // return address is at the top of the stack.
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   738
    __ ret(0);
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   739
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   740
    return start;
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   741
  }
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
   742
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   743
  //----------------------------------------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
   744
  // Support for void verify_mxcsr()
489c9b5090e2 Initial load
duke
parents:
diff changeset
   745
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   746
  // This routine is used with -Xcheck:jni to verify that native
489c9b5090e2 Initial load
duke
parents:
diff changeset
   747
  // JNI code does not return to Java code without restoring the
489c9b5090e2 Initial load
duke
parents:
diff changeset
   748
  // MXCSR register to our expected state.
489c9b5090e2 Initial load
duke
parents:
diff changeset
   749
489c9b5090e2 Initial load
duke
parents:
diff changeset
   750
  address generate_verify_mxcsr() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   751
    StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   752
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   753
489c9b5090e2 Initial load
duke
parents:
diff changeset
   754
    const Address mxcsr_save(rsp, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   755
489c9b5090e2 Initial load
duke
parents:
diff changeset
   756
    if (CheckJNICalls) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   757
      Label ok_ret;
18958
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
   758
      ExternalAddress mxcsr_std(StubRoutines::addr_mxcsr_std());
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   759
      __ push(rax);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   760
      __ subptr(rsp, wordSize);      // allocate a temp location
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   761
      __ stmxcsr(mxcsr_save);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   762
      __ movl(rax, mxcsr_save);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   763
      __ andl(rax, MXCSR_MASK);    // Only check control and mask bits
18958
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
   764
      __ cmp32(rax, mxcsr_std);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   765
      __ jcc(Assembler::equal, ok_ret);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   766
489c9b5090e2 Initial load
duke
parents:
diff changeset
   767
      __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   768
18958
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
   769
      __ ldmxcsr(mxcsr_std);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   770
489c9b5090e2 Initial load
duke
parents:
diff changeset
   771
      __ bind(ok_ret);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   772
      __ addptr(rsp, wordSize);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   773
      __ pop(rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   774
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   775
489c9b5090e2 Initial load
duke
parents:
diff changeset
   776
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   777
489c9b5090e2 Initial load
duke
parents:
diff changeset
   778
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   779
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   780
489c9b5090e2 Initial load
duke
parents:
diff changeset
   781
  address generate_f2i_fixup() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   782
    StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   783
    Address inout(rsp, 5 * wordSize); // return address + 4 saves
489c9b5090e2 Initial load
duke
parents:
diff changeset
   784
489c9b5090e2 Initial load
duke
parents:
diff changeset
   785
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   786
489c9b5090e2 Initial load
duke
parents:
diff changeset
   787
    Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   788
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   789
    __ push(rax);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   790
    __ push(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   791
    __ push(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   792
    __ push(c_rarg1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   793
489c9b5090e2 Initial load
duke
parents:
diff changeset
   794
    __ movl(rax, 0x7f800000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   795
    __ xorl(c_rarg3, c_rarg3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   796
    __ movl(c_rarg2, inout);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   797
    __ movl(c_rarg1, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   798
    __ andl(c_rarg1, 0x7fffffff);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   799
    __ cmpl(rax, c_rarg1); // NaN? -> 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   800
    __ jcc(Assembler::negative, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   801
    __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
489c9b5090e2 Initial load
duke
parents:
diff changeset
   802
    __ movl(c_rarg3, 0x80000000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   803
    __ movl(rax, 0x7fffffff);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   804
    __ cmovl(Assembler::positive, c_rarg3, rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   805
489c9b5090e2 Initial load
duke
parents:
diff changeset
   806
    __ bind(L);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   807
    __ movptr(inout, c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   808
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   809
    __ pop(c_rarg1);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   810
    __ pop(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   811
    __ pop(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   812
    __ pop(rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   813
489c9b5090e2 Initial load
duke
parents:
diff changeset
   814
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   815
489c9b5090e2 Initial load
duke
parents:
diff changeset
   816
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   817
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   818
489c9b5090e2 Initial load
duke
parents:
diff changeset
   819
  address generate_f2l_fixup() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   820
    StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   821
    Address inout(rsp, 5 * wordSize); // return address + 4 saves
489c9b5090e2 Initial load
duke
parents:
diff changeset
   822
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   823
489c9b5090e2 Initial load
duke
parents:
diff changeset
   824
    Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   825
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   826
    __ push(rax);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   827
    __ push(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   828
    __ push(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   829
    __ push(c_rarg1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   830
489c9b5090e2 Initial load
duke
parents:
diff changeset
   831
    __ movl(rax, 0x7f800000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   832
    __ xorl(c_rarg3, c_rarg3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   833
    __ movl(c_rarg2, inout);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   834
    __ movl(c_rarg1, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   835
    __ andl(c_rarg1, 0x7fffffff);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   836
    __ cmpl(rax, c_rarg1); // NaN? -> 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   837
    __ jcc(Assembler::negative, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   838
    __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
489c9b5090e2 Initial load
duke
parents:
diff changeset
   839
    __ mov64(c_rarg3, 0x8000000000000000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   840
    __ mov64(rax, 0x7fffffffffffffff);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   841
    __ cmov(Assembler::positive, c_rarg3, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   842
489c9b5090e2 Initial load
duke
parents:
diff changeset
   843
    __ bind(L);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   844
    __ movptr(inout, c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   845
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   846
    __ pop(c_rarg1);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   847
    __ pop(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   848
    __ pop(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   849
    __ pop(rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   850
489c9b5090e2 Initial load
duke
parents:
diff changeset
   851
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   852
489c9b5090e2 Initial load
duke
parents:
diff changeset
   853
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   854
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   855
489c9b5090e2 Initial load
duke
parents:
diff changeset
   856
  address generate_d2i_fixup() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   857
    StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   858
    Address inout(rsp, 6 * wordSize); // return address + 5 saves
489c9b5090e2 Initial load
duke
parents:
diff changeset
   859
489c9b5090e2 Initial load
duke
parents:
diff changeset
   860
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   861
489c9b5090e2 Initial load
duke
parents:
diff changeset
   862
    Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   863
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   864
    __ push(rax);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   865
    __ push(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   866
    __ push(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   867
    __ push(c_rarg1);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   868
    __ push(c_rarg0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   869
489c9b5090e2 Initial load
duke
parents:
diff changeset
   870
    __ movl(rax, 0x7ff00000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   871
    __ movq(c_rarg2, inout);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   872
    __ movl(c_rarg3, c_rarg2);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   873
    __ mov(c_rarg1, c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   874
    __ mov(c_rarg0, c_rarg2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   875
    __ negl(c_rarg3);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   876
    __ shrptr(c_rarg1, 0x20);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   877
    __ orl(c_rarg3, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   878
    __ andl(c_rarg1, 0x7fffffff);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   879
    __ xorl(c_rarg2, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   880
    __ shrl(c_rarg3, 0x1f);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   881
    __ orl(c_rarg1, c_rarg3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   882
    __ cmpl(rax, c_rarg1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   883
    __ jcc(Assembler::negative, L); // NaN -> 0
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   884
    __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   885
    __ movl(c_rarg2, 0x80000000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   886
    __ movl(rax, 0x7fffffff);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   887
    __ cmov(Assembler::positive, c_rarg2, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   888
489c9b5090e2 Initial load
duke
parents:
diff changeset
   889
    __ bind(L);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   890
    __ movptr(inout, c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   891
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   892
    __ pop(c_rarg0);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   893
    __ pop(c_rarg1);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   894
    __ pop(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   895
    __ pop(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   896
    __ pop(rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   897
489c9b5090e2 Initial load
duke
parents:
diff changeset
   898
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   899
489c9b5090e2 Initial load
duke
parents:
diff changeset
   900
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   901
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   902
489c9b5090e2 Initial load
duke
parents:
diff changeset
   903
  address generate_d2l_fixup() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   904
    StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   905
    Address inout(rsp, 6 * wordSize); // return address + 5 saves
489c9b5090e2 Initial load
duke
parents:
diff changeset
   906
489c9b5090e2 Initial load
duke
parents:
diff changeset
   907
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   908
489c9b5090e2 Initial load
duke
parents:
diff changeset
   909
    Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   910
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   911
    __ push(rax);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   912
    __ push(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   913
    __ push(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   914
    __ push(c_rarg1);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   915
    __ push(c_rarg0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   916
489c9b5090e2 Initial load
duke
parents:
diff changeset
   917
    __ movl(rax, 0x7ff00000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   918
    __ movq(c_rarg2, inout);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   919
    __ movl(c_rarg3, c_rarg2);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   920
    __ mov(c_rarg1, c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   921
    __ mov(c_rarg0, c_rarg2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   922
    __ negl(c_rarg3);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   923
    __ shrptr(c_rarg1, 0x20);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   924
    __ orl(c_rarg3, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   925
    __ andl(c_rarg1, 0x7fffffff);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   926
    __ xorl(c_rarg2, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   927
    __ shrl(c_rarg3, 0x1f);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   928
    __ orl(c_rarg1, c_rarg3);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   929
    __ cmpl(rax, c_rarg1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   930
    __ jcc(Assembler::negative, L); // NaN -> 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
   931
    __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
489c9b5090e2 Initial load
duke
parents:
diff changeset
   932
    __ mov64(c_rarg2, 0x8000000000000000);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   933
    __ mov64(rax, 0x7fffffffffffffff);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   934
    __ cmovq(Assembler::positive, c_rarg2, rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   935
489c9b5090e2 Initial load
duke
parents:
diff changeset
   936
    __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   937
    __ movq(inout, c_rarg2);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   938
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   939
    __ pop(c_rarg0);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   940
    __ pop(c_rarg1);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   941
    __ pop(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   942
    __ pop(c_rarg3);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   943
    __ pop(rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   944
489c9b5090e2 Initial load
duke
parents:
diff changeset
   945
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   946
489c9b5090e2 Initial load
duke
parents:
diff changeset
   947
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   948
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   949
489c9b5090e2 Initial load
duke
parents:
diff changeset
   950
  address generate_fp_mask(const char *stub_name, int64_t mask) {
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5050
diff changeset
   951
    __ align(CodeEntryAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   952
    StubCodeMark mark(this, "StubRoutines", stub_name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   953
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   954
489c9b5090e2 Initial load
duke
parents:
diff changeset
   955
    __ emit_data64( mask, relocInfo::none );
489c9b5090e2 Initial load
duke
parents:
diff changeset
   956
    __ emit_data64( mask, relocInfo::none );
489c9b5090e2 Initial load
duke
parents:
diff changeset
   957
489c9b5090e2 Initial load
duke
parents:
diff changeset
   958
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   959
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   960
489c9b5090e2 Initial load
duke
parents:
diff changeset
   961
  // The following routine generates a subroutine to throw an
489c9b5090e2 Initial load
duke
parents:
diff changeset
   962
  // asynchronous UnknownError when an unsafe access gets a fault that
489c9b5090e2 Initial load
duke
parents:
diff changeset
   963
  // could not be reasonably prevented by the programmer.  (Example:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   964
  // SIGBUS/OBJERR.)
489c9b5090e2 Initial load
duke
parents:
diff changeset
   965
  address generate_handler_for_unsafe_access() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
   966
    StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   967
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
   968
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   969
    __ push(0);                       // hole for return address-to-be
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   970
    __ pusha();                       // push registers
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   971
    Address next_pc(rsp, RegisterImpl::number_of_registers * BytesPerWord);
489c9b5090e2 Initial load
duke
parents:
diff changeset
   972
10545
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10004
diff changeset
   973
    // FIXME: this probably needs alignment logic
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10004
diff changeset
   974
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   975
    __ subptr(rsp, frame::arg_reg_save_area_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   976
    BLOCK_COMMENT("call handle_unsafe_access");
489c9b5090e2 Initial load
duke
parents:
diff changeset
   977
    __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, handle_unsafe_access)));
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   978
    __ addptr(rsp, frame::arg_reg_save_area_bytes);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   979
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   980
    __ movptr(next_pc, rax);          // stuff next address
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
   981
    __ popa();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
   982
    __ ret(0);                        // jump to next address
489c9b5090e2 Initial load
duke
parents:
diff changeset
   983
489c9b5090e2 Initial load
duke
parents:
diff changeset
   984
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
   985
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
   986
489c9b5090e2 Initial load
duke
parents:
diff changeset
   987
  // Non-destructive plausibility checks for oops
489c9b5090e2 Initial load
duke
parents:
diff changeset
   988
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   989
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   990
  //    all args on stack!
489c9b5090e2 Initial load
duke
parents:
diff changeset
   991
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
   992
  // Stack after saving c_rarg3:
489c9b5090e2 Initial load
duke
parents:
diff changeset
   993
  //    [tos + 0]: saved c_rarg3
489c9b5090e2 Initial load
duke
parents:
diff changeset
   994
  //    [tos + 1]: saved c_rarg2
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
   995
  //    [tos + 2]: saved r12 (several TemplateTable methods use it)
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
   996
  //    [tos + 3]: saved flags
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
   997
  //    [tos + 4]: return address
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
   998
  //  * [tos + 5]: error message (char*)
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
   999
  //  * [tos + 6]: object to verify (oop)
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1000
  //  * [tos + 7]: saved rax - saved by caller and bashed
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1001
  //  * [tos + 8]: saved r10 (rscratch1) - saved by caller
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1002
  //  * = popped on exit
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1003
  address generate_verify_oop() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1004
    StubCodeMark mark(this, "StubRoutines", "verify_oop");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1005
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1006
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1007
    Label exit, error;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1008
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1009
    __ pushf();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1010
    __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1011
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1012
    __ push(r12);
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1013
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1014
    // save c_rarg2 and c_rarg3
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1015
    __ push(c_rarg2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1016
    __ push(c_rarg3);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1017
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1018
    enum {
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1019
           // After previous pushes.
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1020
           oop_to_verify = 6 * wordSize,
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1021
           saved_rax     = 7 * wordSize,
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1022
           saved_r10     = 8 * wordSize,
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1023
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1024
           // Before the call to MacroAssembler::debug(), see below.
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1025
           return_addr   = 16 * wordSize,
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1026
           error_msg     = 17 * wordSize
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1027
    };
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1028
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1029
    // get object
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1030
    __ movptr(rax, Address(rsp, oop_to_verify));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1031
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1032
    // make sure object is 'reasonable'
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1033
    __ testptr(rax, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1034
    __ jcc(Assembler::zero, exit); // if obj is NULL it is OK
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1035
    // Check if the oop is in the right area of memory
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1036
    __ movptr(c_rarg2, rax);
1888
bbf498fb4354 6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents: 1437
diff changeset
  1037
    __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask());
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1038
    __ andptr(c_rarg2, c_rarg3);
1888
bbf498fb4354 6787106: Hotspot 32 bit build fails on platforms having different definitions for intptr_t & int32_t
xlu
parents: 1437
diff changeset
  1039
    __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits());
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1040
    __ cmpptr(c_rarg2, c_rarg3);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1041
    __ jcc(Assembler::notZero, error);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1042
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1043
    // set r12 to heapbase for load_klass()
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1044
    __ reinit_heapbase();
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1045
13728
882756847a04 6964458: Reimplement class meta-data storage to use native memory
coleenp
parents: 13391
diff changeset
  1046
    // make sure klass is 'reasonable', which is not zero.
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1047
    __ load_klass(rax, rax);  // get klass
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1048
    __ testptr(rax, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1049
    __ jcc(Assembler::zero, error); // if klass is NULL it is broken
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1050
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1051
    // return if everything seems ok
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1052
    __ bind(exit);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1053
    __ movptr(rax, Address(rsp, saved_rax));     // get saved rax back
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1054
    __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1055
    __ pop(c_rarg3);                             // restore c_rarg3
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1056
    __ pop(c_rarg2);                             // restore c_rarg2
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1057
    __ pop(r12);                                 // restore r12
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1058
    __ popf();                                   // restore flags
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1059
    __ ret(4 * wordSize);                        // pop caller saved stuff
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1060
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1061
    // handle errors
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1062
    __ bind(error);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1063
    __ movptr(rax, Address(rsp, saved_rax));     // get saved rax back
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1064
    __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1065
    __ pop(c_rarg3);                             // get saved c_rarg3 back
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1066
    __ pop(c_rarg2);                             // get saved c_rarg2 back
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1067
    __ pop(r12);                                 // get saved r12 back
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1068
    __ popf();                                   // get saved flags off stack --
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1069
                                                 // will be ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1070
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1071
    __ pusha();                                  // push registers
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1072
                                                 // (rip is already
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1073
                                                 // already pushed)
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1074
    // debug(char* msg, int64_t pc, int64_t regs[])
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1075
    // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1076
    // pushed all the registers, so now the stack looks like:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1077
    //     [tos +  0] 16 saved registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1078
    //     [tos + 16] return address
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1079
    //   * [tos + 17] error message (char*)
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1080
    //   * [tos + 18] object to verify (oop)
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1081
    //   * [tos + 19] saved rax - saved by caller and bashed
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1082
    //   * [tos + 20] saved r10 (rscratch1) - saved by caller
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1083
    //   * = popped on exit
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1084
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1085
    __ movptr(c_rarg0, Address(rsp, error_msg));    // pass address of error message
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1086
    __ movptr(c_rarg1, Address(rsp, return_addr));  // pass return address
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1087
    __ movq(c_rarg2, rsp);                          // pass address of regs on stack
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1088
    __ mov(r12, rsp);                               // remember rsp
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1089
    __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1090
    __ andptr(rsp, -16);                            // align stack as required by ABI
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1091
    BLOCK_COMMENT("call MacroAssembler::debug");
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1092
    __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1093
    __ mov(rsp, r12);                               // restore rsp
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1094
    __ popa();                                      // pop registers (includes r12)
5706
0c91076143f9 6958254: -XX:+VerifyOops is broken on x86
kvn
parents: 5547
diff changeset
  1095
    __ ret(4 * wordSize);                           // pop caller saved stuff
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1096
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1097
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1098
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1099
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1100
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1101
  // Verify that a register contains clean 32-bits positive value
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1102
  // (high 32-bits are 0) so it could be used in 64-bits shifts.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1103
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1104
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1105
  //    Rint  -  32-bits value
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1106
  //    Rtmp  -  scratch
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1107
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1108
  void assert_clean_int(Register Rint, Register Rtmp) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1109
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1110
    Label L;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1111
    assert_different_registers(Rtmp, Rint);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1112
    __ movslq(Rtmp, Rint);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1113
    __ cmpq(Rtmp, Rint);
371
1aacedc9db7c 6689060: Escape Analysis does not work with Compressed Oops
kvn
parents: 360
diff changeset
  1114
    __ jcc(Assembler::equal, L);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1115
    __ stop("high 32-bits of int value are not 0");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1116
    __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1117
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1118
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1119
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1120
  //  Generate overlap test for array copy stubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1121
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1122
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1123
  //     c_rarg0 - from
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1124
  //     c_rarg1 - to
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1125
  //     c_rarg2 - element count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1126
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1127
  //  Output:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1128
  //     rax   - &from[element count - 1]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1129
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1130
  void array_overlap_test(address no_overlap_target, Address::ScaleFactor sf) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1131
    assert(no_overlap_target != NULL, "must be generated");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1132
    array_overlap_test(no_overlap_target, NULL, sf);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1133
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1134
  void array_overlap_test(Label& L_no_overlap, Address::ScaleFactor sf) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1135
    array_overlap_test(NULL, &L_no_overlap, sf);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1136
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1137
  void array_overlap_test(address no_overlap_target, Label* NOLp, Address::ScaleFactor sf) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1138
    const Register from     = c_rarg0;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1139
    const Register to       = c_rarg1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1140
    const Register count    = c_rarg2;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1141
    const Register end_from = rax;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1142
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1143
    __ cmpptr(to, from);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1144
    __ lea(end_from, Address(from, count, sf, 0));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1145
    if (NOLp == NULL) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1146
      ExternalAddress no_overlap(no_overlap_target);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1147
      __ jump_cc(Assembler::belowEqual, no_overlap);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1148
      __ cmpptr(to, end_from);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1149
      __ jump_cc(Assembler::aboveEqual, no_overlap);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1150
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1151
      __ jcc(Assembler::belowEqual, (*NOLp));
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1152
      __ cmpptr(to, end_from);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1153
      __ jcc(Assembler::aboveEqual, (*NOLp));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1154
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1155
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1156
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1157
  // Shuffle first three arg regs on Windows into Linux/Solaris locations.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1158
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1159
  // Outputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1160
  //    rdi - rcx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1161
  //    rsi - rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1162
  //    rdx - r8
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1163
  //    rcx - r9
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1164
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1165
  // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1166
  // are non-volatile.  r9 and r10 should not be used by the caller.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1167
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1168
  void setup_arg_regs(int nargs = 3) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1169
    const Register saved_rdi = r9;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1170
    const Register saved_rsi = r10;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1171
    assert(nargs == 3 || nargs == 4, "else fix");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1172
#ifdef _WIN64
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1173
    assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1174
           "unexpected argument registers");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1175
    if (nargs >= 4)
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1176
      __ mov(rax, r9);  // r9 is also saved_rdi
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1177
    __ movptr(saved_rdi, rdi);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1178
    __ movptr(saved_rsi, rsi);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1179
    __ mov(rdi, rcx); // c_rarg0
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1180
    __ mov(rsi, rdx); // c_rarg1
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1181
    __ mov(rdx, r8);  // c_rarg2
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1182
    if (nargs >= 4)
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1183
      __ mov(rcx, rax); // c_rarg3 (via rax)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1184
#else
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1185
    assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1186
           "unexpected argument registers");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1187
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1188
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1189
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1190
  void restore_arg_regs() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1191
    const Register saved_rdi = r9;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1192
    const Register saved_rsi = r10;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1193
#ifdef _WIN64
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1194
    __ movptr(rdi, saved_rdi);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1195
    __ movptr(rsi, saved_rsi);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1196
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1197
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1198
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1199
  // Generate code for an array write pre barrier
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1200
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1201
  //     addr    -  starting address
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1202
  //     count   -  element count
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1203
  //     tmp     - scratch register
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1204
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1205
  //     Destroy no registers!
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1206
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1207
  void  gen_write_ref_array_pre_barrier(Register addr, Register count, bool dest_uninitialized) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1208
    BarrierSet* bs = Universe::heap()->barrier_set();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1209
    switch (bs->kind()) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1210
      case BarrierSet::G1SATBCT:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1211
      case BarrierSet::G1SATBCTLogging:
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1212
        // 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
  1213
        if (!dest_uninitialized) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1214
           __ pusha();                      // push registers
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1215
           if (count == c_rarg0) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1216
             if (addr == c_rarg1) {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1217
               // exactly backwards!!
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1218
               __ xchgptr(c_rarg1, c_rarg0);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1219
             } else {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1220
               __ movptr(c_rarg1, count);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1221
               __ movptr(c_rarg0, addr);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1222
             }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1223
           } else {
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1224
             __ movptr(c_rarg0, addr);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1225
             __ movptr(c_rarg1, count);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1226
           }
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1227
           __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), 2);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1228
           __ popa();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1229
        }
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1230
         break;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1231
      case BarrierSet::CardTableModRef:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1232
      case BarrierSet::CardTableExtension:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1233
      case BarrierSet::ModRef:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1234
        break;
1374
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1235
      default:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1236
        ShouldNotReachHere();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1237
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1238
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1239
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1240
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1241
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1242
  // Generate code for an array write post barrier
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1243
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1244
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1245
  //     start    - register containing starting address of destination array
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1246
  //     count    - elements count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1247
  //     scratch  - scratch register
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1248
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1249
  //  The input registers are overwritten.
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1250
  //
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1251
  void  gen_write_ref_array_post_barrier(Register start, Register count, Register scratch) {
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1252
    assert_different_registers(start, count, scratch);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1253
    BarrierSet* bs = Universe::heap()->barrier_set();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1254
    switch (bs->kind()) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1255
      case BarrierSet::G1SATBCT:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1256
      case BarrierSet::G1SATBCTLogging:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1257
        {
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1258
          __ pusha();             // push registers (overkill)
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1259
          if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1260
            assert_different_registers(c_rarg1, start);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1261
            __ mov(c_rarg1, count);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1262
            __ mov(c_rarg0, start);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1263
          } else {
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1264
            assert_different_registers(c_rarg0, count);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1265
            __ mov(c_rarg0, start);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1266
            __ mov(c_rarg1, count);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1267
          }
4740
d708800308b7 6918006: G1: spill space must be reserved on the stack for barrier calls on Windows x64
apetrusenko
parents: 4645
diff changeset
  1268
          __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), 2);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1269
          __ popa();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1270
        }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1271
        break;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1272
      case BarrierSet::CardTableModRef:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1273
      case BarrierSet::CardTableExtension:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1274
        {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1275
          CardTableModRefBS* ct = (CardTableModRefBS*)bs;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1276
          assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1277
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1278
          Label L_loop;
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1279
          const Register end = count;
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1280
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1281
          __ leaq(end, Address(start, count, TIMES_OOP, 0));  // end == start+count*oop_size
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1282
          __ subptr(end, BytesPerHeapOop); // end - 1 to make inclusive
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1283
          __ shrptr(start, CardTableModRefBS::card_shift);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1284
          __ shrptr(end,   CardTableModRefBS::card_shift);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1285
          __ subptr(end, start); // end --> cards count
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1286
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1287
          int64_t disp = (int64_t) ct->byte_map_base;
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1288
          __ mov64(scratch, disp);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1289
          __ addptr(start, scratch);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1290
        __ BIND(L_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1291
          __ movb(Address(start, count, Address::times_1), 0);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1292
          __ decrement(count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1293
          __ jcc(Assembler::greaterEqual, L_loop);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1294
        }
1374
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1295
        break;
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1296
      default:
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1297
        ShouldNotReachHere();
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1298
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1299
    }
4c24294029a9 6711316: Open source the Garbage-First garbage collector
ysr
parents: 371
diff changeset
  1300
  }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1301
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1302
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1303
  // Copy big chunks forward
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1304
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1305
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1306
  //   end_from     - source arrays end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1307
  //   end_to       - destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1308
  //   qword_count  - 64-bits element count, negative
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1309
  //   to           - scratch
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1310
  //   L_copy_bytes - entry label
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1311
  //   L_copy_8_bytes  - exit  label
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1312
  //
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1313
  void copy_bytes_forward(Register end_from, Register end_to,
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1314
                             Register qword_count, Register to,
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1315
                             Label& L_copy_bytes, Label& L_copy_8_bytes) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1316
    DEBUG_ONLY(__ stop("enter at entry label, not here"));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1317
    Label L_loop;
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5050
diff changeset
  1318
    __ align(OptoLoopAlignment);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1319
    if (UseUnalignedLoadStores) {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1320
      Label L_end;
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1321
      // Copy 64-bytes per iteration
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1322
      __ BIND(L_loop);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1323
      if (UseAVX >= 2) {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1324
        __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1325
        __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1326
        __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1327
        __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1328
      } else {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1329
        __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1330
        __ movdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1331
        __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, -40));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1332
        __ movdqu(Address(end_to, qword_count, Address::times_8, -40), xmm1);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1333
        __ movdqu(xmm2, Address(end_from, qword_count, Address::times_8, -24));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1334
        __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm2);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1335
        __ movdqu(xmm3, Address(end_from, qword_count, Address::times_8, - 8));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1336
        __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm3);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1337
      }
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1338
      __ BIND(L_copy_bytes);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1339
      __ addptr(qword_count, 8);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1340
      __ jcc(Assembler::lessEqual, L_loop);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1341
      __ subptr(qword_count, 4);  // sub(8) and add(4)
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1342
      __ jccb(Assembler::greater, L_end);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1343
      // Copy trailing 32 bytes
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1344
      if (UseAVX >= 2) {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1345
        __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1346
        __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1347
      } else {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1348
        __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1349
        __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1350
        __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1351
        __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1352
      }
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1353
      __ addptr(qword_count, 4);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1354
      __ BIND(L_end);
16624
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1355
      if (UseAVX >= 2) {
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1356
        // clean upper bits of YMM registers
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1357
        __ vzeroupper();
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1358
      }
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1359
    } else {
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1360
      // Copy 32-bytes per iteration
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1361
      __ BIND(L_loop);
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1362
      __ movq(to, Address(end_from, qword_count, Address::times_8, -24));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1363
      __ movq(Address(end_to, qword_count, Address::times_8, -24), to);
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1364
      __ movq(to, Address(end_from, qword_count, Address::times_8, -16));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1365
      __ movq(Address(end_to, qword_count, Address::times_8, -16), to);
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1366
      __ movq(to, Address(end_from, qword_count, Address::times_8, - 8));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1367
      __ movq(Address(end_to, qword_count, Address::times_8, - 8), to);
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1368
      __ movq(to, Address(end_from, qword_count, Address::times_8, - 0));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1369
      __ movq(Address(end_to, qword_count, Address::times_8, - 0), to);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1370
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1371
      __ BIND(L_copy_bytes);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1372
      __ addptr(qword_count, 4);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1373
      __ jcc(Assembler::lessEqual, L_loop);
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1374
    }
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1375
    __ subptr(qword_count, 4);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1376
    __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1377
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1378
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1379
  // Copy big chunks backward
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1380
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1381
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1382
  //   from         - source arrays address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1383
  //   dest         - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1384
  //   qword_count  - 64-bits element count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1385
  //   to           - scratch
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1386
  //   L_copy_bytes - entry label
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1387
  //   L_copy_8_bytes  - exit  label
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1388
  //
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1389
  void copy_bytes_backward(Register from, Register dest,
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1390
                              Register qword_count, Register to,
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1391
                              Label& L_copy_bytes, Label& L_copy_8_bytes) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1392
    DEBUG_ONLY(__ stop("enter at entry label, not here"));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1393
    Label L_loop;
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5050
diff changeset
  1394
    __ align(OptoLoopAlignment);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1395
    if (UseUnalignedLoadStores) {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1396
      Label L_end;
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1397
      // Copy 64-bytes per iteration
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1398
      __ BIND(L_loop);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1399
      if (UseAVX >= 2) {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1400
        __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1401
        __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1402
        __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8,  0));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1403
        __ vmovdqu(Address(dest, qword_count, Address::times_8,  0), xmm1);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1404
      } else {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1405
        __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1406
        __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1407
        __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1408
        __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1409
        __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1410
        __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1411
        __ movdqu(xmm3, Address(from, qword_count, Address::times_8,  0));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1412
        __ movdqu(Address(dest, qword_count, Address::times_8,  0), xmm3);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1413
      }
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1414
      __ BIND(L_copy_bytes);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1415
      __ subptr(qword_count, 8);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1416
      __ jcc(Assembler::greaterEqual, L_loop);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1417
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1418
      __ addptr(qword_count, 4);  // add(8) and sub(4)
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1419
      __ jccb(Assembler::less, L_end);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1420
      // Copy trailing 32 bytes
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1421
      if (UseAVX >= 2) {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1422
        __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 0));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1423
        __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1424
      } else {
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1425
        __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1426
        __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1427
        __ movdqu(xmm1, Address(from, qword_count, Address::times_8,  0));
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1428
        __ movdqu(Address(dest, qword_count, Address::times_8,  0), xmm1);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1429
      }
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1430
      __ subptr(qword_count, 4);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1431
      __ BIND(L_end);
16624
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1432
      if (UseAVX >= 2) {
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1433
        // clean upper bits of YMM registers
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1434
        __ vzeroupper();
9dbd4b210bf9 8011102: Clear AVX registers after return from JNI call
kvn
parents: 15115
diff changeset
  1435
      }
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1436
    } else {
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1437
      // Copy 32-bytes per iteration
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1438
      __ BIND(L_loop);
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1439
      __ movq(to, Address(from, qword_count, Address::times_8, 24));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1440
      __ movq(Address(dest, qword_count, Address::times_8, 24), to);
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1441
      __ movq(to, Address(from, qword_count, Address::times_8, 16));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1442
      __ movq(Address(dest, qword_count, Address::times_8, 16), to);
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1443
      __ movq(to, Address(from, qword_count, Address::times_8,  8));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1444
      __ movq(Address(dest, qword_count, Address::times_8,  8), to);
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1445
      __ movq(to, Address(from, qword_count, Address::times_8,  0));
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1446
      __ movq(Address(dest, qword_count, Address::times_8,  0), to);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1447
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1448
      __ BIND(L_copy_bytes);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1449
      __ subptr(qword_count, 4);
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1450
      __ jcc(Assembler::greaterEqual, L_loop);
1437
d1846c1c04c4 6532536: Optimize arraycopy stubs for Intel cpus
kvn
parents: 1394
diff changeset
  1451
    }
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1452
    __ addptr(qword_count, 4);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1453
    __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1454
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1455
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1456
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1457
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1458
  //   aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1459
  //             ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1460
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1461
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1462
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1463
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1464
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1465
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1466
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1467
  // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1468
  // we let the hardware handle it.  The one to eight bytes within words,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1469
  // dwords or qwords that span cache line boundaries will still be loaded
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1470
  // and stored atomically.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1471
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1472
  // Side Effects:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1473
  //   disjoint_byte_copy_entry is set to the no-overlap entry point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1474
  //   used by generate_conjoint_byte_copy().
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1475
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1476
  address generate_disjoint_byte_copy(bool aligned, address* entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1477
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1478
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1479
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1480
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1481
    Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1482
    Label L_copy_byte, L_exit;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1483
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1484
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1485
    const Register count       = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1486
    const Register byte_count  = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1487
    const Register qword_count = count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1488
    const Register end_from    = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1489
    const Register end_to      = to;   // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1490
    // End pointers are inclusive, and if count is not zero they point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1491
    // to the last unit copied:  end_to[0] := end_from[0]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1492
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1493
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1494
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1495
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1496
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1497
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1498
       // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1499
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1500
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1501
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1502
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1503
                      // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1504
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1505
    // 'from', 'to' and 'count' are now valid
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1506
    __ movptr(byte_count, count);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1507
    __ shrptr(count, 3); // count => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1508
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1509
    // Copy from low to high addresses.  Use 'to' as scratch.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1510
    __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1511
    __ lea(end_to,   Address(to,   qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1512
    __ negptr(qword_count); // make the count negative
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1513
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1514
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1515
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1516
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1517
    __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1518
    __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1519
    __ increment(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1520
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1521
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1522
    // Check for and copy trailing dword
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1523
  __ BIND(L_copy_4_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1524
    __ testl(byte_count, 4);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1525
    __ jccb(Assembler::zero, L_copy_2_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1526
    __ movl(rax, Address(end_from, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1527
    __ movl(Address(end_to, 8), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1528
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1529
    __ addptr(end_from, 4);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1530
    __ addptr(end_to, 4);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1531
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1532
    // Check for and copy trailing word
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1533
  __ BIND(L_copy_2_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1534
    __ testl(byte_count, 2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1535
    __ jccb(Assembler::zero, L_copy_byte);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1536
    __ movw(rax, Address(end_from, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1537
    __ movw(Address(end_to, 8), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1538
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1539
    __ addptr(end_from, 2);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1540
    __ addptr(end_to, 2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1541
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1542
    // Check for and copy trailing byte
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1543
  __ BIND(L_copy_byte);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1544
    __ testl(byte_count, 1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1545
    __ jccb(Assembler::zero, L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1546
    __ movb(rax, Address(end_from, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1547
    __ movb(Address(end_to, 8), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1548
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1549
  __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1550
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1551
    inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1552
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1553
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1554
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1555
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1556
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1557
    copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1558
    __ jmp(L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1559
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1560
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1561
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1562
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1563
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1564
  //   aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1565
  //             ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1566
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1567
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1568
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1569
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1570
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1571
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1572
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1573
  // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1574
  // we let the hardware handle it.  The one to eight bytes within words,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1575
  // dwords or qwords that span cache line boundaries will still be loaded
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1576
  // and stored atomically.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1577
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1578
  address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1579
                                      address* entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1580
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1581
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1582
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1583
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1584
    Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1585
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1586
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1587
    const Register count       = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1588
    const Register byte_count  = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1589
    const Register qword_count = count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1590
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1591
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1592
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1593
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1594
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1595
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1596
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1597
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1598
    }
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1599
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1600
    array_overlap_test(nooverlap_target, Address::times_1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1601
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1602
                      // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1603
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1604
    // 'from', 'to' and 'count' are now valid
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1605
    __ movptr(byte_count, count);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1606
    __ shrptr(count, 3);   // count => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1607
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1608
    // Copy from high to low addresses.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1609
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1610
    // Check for and copy trailing byte
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1611
    __ testl(byte_count, 1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1612
    __ jcc(Assembler::zero, L_copy_2_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1613
    __ movb(rax, Address(from, byte_count, Address::times_1, -1));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1614
    __ movb(Address(to, byte_count, Address::times_1, -1), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1615
    __ decrement(byte_count); // Adjust for possible trailing word
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1616
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1617
    // Check for and copy trailing word
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1618
  __ BIND(L_copy_2_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1619
    __ testl(byte_count, 2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1620
    __ jcc(Assembler::zero, L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1621
    __ movw(rax, Address(from, byte_count, Address::times_1, -2));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1622
    __ movw(Address(to, byte_count, Address::times_1, -2), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1623
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1624
    // Check for and copy trailing dword
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1625
  __ BIND(L_copy_4_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1626
    __ testl(byte_count, 4);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1627
    __ jcc(Assembler::zero, L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1628
    __ movl(rax, Address(from, qword_count, Address::times_8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1629
    __ movl(Address(to, qword_count, Address::times_8), rax);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1630
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1631
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1632
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1633
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1634
    __ movq(rax, Address(from, qword_count, Address::times_8, -8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1635
    __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1636
    __ decrement(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1637
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1638
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1639
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1640
    inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1641
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1642
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1643
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1644
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1645
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1646
    copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1647
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1648
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1649
    inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1650
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1651
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1652
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1653
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1654
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1655
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1656
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1657
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1658
  //   aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1659
  //             ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1660
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1661
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1662
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1663
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1664
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1665
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1666
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1667
  // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1668
  // let the hardware handle it.  The two or four words within dwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1669
  // or qwords that span cache line boundaries will still be loaded
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1670
  // and stored atomically.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1671
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1672
  // Side Effects:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1673
  //   disjoint_short_copy_entry is set to the no-overlap entry point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1674
  //   used by generate_conjoint_short_copy().
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1675
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1676
  address generate_disjoint_short_copy(bool aligned, address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1677
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1678
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1679
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1680
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1681
    Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1682
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1683
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1684
    const Register count       = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1685
    const Register word_count  = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1686
    const Register qword_count = count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1687
    const Register end_from    = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1688
    const Register end_to      = to;   // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1689
    // End pointers are inclusive, and if count is not zero they point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1690
    // to the last unit copied:  end_to[0] := end_from[0]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1691
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1692
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1693
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1694
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1695
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1696
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1697
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1698
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1699
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1700
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1701
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1702
                      // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1703
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1704
    // 'from', 'to' and 'count' are now valid
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1705
    __ movptr(word_count, count);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1706
    __ shrptr(count, 2); // count => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1707
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1708
    // Copy from low to high addresses.  Use 'to' as scratch.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1709
    __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1710
    __ lea(end_to,   Address(to,   qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1711
    __ negptr(qword_count);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1712
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1713
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1714
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1715
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1716
    __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1717
    __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1718
    __ increment(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1719
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1720
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1721
    // Original 'dest' is trashed, so we can't use it as a
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1722
    // base register for a possible trailing word copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1723
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1724
    // Check for and copy trailing dword
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1725
  __ BIND(L_copy_4_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1726
    __ testl(word_count, 2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1727
    __ jccb(Assembler::zero, L_copy_2_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1728
    __ movl(rax, Address(end_from, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1729
    __ movl(Address(end_to, 8), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1730
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1731
    __ addptr(end_from, 4);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1732
    __ addptr(end_to, 4);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1733
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1734
    // Check for and copy trailing word
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1735
  __ BIND(L_copy_2_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1736
    __ testl(word_count, 1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1737
    __ jccb(Assembler::zero, L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1738
    __ movw(rax, Address(end_from, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1739
    __ movw(Address(end_to, 8), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1740
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1741
  __ BIND(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1742
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1743
    inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1744
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1745
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1746
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1747
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1748
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1749
    copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1750
    __ jmp(L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1751
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1752
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1753
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1754
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1755
  address generate_fill(BasicType t, bool aligned, const char *name) {
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1756
    __ align(CodeEntryAlignment);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1757
    StubCodeMark mark(this, "StubRoutines", name);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1758
    address start = __ pc();
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1759
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1760
    BLOCK_COMMENT("Entry:");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1761
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1762
    const Register to       = c_rarg0;  // source array address
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1763
    const Register value    = c_rarg1;  // value
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1764
    const Register count    = c_rarg2;  // elements count
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1765
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1766
    __ enter(); // required for proper stackwalking of RuntimeStub frame
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1767
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1768
    __ generate_fill(t, aligned, to, value, count, rax, xmm0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1769
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1770
    __ leave(); // required for proper stackwalking of RuntimeStub frame
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1771
    __ ret(0);
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1772
    return start;
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1773
  }
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  1774
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1775
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1776
  //   aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1777
  //             ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1778
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1779
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1780
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1781
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1782
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1783
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1784
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1785
  // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1786
  // let the hardware handle it.  The two or four words within dwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1787
  // or qwords that span cache line boundaries will still be loaded
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1788
  // and stored atomically.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1789
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1790
  address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1791
                                       address *entry, const char *name) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1792
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1793
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1794
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1795
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1796
    Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1797
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1798
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1799
    const Register count       = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1800
    const Register word_count  = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1801
    const Register qword_count = count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1802
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1803
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1804
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1805
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1806
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1807
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1808
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1809
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1810
    }
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1811
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1812
    array_overlap_test(nooverlap_target, Address::times_2);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1813
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1814
                      // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1815
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1816
    // 'from', 'to' and 'count' are now valid
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1817
    __ movptr(word_count, count);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1818
    __ shrptr(count, 2); // count => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1819
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1820
    // Copy from high to low addresses.  Use 'to' as scratch.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1821
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1822
    // Check for and copy trailing word
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1823
    __ testl(word_count, 1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1824
    __ jccb(Assembler::zero, L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1825
    __ movw(rax, Address(from, word_count, Address::times_2, -2));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1826
    __ movw(Address(to, word_count, Address::times_2, -2), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1827
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1828
    // Check for and copy trailing dword
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1829
  __ BIND(L_copy_4_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1830
    __ testl(word_count, 2);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1831
    __ jcc(Assembler::zero, L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1832
    __ movl(rax, Address(from, qword_count, Address::times_8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1833
    __ movl(Address(to, qword_count, Address::times_8), rax);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1834
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1835
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1836
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1837
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1838
    __ movq(rax, Address(from, qword_count, Address::times_8, -8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1839
    __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1840
    __ decrement(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1841
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1842
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1843
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1844
    inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1845
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1846
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1847
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1848
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1849
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1850
    copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1851
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1852
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1853
    inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1854
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1855
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1856
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1857
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1858
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1859
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1860
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1861
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1862
  //   aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1863
  //             ignored
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1864
  //   is_oop  - true => oop array, so generate store check code
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1865
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1866
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1867
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1868
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1869
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1870
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1871
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1872
  // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1873
  // the hardware handle it.  The two dwords within qwords that span
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1874
  // cache line boundaries will still be loaded and stored atomicly.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1875
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1876
  // Side Effects:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1877
  //   disjoint_int_copy_entry is set to the no-overlap entry point
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1878
  //   used by generate_conjoint_int_oop_copy().
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1879
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1880
  address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, address* entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1881
                                         const char *name, bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1882
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1883
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1884
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1885
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1886
    Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1887
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1888
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1889
    const Register count       = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1890
    const Register dword_count = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1891
    const Register qword_count = count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1892
    const Register end_from    = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1893
    const Register end_to      = to;   // destination array end address
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1894
    const Register saved_to    = r11;  // saved destination array address
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1895
    // End pointers are inclusive, and if count is not zero they point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1896
    // to the last unit copied:  end_to[0] := end_from[0]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1897
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1898
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1899
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1900
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1901
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1902
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1903
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1904
      BLOCK_COMMENT("Entry:");
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1905
    }
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1906
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1907
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1908
                      // r9 and r10 may be used to save non-volatile registers
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1909
    if (is_oop) {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1910
      __ movq(saved_to, to);
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1911
      gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1912
    }
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1913
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1914
    // 'from', 'to' and 'count' are now valid
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1915
    __ movptr(dword_count, count);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1916
    __ shrptr(count, 1); // count => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1917
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1918
    // Copy from low to high addresses.  Use 'to' as scratch.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1919
    __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1920
    __ lea(end_to,   Address(to,   qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1921
    __ negptr(qword_count);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1922
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1923
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1924
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1925
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1926
    __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1927
    __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1928
    __ increment(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1929
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1930
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1931
    // Check for and copy trailing dword
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1932
  __ BIND(L_copy_4_bytes);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1933
    __ testl(dword_count, 1); // Only byte test since the value is 0 or 1
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1934
    __ jccb(Assembler::zero, L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1935
    __ movl(rax, Address(end_from, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1936
    __ movl(Address(end_to, 8), rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1937
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1938
  __ BIND(L_exit);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1939
    if (is_oop) {
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  1940
      gen_write_ref_array_post_barrier(saved_to, dword_count, rax);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1941
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1942
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  1943
    inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  1944
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1945
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1946
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1947
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1948
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1949
    copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1950
    __ jmp(L_copy_4_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1951
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1952
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1953
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1954
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1955
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1956
  //   aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1957
  //             ignored
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1958
  //   is_oop  - true => oop array, so generate store check code
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1959
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1960
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1961
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1962
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1963
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1964
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1965
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1966
  // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1967
  // the hardware handle it.  The two dwords within qwords that span
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1968
  // cache line boundaries will still be loaded and stored atomicly.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1969
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1970
  address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, address nooverlap_target,
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1971
                                         address *entry, const char *name,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1972
                                         bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1973
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1974
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1975
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1976
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  1977
    Label L_copy_bytes, L_copy_8_bytes, L_copy_2_bytes, L_exit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1978
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1979
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1980
    const Register count       = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1981
    const Register dword_count = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1982
    const Register qword_count = count;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1983
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1984
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1985
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  1986
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1987
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1988
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1989
       // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1990
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1991
    }
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1992
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1993
    array_overlap_test(nooverlap_target, Address::times_4);
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1994
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1995
                      // r9 and r10 may be used to save non-volatile registers
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  1996
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1997
    if (is_oop) {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  1998
      // no registers are destroyed by this call
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  1999
      gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2000
    }
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2001
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2002
    assert_clean_int(count, rax); // Make sure 'count' is clean int.
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2003
    // 'from', 'to' and 'count' are now valid
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2004
    __ movptr(dword_count, count);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2005
    __ shrptr(count, 1); // count => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2006
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2007
    // Copy from high to low addresses.  Use 'to' as scratch.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2008
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2009
    // Check for and copy trailing dword
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2010
    __ testl(dword_count, 1);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2011
    __ jcc(Assembler::zero, L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2012
    __ movl(rax, Address(from, dword_count, Address::times_4, -4));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2013
    __ movl(Address(to, dword_count, Address::times_4, -4), rax);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2014
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2015
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2016
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2017
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2018
    __ movq(rax, Address(from, qword_count, Address::times_8, -8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2019
    __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2020
    __ decrement(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2021
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2022
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2023
    if (is_oop) {
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2024
      __ jmp(L_exit);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2025
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2026
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2027
    inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2028
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2029
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2030
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2031
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2032
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2033
    copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2034
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2035
  __ BIND(L_exit);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2036
    if (is_oop) {
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2037
      gen_write_ref_array_post_barrier(to, dword_count, rax);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2038
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2039
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2040
    inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2041
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2042
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2043
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2044
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2045
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2046
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2047
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2048
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2049
  //   aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2050
  //             ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2051
  //   is_oop  - true => oop array, so generate store check code
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2052
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2053
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2054
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2055
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2056
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2057
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2058
  //
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2059
 // Side Effects:
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2060
  //   disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2061
  //   no-overlap entry point used by generate_conjoint_long_oop_copy().
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2062
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2063
  address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, address *entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2064
                                          const char *name, bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2065
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2066
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2067
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2068
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2069
    Label L_copy_bytes, L_copy_8_bytes, L_exit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2070
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2071
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2072
    const Register qword_count = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2073
    const Register end_from    = from; // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2074
    const Register end_to      = rcx;  // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2075
    const Register saved_to    = to;
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2076
    const Register saved_count = r11;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2077
    // End pointers are inclusive, and if count is not zero they point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2078
    // to the last unit copied:  end_to[0] := end_from[0]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2079
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2080
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2081
    // Save no-overlap entry point for generate_conjoint_long_oop_copy()
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2082
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2083
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2084
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2085
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2086
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2087
      BLOCK_COMMENT("Entry:");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2088
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2089
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2090
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2091
                      // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2092
    // 'from', 'to' and 'qword_count' are now valid
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2093
    if (is_oop) {
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2094
      // Save to and count for store barrier
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2095
      __ movptr(saved_count, qword_count);
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2096
      // no registers are destroyed by this call
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2097
      gen_write_ref_array_pre_barrier(to, qword_count, dest_uninitialized);
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2098
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2099
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2100
    // Copy from low to high addresses.  Use 'to' as scratch.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2101
    __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2102
    __ lea(end_to,   Address(to,   qword_count, Address::times_8, -8));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2103
    __ negptr(qword_count);
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2104
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2105
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2106
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2107
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2108
    __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2109
    __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2110
    __ increment(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2111
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2112
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2113
    if (is_oop) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2114
      __ jmp(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2115
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2116
      restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2117
      inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2118
      __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2119
      __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2120
      __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2121
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2122
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2123
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2124
    copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2125
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2126
    if (is_oop) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2127
    __ BIND(L_exit);
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2128
      gen_write_ref_array_post_barrier(saved_to, saved_count, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2129
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2130
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2131
    if (is_oop) {
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2132
      inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2133
    } else {
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2134
      inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2135
    }
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2136
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2137
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2138
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2139
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2140
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2141
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2142
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2143
  // Arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2144
  //   aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2145
  //             ignored
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2146
  //   is_oop  - true => oop array, so generate store check code
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2147
  //   name    - stub name string
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2148
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2149
  // Inputs:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2150
  //   c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2151
  //   c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2152
  //   c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2153
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2154
  address generate_conjoint_long_oop_copy(bool aligned, bool is_oop,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2155
                                          address nooverlap_target, address *entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2156
                                          const char *name, bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2157
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2158
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2159
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2160
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2161
    Label L_copy_bytes, L_copy_8_bytes, L_exit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2162
    const Register from        = rdi;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2163
    const Register to          = rsi;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2164
    const Register qword_count = rdx;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2165
    const Register saved_count = rcx;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2166
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2167
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2168
    assert_clean_int(c_rarg2, rax);    // Make sure 'count' is clean int.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2169
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2170
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2171
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2172
      // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2173
      BLOCK_COMMENT("Entry:");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2174
    }
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2175
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2176
    array_overlap_test(nooverlap_target, Address::times_8);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2177
    setup_arg_regs(); // from => rdi, to => rsi, count => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2178
                      // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2179
    // 'from', 'to' and 'qword_count' are now valid
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2180
    if (is_oop) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2181
      // Save to and count for store barrier
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2182
      __ movptr(saved_count, qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2183
      // No registers are destroyed by this call
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2184
      gen_write_ref_array_pre_barrier(to, saved_count, dest_uninitialized);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2185
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2186
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2187
    __ jmp(L_copy_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2188
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2189
    // Copy trailing qwords
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2190
  __ BIND(L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2191
    __ movq(rax, Address(from, qword_count, Address::times_8, -8));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2192
    __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2193
    __ decrement(qword_count);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2194
    __ jcc(Assembler::notZero, L_copy_8_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2195
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2196
    if (is_oop) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2197
      __ jmp(L_exit);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2198
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2199
      restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2200
      inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2201
      __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2202
      __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2203
      __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2204
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2205
15115
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2206
    // Copy in multi-bytes chunks
f8ef87f6f07f 8005544: Use 256bit YMM registers in arraycopy stubs on x86
kvn
parents: 14834
diff changeset
  2207
    copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2208
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2209
    if (is_oop) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2210
    __ BIND(L_exit);
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2211
      gen_write_ref_array_post_barrier(to, saved_count, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2212
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2213
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2214
    if (is_oop) {
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2215
      inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2216
    } else {
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2217
      inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2218
    }
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2219
    __ xorptr(rax, rax); // return 0
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2220
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2221
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2222
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2223
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2224
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2225
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2226
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2227
  // Helper for generating a dynamic type check.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2228
  // Smashes no registers.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2229
  void generate_type_check(Register sub_klass,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2230
                           Register super_check_offset,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2231
                           Register super_klass,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2232
                           Label& L_success) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2233
    assert_different_registers(sub_klass, super_check_offset, super_klass);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2234
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2235
    BLOCK_COMMENT("type_check:");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2236
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2237
    Label L_miss;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2238
2256
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 1888
diff changeset
  2239
    __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg,        &L_success, &L_miss, NULL,
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 1888
diff changeset
  2240
                                     super_check_offset);
82d4e10b7c6b 6813212: factor duplicated assembly code for general subclass check (for 6655638)
jrose
parents: 1888
diff changeset
  2241
    __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2242
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2243
    // Fall through on failure!
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2244
    __ BIND(L_miss);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2245
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2246
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2247
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2248
  //  Generate checkcasting array copy stub
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2249
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2250
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2251
  //    c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2252
  //    c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2253
  //    c_rarg2   - element count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2254
  //    c_rarg3   - size_t ckoff (super_check_offset)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2255
  // not Win64
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2256
  //    c_rarg4   - oop ckval (super_klass)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2257
  // Win64
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2258
  //    rsp+40    - oop ckval (super_klass)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2259
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2260
  //  Output:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2261
  //    rax ==  0  -  success
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2262
  //    rax == -1^K - failure, where K is partial transfer count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2263
  //
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2264
  address generate_checkcast_copy(const char *name, address *entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2265
                                  bool dest_uninitialized = false) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2266
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2267
    Label L_load_element, L_store_element, L_do_card_marks, L_done;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2268
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2269
    // Input registers (after setup_arg_regs)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2270
    const Register from        = rdi;   // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2271
    const Register to          = rsi;   // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2272
    const Register length      = rdx;   // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2273
    const Register ckoff       = rcx;   // super_check_offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2274
    const Register ckval       = r8;    // super_klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2275
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2276
    // Registers used as temps (r13, r14 are save-on-entry)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2277
    const Register end_from    = from;  // source array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2278
    const Register end_to      = r13;   // destination array end address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2279
    const Register count       = rdx;   // -(count_remaining)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2280
    const Register r14_length  = r14;   // saved copy of length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2281
    // End pointers are inclusive, and if length is not zero they point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2282
    // to the last unit copied:  end_to[0] := end_from[0]
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2283
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2284
    const Register rax_oop    = rax;    // actual oop copied
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2285
    const Register r11_klass  = r11;    // oop._klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2286
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2287
    //---------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2288
    // Assembler stub will be used for this call to arraycopy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2289
    // if the two arrays are subtypes of Object[] but the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2290
    // destination array type is not equal to or a supertype
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2291
    // of the source type.  Each element must be separately
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2292
    // checked.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2293
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2294
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2295
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2296
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2297
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2298
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2299
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2300
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2301
    // caller guarantees that the arrays really are different
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2302
    // otherwise, we would have to make conjoint checks
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2303
    { Label L;
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2304
      array_overlap_test(L, TIMES_OOP);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2305
      __ stop("checkcast_copy within a single array");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2306
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2307
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2308
#endif //ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2309
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2310
    setup_arg_regs(4); // from => rdi, to => rsi, length => rdx
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2311
                       // ckoff => rcx, ckval => r8
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2312
                       // r9 and r10 may be used to save non-volatile registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2313
#ifdef _WIN64
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2314
    // last argument (#4) is on stack on Win64
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2315
    __ movptr(ckval, Address(rsp, 6 * wordSize));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2316
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2317
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2318
    // Caller of this entry point must set up the argument registers.
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2319
    if (entry != NULL) {
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2320
      *entry = __ pc();
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2321
      BLOCK_COMMENT("Entry:");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2322
    }
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2323
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2324
    // allocate spill slots for r13, r14
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2325
    enum {
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2326
      saved_r13_offset,
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2327
      saved_r14_offset,
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2328
      saved_rbp_offset
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2329
    };
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2330
    __ subptr(rsp, saved_rbp_offset * wordSize);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2331
    __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2332
    __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2333
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2334
    // check that int operands are properly extended to size_t
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2335
    assert_clean_int(length, rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2336
    assert_clean_int(ckoff, rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2337
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2338
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2339
    BLOCK_COMMENT("assert consistent ckoff/ckval");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2340
    // The ckoff and ckval must be mutually consistent,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2341
    // even though caller generates both.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2342
    { Label L;
11430
718fc06da49a 7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents: 11194
diff changeset
  2343
      int sco_offset = in_bytes(Klass::super_check_offset_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2344
      __ cmpl(ckoff, Address(ckval, sco_offset));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2345
      __ jcc(Assembler::equal, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2346
      __ stop("super_check_offset inconsistent");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2347
      __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2348
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2349
#endif //ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2350
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2351
    // Loop-invariant addresses.  They are exclusive end pointers.
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2352
    Address end_from_addr(from, length, TIMES_OOP, 0);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2353
    Address   end_to_addr(to,   length, TIMES_OOP, 0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2354
    // Loop-variant addresses.  They assume post-incremented count < 0.
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2355
    Address from_element_addr(end_from, count, TIMES_OOP, 0);
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2356
    Address   to_element_addr(end_to,   count, TIMES_OOP, 0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2357
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2358
    gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2359
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2360
    // Copy from low to high addresses, indexed from the end of each array.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2361
    __ lea(end_from, end_from_addr);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2362
    __ lea(end_to,   end_to_addr);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2363
    __ movptr(r14_length, length);        // save a copy of the length
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2364
    assert(length == count, "");          // else fix next line:
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2365
    __ negptr(count);                     // negate and test the length
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2366
    __ jcc(Assembler::notZero, L_load_element);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2367
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2368
    // Empty array:  Nothing to do.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2369
    __ xorptr(rax, rax);                  // return 0 on (trivial) success
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2370
    __ jmp(L_done);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2371
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2372
    // ======== begin loop ========
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2373
    // (Loop is rotated; its entry is L_load_element.)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2374
    // Loop control:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2375
    //   for (count = -count; count != 0; count++)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2376
    // Base pointers src, dst are biased by 8*(count-1),to last element.
5249
5cac34e6fe54 6940701: Don't align loops in stubs for Niagara sparc
kvn
parents: 5050
diff changeset
  2377
    __ align(OptoLoopAlignment);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2378
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2379
    __ BIND(L_store_element);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2380
    __ store_heap_oop(to_element_addr, rax_oop);  // store the oop
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2381
    __ increment(count);               // increment the count toward zero
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2382
    __ jcc(Assembler::zero, L_do_card_marks);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2383
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2384
    // ======== loop entry is here ========
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2385
    __ BIND(L_load_element);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2386
    __ load_heap_oop(rax_oop, from_element_addr); // load the oop
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2387
    __ testptr(rax_oop, rax_oop);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2388
    __ jcc(Assembler::zero, L_store_element);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2389
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2390
    __ load_klass(r11_klass, rax_oop);// query the object klass
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2391
    generate_type_check(r11_klass, ckoff, ckval, L_store_element);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2392
    // ======== end loop ========
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2393
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2394
    // It was a real error; we must depend on the caller to finish the job.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2395
    // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2396
    // Emit GC store barriers for the oops we have copied (r14 + rdx),
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2397
    // and report their number to the caller.
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2398
    assert_different_registers(rax, r14_length, count, to, end_to, rcx, rscratch1);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2399
    Label L_post_barrier;
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2400
    __ addptr(r14_length, count);     // K = (original - remaining) oops
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2401
    __ movptr(rax, r14_length);       // save the value
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2402
    __ notptr(rax);                   // report (-1^K) to caller (does not affect flags)
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2403
    __ jccb(Assembler::notZero, L_post_barrier);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2404
    __ jmp(L_done); // K == 0, nothing was copied, skip post barrier
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2405
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2406
    // Come here on success only.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2407
    __ BIND(L_do_card_marks);
17622
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2408
    __ xorptr(rax, rax);              // return 0 on success
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2409
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2410
    __ BIND(L_post_barrier);
4037daf22a17 8010927: Kitchensink crashed with SIGSEGV, Problematic frame: v ~StubRoutines::checkcast_arraycopy
kvn
parents: 16624
diff changeset
  2411
    gen_write_ref_array_post_barrier(to, r14_length, rscratch1);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2412
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2413
    // Common exit point (success or failure).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2414
    __ BIND(L_done);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2415
    __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2416
    __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2417
    restore_arg_regs();
11194
ee1235a09fc3 7110489: C1: 64-bit tiered with ForceUnreachable: assert(reachable(src)) failed: Address should be reachable
never
parents: 11190
diff changeset
  2418
    inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); // Update counter after rscratch1 is free
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2419
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2420
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2421
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2422
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2423
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2424
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2425
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2426
  //  Generate 'unsafe' array copy stub
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2427
  //  Though just as safe as the other stubs, it takes an unscaled
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2428
  //  size_t argument instead of an element count.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2429
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2430
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2431
  //    c_rarg0   - source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2432
  //    c_rarg1   - destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2433
  //    c_rarg2   - byte count, treated as ssize_t, can be zero
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2434
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2435
  // Examines the alignment of the operands and dispatches
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2436
  // to a long, int, short, or byte copy loop.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2437
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2438
  address generate_unsafe_copy(const char *name,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2439
                               address byte_copy_entry, address short_copy_entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2440
                               address int_copy_entry, address long_copy_entry) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2441
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2442
    Label L_long_aligned, L_int_aligned, L_short_aligned;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2443
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2444
    // Input registers (before setup_arg_regs)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2445
    const Register from        = c_rarg0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2446
    const Register to          = c_rarg1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2447
    const Register size        = c_rarg2;  // byte count (size_t)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2448
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2449
    // Register used as a temp
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2450
    const Register bits        = rax;      // test copy of low bits
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2451
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2452
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2453
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2454
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2455
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2456
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2457
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2458
    // bump this on entry, not on exit:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2459
    inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2460
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2461
    __ mov(bits, from);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2462
    __ orptr(bits, to);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2463
    __ orptr(bits, size);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2464
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2465
    __ testb(bits, BytesPerLong-1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2466
    __ jccb(Assembler::zero, L_long_aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2467
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2468
    __ testb(bits, BytesPerInt-1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2469
    __ jccb(Assembler::zero, L_int_aligned);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2470
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2471
    __ testb(bits, BytesPerShort-1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2472
    __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2473
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2474
    __ BIND(L_short_aligned);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2475
    __ shrptr(size, LogBytesPerShort); // size => short_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2476
    __ jump(RuntimeAddress(short_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2477
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2478
    __ BIND(L_int_aligned);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2479
    __ shrptr(size, LogBytesPerInt); // size => int_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2480
    __ jump(RuntimeAddress(int_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2481
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2482
    __ BIND(L_long_aligned);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2483
    __ shrptr(size, LogBytesPerLong); // size => qword_count
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2484
    __ jump(RuntimeAddress(long_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2485
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2486
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2487
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2488
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2489
  // Perform range checks on the proposed arraycopy.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2490
  // Kills temp, but nothing else.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2491
  // Also, clean the sign bits of src_pos and dst_pos.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2492
  void arraycopy_range_checks(Register src,     // source array oop (c_rarg0)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2493
                              Register src_pos, // source position (c_rarg1)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2494
                              Register dst,     // destination array oo (c_rarg2)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2495
                              Register dst_pos, // destination position (c_rarg3)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2496
                              Register length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2497
                              Register temp,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2498
                              Label& L_failed) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2499
    BLOCK_COMMENT("arraycopy_range_checks:");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2500
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2501
    //  if (src_pos + length > arrayOop(src)->length())  FAIL;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2502
    __ movl(temp, length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2503
    __ addl(temp, src_pos);             // src_pos + length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2504
    __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes()));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2505
    __ jcc(Assembler::above, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2506
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2507
    //  if (dst_pos + length > arrayOop(dst)->length())  FAIL;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2508
    __ movl(temp, length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2509
    __ addl(temp, dst_pos);             // dst_pos + length
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2510
    __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes()));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2511
    __ jcc(Assembler::above, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2512
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2513
    // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2514
    // Move with sign extension can be used since they are positive.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2515
    __ movslq(src_pos, src_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2516
    __ movslq(dst_pos, dst_pos);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2517
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2518
    BLOCK_COMMENT("arraycopy_range_checks done");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2519
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2520
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2521
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2522
  //  Generate generic array copy stubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2523
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2524
  //  Input:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2525
  //    c_rarg0    -  src oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2526
  //    c_rarg1    -  src_pos (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2527
  //    c_rarg2    -  dst oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2528
  //    c_rarg3    -  dst_pos (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2529
  // not Win64
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2530
  //    c_rarg4    -  element count (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2531
  // Win64
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2532
  //    rsp+40     -  element count (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2533
  //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2534
  //  Output:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2535
  //    rax ==  0  -  success
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2536
  //    rax == -1^K - failure, where K is partial transfer count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2537
  //
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2538
  address generate_generic_copy(const char *name,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2539
                                address byte_copy_entry, address short_copy_entry,
8876
f82367de21f5 7026307: DEBUG MESSAGE: broken null klass on amd64
iveresov
parents: 8874
diff changeset
  2540
                                address int_copy_entry, address oop_copy_entry,
f82367de21f5 7026307: DEBUG MESSAGE: broken null klass on amd64
iveresov
parents: 8874
diff changeset
  2541
                                address long_copy_entry, address checkcast_copy_entry) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2542
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2543
    Label L_failed, L_failed_0, L_objArray;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2544
    Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2545
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2546
    // Input registers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2547
    const Register src        = c_rarg0;  // source array oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2548
    const Register src_pos    = c_rarg1;  // source position
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2549
    const Register dst        = c_rarg2;  // destination array oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2550
    const Register dst_pos    = c_rarg3;  // destination position
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2551
#ifndef _WIN64
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2552
    const Register length     = c_rarg4;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2553
#else
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2554
    const Address  length(rsp, 6 * wordSize);  // elements count is on stack on Win64
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2555
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2556
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2557
    { int modulus = CodeEntryAlignment;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2558
      int target  = modulus - 5; // 5 = sizeof jmp(L_failed)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2559
      int advance = target - (__ offset() % modulus);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2560
      if (advance < 0)  advance += modulus;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2561
      if (advance > 0)  __ nop(advance);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2562
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2563
    StubCodeMark mark(this, "StubRoutines", name);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2564
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2565
    // Short-hop target to L_failed.  Makes for denser prologue code.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2566
    __ BIND(L_failed_0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2567
    __ jmp(L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2568
    assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2569
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2570
    __ align(CodeEntryAlignment);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2571
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2572
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2573
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2574
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2575
    // bump this on entry, not on exit:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2576
    inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2577
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2578
    //-----------------------------------------------------------------------
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2579
    // Assembler stub will be used for this call to arraycopy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2580
    // if the following conditions are met:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2581
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2582
    // (1) src and dst must not be null.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2583
    // (2) src_pos must not be negative.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2584
    // (3) dst_pos must not be negative.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2585
    // (4) length  must not be negative.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2586
    // (5) src klass and dst klass should be the same and not NULL.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2587
    // (6) src and dst should be arrays.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2588
    // (7) src_pos + length must not exceed length of src.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2589
    // (8) dst_pos + length must not exceed length of dst.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2590
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2591
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2592
    //  if (src == NULL) return -1;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2593
    __ testptr(src, src);         // src oop
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2594
    size_t j1off = __ offset();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2595
    __ jccb(Assembler::zero, L_failed_0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2596
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2597
    //  if (src_pos < 0) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2598
    __ testl(src_pos, src_pos); // src_pos (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2599
    __ jccb(Assembler::negative, L_failed_0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2600
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2601
    //  if (dst == NULL) return -1;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2602
    __ testptr(dst, dst);         // dst oop
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2603
    __ jccb(Assembler::zero, L_failed_0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2604
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2605
    //  if (dst_pos < 0) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2606
    __ testl(dst_pos, dst_pos); // dst_pos (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2607
    size_t j4off = __ offset();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2608
    __ jccb(Assembler::negative, L_failed_0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2609
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2610
    // The first four tests are very dense code,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2611
    // but not quite dense enough to put four
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2612
    // jumps in a 16-byte instruction fetch buffer.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2613
    // That's good, because some branch predicters
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2614
    // do not like jumps so close together.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2615
    // Make sure of this.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2616
    guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2617
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2618
    // registers used as temp
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2619
    const Register r11_length    = r11; // elements count to copy
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2620
    const Register r10_src_klass = r10; // array klass
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2621
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2622
    //  if (length < 0) return -1;
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2623
    __ movl(r11_length, length);        // length (elements count, 32-bits value)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2624
    __ testl(r11_length, r11_length);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2625
    __ jccb(Assembler::negative, L_failed_0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2626
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2627
    __ load_klass(r10_src_klass, src);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2628
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2629
    //  assert(src->klass() != NULL);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2630
    {
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2631
      BLOCK_COMMENT("assert klasses not null {");
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2632
      Label L1, L2;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2633
      __ testptr(r10_src_klass, r10_src_klass);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2634
      __ jcc(Assembler::notZero, L2);   // it is broken if klass is NULL
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2635
      __ bind(L1);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2636
      __ stop("broken null klass");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2637
      __ bind(L2);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2638
      __ load_klass(rax, dst);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2639
      __ cmpq(rax, 0);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2640
      __ jcc(Assembler::equal, L1);     // this would be broken also
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2641
      BLOCK_COMMENT("} assert klasses not null done");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2642
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2643
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2644
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2645
    // Load layout helper (32-bits)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2646
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2647
    //  |array_tag|     | header_size | element_type |     |log2_element_size|
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2648
    // 32        30    24            16              8     2                 0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2649
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2650
    //   array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2651
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2652
11430
718fc06da49a 7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents: 11194
diff changeset
  2653
    const int lh_offset = in_bytes(Klass::layout_helper_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2654
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2655
    // Handle objArrays completely differently...
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2656
    const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2657
    __ cmpl(Address(r10_src_klass, lh_offset), objArray_lh);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2658
    __ jcc(Assembler::equal, L_objArray);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2659
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2660
    //  if (src->klass() != dst->klass()) return -1;
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2661
    __ load_klass(rax, dst);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2662
    __ cmpq(r10_src_klass, rax);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2663
    __ jcc(Assembler::notEqual, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2664
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2665
    const Register rax_lh = rax;  // layout helper
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2666
    __ movl(rax_lh, Address(r10_src_klass, lh_offset));
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2667
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2668
    //  if (!src->is_Array()) return -1;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2669
    __ cmpl(rax_lh, Klass::_lh_neutral_value);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2670
    __ jcc(Assembler::greaterEqual, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2671
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2672
    // At this point, it is known to be a typeArray (array_tag 0x3).
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2673
#ifdef ASSERT
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2674
    {
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2675
      BLOCK_COMMENT("assert primitive array {");
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2676
      Label L;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2677
      __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2678
      __ jcc(Assembler::greaterEqual, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2679
      __ stop("must be a primitive array");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2680
      __ bind(L);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2681
      BLOCK_COMMENT("} assert primitive array done");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2682
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2683
#endif
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2684
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2685
    arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2686
                           r10, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2687
13952
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2688
    // TypeArrayKlass
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2689
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2690
    // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2691
    // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2692
    //
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2693
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2694
    const Register r10_offset = r10;    // array offset
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2695
    const Register rax_elsize = rax_lh; // element size
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2696
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2697
    __ movl(r10_offset, rax_lh);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2698
    __ shrl(r10_offset, Klass::_lh_header_size_shift);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2699
    __ andptr(r10_offset, Klass::_lh_header_size_mask);   // array_offset
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2700
    __ addptr(src, r10_offset);           // src array offset
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2701
    __ addptr(dst, r10_offset);           // dst array offset
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2702
    BLOCK_COMMENT("choose copy loop based on element size");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2703
    __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2704
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2705
    // next registers should be set before the jump to corresponding stub
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2706
    const Register from     = c_rarg0;  // source array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2707
    const Register to       = c_rarg1;  // destination array address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2708
    const Register count    = c_rarg2;  // elements count
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2709
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2710
    // 'from', 'to', 'count' registers should be set in such order
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2711
    // since they are the same as 'src', 'src_pos', 'dst'.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2712
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2713
  __ BIND(L_copy_bytes);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2714
    __ cmpl(rax_elsize, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2715
    __ jccb(Assembler::notEqual, L_copy_shorts);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2716
    __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2717
    __ lea(to,   Address(dst, dst_pos, Address::times_1, 0));// dst_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2718
    __ movl2ptr(count, r11_length); // length
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2719
    __ jump(RuntimeAddress(byte_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2720
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2721
  __ BIND(L_copy_shorts);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2722
    __ cmpl(rax_elsize, LogBytesPerShort);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2723
    __ jccb(Assembler::notEqual, L_copy_ints);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2724
    __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2725
    __ lea(to,   Address(dst, dst_pos, Address::times_2, 0));// dst_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2726
    __ movl2ptr(count, r11_length); // length
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2727
    __ jump(RuntimeAddress(short_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2728
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2729
  __ BIND(L_copy_ints);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2730
    __ cmpl(rax_elsize, LogBytesPerInt);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2731
    __ jccb(Assembler::notEqual, L_copy_longs);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2732
    __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2733
    __ lea(to,   Address(dst, dst_pos, Address::times_4, 0));// dst_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2734
    __ movl2ptr(count, r11_length); // length
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2735
    __ jump(RuntimeAddress(int_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2736
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2737
  __ BIND(L_copy_longs);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2738
#ifdef ASSERT
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2739
    {
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2740
      BLOCK_COMMENT("assert long copy {");
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2741
      Label L;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2742
      __ cmpl(rax_elsize, LogBytesPerLong);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2743
      __ jcc(Assembler::equal, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2744
      __ stop("must be long copy, but elsize is wrong");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2745
      __ bind(L);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2746
      BLOCK_COMMENT("} assert long copy done");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2747
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2748
#endif
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2749
    __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2750
    __ lea(to,   Address(dst, dst_pos, Address::times_8, 0));// dst_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2751
    __ movl2ptr(count, r11_length); // length
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2752
    __ jump(RuntimeAddress(long_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2753
13952
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2754
    // ObjArrayKlass
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2755
  __ BIND(L_objArray);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2756
    // live at this point:  r10_src_klass, r11_length, src[_pos], dst[_pos]
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2757
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2758
    Label L_plain_copy, L_checkcast_copy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2759
    //  test array classes for subtyping
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2760
    __ load_klass(rax, dst);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2761
    __ cmpq(r10_src_klass, rax); // usual case is exact equality
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2762
    __ jcc(Assembler::notEqual, L_checkcast_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2763
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2764
    // Identically typed arrays can be copied without element-wise checks.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2765
    arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2766
                           r10, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2767
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2768
    __ lea(from, Address(src, src_pos, TIMES_OOP,
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2769
                 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2770
    __ lea(to,   Address(dst, dst_pos, TIMES_OOP,
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2771
                 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2772
    __ movl2ptr(count, r11_length); // length
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2773
  __ BIND(L_plain_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2774
    __ jump(RuntimeAddress(oop_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2775
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2776
  __ BIND(L_checkcast_copy);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2777
    // live at this point:  r10_src_klass, r11_length, rax (dst_klass)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2778
    {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2779
      // Before looking at dst.length, make sure dst is also an objArray.
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2780
      __ cmpl(Address(rax, lh_offset), objArray_lh);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2781
      __ jcc(Assembler::notEqual, L_failed);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2782
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2783
      // It is safe to examine both src.length and dst.length.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2784
      arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2785
                             rax, L_failed);
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2786
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2787
      const Register r11_dst_klass = r11;
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2788
      __ load_klass(r11_dst_klass, dst); // reload
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2789
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2790
      // Marshal the base address arguments now, freeing registers.
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2791
      __ lea(from, Address(src, src_pos, TIMES_OOP,
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2792
                   arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2793
      __ lea(to,   Address(dst, dst_pos, TIMES_OOP,
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2794
                   arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2795
      __ movl(count, length);           // length (reloaded)
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2796
      Register sco_temp = c_rarg3;      // this register is free now
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2797
      assert_different_registers(from, to, count, sco_temp,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2798
                                 r11_dst_klass, r10_src_klass);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2799
      assert_clean_int(count, sco_temp);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2800
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2801
      // Generate the type check.
11430
718fc06da49a 7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents: 11194
diff changeset
  2802
      const int sco_offset = in_bytes(Klass::super_check_offset_offset());
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2803
      __ movl(sco_temp, Address(r11_dst_klass, sco_offset));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2804
      assert_clean_int(sco_temp, rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2805
      generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2806
13952
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2807
      // Fetch destination element klass from the ObjArrayKlass header.
e3cf184080bc 8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents: 13728
diff changeset
  2808
      int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2809
      __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset));
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2810
      __ movl(  sco_temp,      Address(r11_dst_klass, sco_offset));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2811
      assert_clean_int(sco_temp, rax);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2812
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2813
      // the checkcast_copy loop needs two extra arguments:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2814
      assert(c_rarg3 == sco_temp, "#3 already in place");
7431
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2815
      // Set up arguments for checkcast_copy_entry.
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2816
      setup_arg_regs(4);
e9f07f8aef47 6998985: faulty generic arraycopy on windows x86_64: 4th arg overwritten with oop
twisti
parents: 7397
diff changeset
  2817
      __ movptr(r8, r11_dst_klass);  // dst.klass.element_klass, r8 is c_rarg4 on Linux/Solaris
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2818
      __ jump(RuntimeAddress(checkcast_copy_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2819
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2820
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2821
  __ BIND(L_failed);
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2822
    __ xorptr(rax, rax);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  2823
    __ notptr(rax); // return -1
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2824
    __ leave();   // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2825
    __ ret(0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2826
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2827
    return start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2828
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2829
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2830
  void generate_arraycopy_stubs() {
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2831
    address entry;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2832
    address entry_jbyte_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2833
    address entry_jshort_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2834
    address entry_jint_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2835
    address entry_oop_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2836
    address entry_jlong_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2837
    address entry_checkcast_arraycopy;
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2838
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2839
    StubRoutines::_jbyte_disjoint_arraycopy  = generate_disjoint_byte_copy(false, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2840
                                                                           "jbyte_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2841
    StubRoutines::_jbyte_arraycopy           = generate_conjoint_byte_copy(false, entry, &entry_jbyte_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2842
                                                                           "jbyte_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2843
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2844
    StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2845
                                                                            "jshort_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2846
    StubRoutines::_jshort_arraycopy          = generate_conjoint_short_copy(false, entry, &entry_jshort_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2847
                                                                            "jshort_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2848
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2849
    StubRoutines::_jint_disjoint_arraycopy   = generate_disjoint_int_oop_copy(false, false, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2850
                                                                              "jint_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2851
    StubRoutines::_jint_arraycopy            = generate_conjoint_int_oop_copy(false, false, entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2852
                                                                              &entry_jint_arraycopy, "jint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2853
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2854
    StubRoutines::_jlong_disjoint_arraycopy  = generate_disjoint_long_oop_copy(false, false, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2855
                                                                               "jlong_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2856
    StubRoutines::_jlong_arraycopy           = generate_conjoint_long_oop_copy(false, false, entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2857
                                                                               &entry_jlong_arraycopy, "jlong_arraycopy");
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2858
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2859
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2860
    if (UseCompressedOops) {
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2861
      StubRoutines::_oop_disjoint_arraycopy  = generate_disjoint_int_oop_copy(false, true, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2862
                                                                              "oop_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2863
      StubRoutines::_oop_arraycopy           = generate_conjoint_int_oop_copy(false, true, entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2864
                                                                              &entry_oop_arraycopy, "oop_arraycopy");
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2865
      StubRoutines::_oop_disjoint_arraycopy_uninit  = generate_disjoint_int_oop_copy(false, true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2866
                                                                                     "oop_disjoint_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2867
                                                                                     /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2868
      StubRoutines::_oop_arraycopy_uninit           = generate_conjoint_int_oop_copy(false, true, entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2869
                                                                                     NULL, "oop_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2870
                                                                                     /*dest_uninitialized*/true);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2871
    } else {
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2872
      StubRoutines::_oop_disjoint_arraycopy  = generate_disjoint_long_oop_copy(false, true, &entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2873
                                                                               "oop_disjoint_arraycopy");
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2874
      StubRoutines::_oop_arraycopy           = generate_conjoint_long_oop_copy(false, true, entry,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2875
                                                                               &entry_oop_arraycopy, "oop_arraycopy");
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2876
      StubRoutines::_oop_disjoint_arraycopy_uninit  = generate_disjoint_long_oop_copy(false, true, &entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2877
                                                                                      "oop_disjoint_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2878
                                                                                      /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2879
      StubRoutines::_oop_arraycopy_uninit           = generate_conjoint_long_oop_copy(false, true, entry,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2880
                                                                                      NULL, "oop_arraycopy_uninit",
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2881
                                                                                      /*dest_uninitialized*/true);
360
21d113ecbf6a 6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents: 189
diff changeset
  2882
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2883
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2884
    StubRoutines::_checkcast_arraycopy        = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2885
    StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2886
                                                                        /*dest_uninitialized*/true);
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2887
8487
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2888
    StubRoutines::_unsafe_arraycopy    = generate_unsafe_copy("unsafe_arraycopy",
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2889
                                                              entry_jbyte_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2890
                                                              entry_jshort_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2891
                                                              entry_jint_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2892
                                                              entry_jlong_arraycopy);
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2893
    StubRoutines::_generic_arraycopy   = generate_generic_copy("generic_arraycopy",
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2894
                                                               entry_jbyte_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2895
                                                               entry_jshort_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2896
                                                               entry_jint_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2897
                                                               entry_oop_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2898
                                                               entry_jlong_arraycopy,
bf96596f06d2 7020521: arraycopy stubs place prebarriers incorrectly
iveresov
parents: 7431
diff changeset
  2899
                                                               entry_checkcast_arraycopy);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2900
6433
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2901
    StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2902
    StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2903
    StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2904
    StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2905
    StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2906
    StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
b0e4fafdc38b 4809552: Optimize Arrays.fill(...)
never
parents: 5706
diff changeset
  2907
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2908
    // We don't generate specialized code for HeapWord-aligned source
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2909
    // arrays, so just use the code we've already generated
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2910
    StubRoutines::_arrayof_jbyte_disjoint_arraycopy  = StubRoutines::_jbyte_disjoint_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2911
    StubRoutines::_arrayof_jbyte_arraycopy           = StubRoutines::_jbyte_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2912
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2913
    StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2914
    StubRoutines::_arrayof_jshort_arraycopy          = StubRoutines::_jshort_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2915
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2916
    StubRoutines::_arrayof_jint_disjoint_arraycopy   = StubRoutines::_jint_disjoint_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2917
    StubRoutines::_arrayof_jint_arraycopy            = StubRoutines::_jint_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2918
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2919
    StubRoutines::_arrayof_jlong_disjoint_arraycopy  = StubRoutines::_jlong_disjoint_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2920
    StubRoutines::_arrayof_jlong_arraycopy           = StubRoutines::_jlong_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2921
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2922
    StubRoutines::_arrayof_oop_disjoint_arraycopy    = StubRoutines::_oop_disjoint_arraycopy;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2923
    StubRoutines::_arrayof_oop_arraycopy             = StubRoutines::_oop_arraycopy;
8498
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2924
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2925
    StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit    = StubRoutines::_oop_disjoint_arraycopy_uninit;
6398004126b9 6627983: G1: Bad oop deference during marking
iveresov
parents: 8487
diff changeset
  2926
    StubRoutines::_arrayof_oop_arraycopy_uninit             = StubRoutines::_oop_arraycopy_uninit;
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2927
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  2928
4645
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2929
  void generate_math_stubs() {
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2930
    {
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2931
      StubCodeMark mark(this, "StubRoutines", "log");
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2932
      StubRoutines::_intrinsic_log = (double (*)(double)) __ pc();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2933
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2934
      __ subq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2935
      __ movdbl(Address(rsp, 0), xmm0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2936
      __ fld_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2937
      __ flog();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2938
      __ fstp_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2939
      __ movdbl(xmm0, Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2940
      __ addq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2941
      __ ret(0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2942
    }
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2943
    {
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2944
      StubCodeMark mark(this, "StubRoutines", "log10");
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2945
      StubRoutines::_intrinsic_log10 = (double (*)(double)) __ pc();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2946
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2947
      __ subq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2948
      __ movdbl(Address(rsp, 0), xmm0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2949
      __ fld_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2950
      __ flog10();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2951
      __ fstp_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2952
      __ movdbl(xmm0, Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2953
      __ addq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2954
      __ ret(0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2955
    }
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2956
    {
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2957
      StubCodeMark mark(this, "StubRoutines", "sin");
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2958
      StubRoutines::_intrinsic_sin = (double (*)(double)) __ pc();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2959
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2960
      __ subq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2961
      __ movdbl(Address(rsp, 0), xmm0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2962
      __ fld_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2963
      __ trigfunc('s');
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2964
      __ fstp_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2965
      __ movdbl(xmm0, Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2966
      __ addq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2967
      __ ret(0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2968
    }
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2969
    {
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2970
      StubCodeMark mark(this, "StubRoutines", "cos");
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2971
      StubRoutines::_intrinsic_cos = (double (*)(double)) __ pc();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2972
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2973
      __ subq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2974
      __ movdbl(Address(rsp, 0), xmm0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2975
      __ fld_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2976
      __ trigfunc('c');
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2977
      __ fstp_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2978
      __ movdbl(xmm0, Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2979
      __ addq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2980
      __ ret(0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2981
    }
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2982
    {
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2983
      StubCodeMark mark(this, "StubRoutines", "tan");
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2984
      StubRoutines::_intrinsic_tan = (double (*)(double)) __ pc();
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2985
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2986
      __ subq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2987
      __ movdbl(Address(rsp, 0), xmm0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2988
      __ fld_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2989
      __ trigfunc('t');
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2990
      __ fstp_d(Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2991
      __ movdbl(xmm0, Address(rsp, 0));
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2992
      __ addq(rsp, 8);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2993
      __ ret(0);
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  2994
    }
12739
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  2995
    {
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  2996
      StubCodeMark mark(this, "StubRoutines", "exp");
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  2997
      StubRoutines::_intrinsic_exp = (double (*)(double)) __ pc();
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  2998
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  2999
      __ subq(rsp, 8);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3000
      __ movdbl(Address(rsp, 0), xmm0);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3001
      __ fld_d(Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3002
      __ exp_with_fallback(0);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3003
      __ fstp_d(Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3004
      __ movdbl(xmm0, Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3005
      __ addq(rsp, 8);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3006
      __ ret(0);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3007
    }
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3008
    {
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3009
      StubCodeMark mark(this, "StubRoutines", "pow");
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3010
      StubRoutines::_intrinsic_pow = (double (*)(double,double)) __ pc();
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3011
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3012
      __ subq(rsp, 8);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3013
      __ movdbl(Address(rsp, 0), xmm1);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3014
      __ fld_d(Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3015
      __ movdbl(Address(rsp, 0), xmm0);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3016
      __ fld_d(Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3017
      __ pow_with_fallback(0);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3018
      __ fstp_d(Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3019
      __ movdbl(xmm0, Address(rsp, 0));
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3020
      __ addq(rsp, 8);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3021
      __ ret(0);
09f26b73ae66 7133857: exp() and pow() should use the x87 ISA on x86
roland
parents: 11961
diff changeset
  3022
    }
4645
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  3023
  }
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  3024
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3025
  // AES intrinsic stubs
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3026
  enum {AESBlockSize = 16};
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3027
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3028
  address generate_key_shuffle_mask() {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3029
    __ align(16);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3030
    StubCodeMark mark(this, "StubRoutines", "key_shuffle_mask");
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3031
    address start = __ pc();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3032
    __ emit_data64( 0x0405060700010203, relocInfo::none );
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3033
    __ emit_data64( 0x0c0d0e0f08090a0b, relocInfo::none );
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3034
    return start;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3035
  }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3036
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3037
  // Utility routine for loading a 128-bit key word in little endian format
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3038
  // can optionally specify that the shuffle mask is already in an xmmregister
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3039
  void load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask=NULL) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3040
    __ movdqu(xmmdst, Address(key, offset));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3041
    if (xmm_shuf_mask != NULL) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3042
      __ pshufb(xmmdst, xmm_shuf_mask);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3043
    } else {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3044
      __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3045
    }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3046
  }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3047
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3048
  // Arguments:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3049
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3050
  // Inputs:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3051
  //   c_rarg0   - source byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3052
  //   c_rarg1   - destination byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3053
  //   c_rarg2   - K (key) in little endian int array
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3054
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3055
  address generate_aescrypt_encryptBlock() {
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3056
    assert(UseAES, "need AES instructions and misaligned SSE support");
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3057
    __ align(CodeEntryAlignment);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3058
    StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3059
    Label L_doLast;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3060
    address start = __ pc();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3061
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3062
    const Register from        = c_rarg0;  // source array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3063
    const Register to          = c_rarg1;  // destination array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3064
    const Register key         = c_rarg2;  // key array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3065
    const Register keylen      = rax;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3066
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3067
    const XMMRegister xmm_result = xmm0;
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3068
    const XMMRegister xmm_key_shuf_mask = xmm1;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3069
    // On win64 xmm6-xmm15 must be preserved so don't use them.
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3070
    const XMMRegister xmm_temp1  = xmm2;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3071
    const XMMRegister xmm_temp2  = xmm3;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3072
    const XMMRegister xmm_temp3  = xmm4;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3073
    const XMMRegister xmm_temp4  = xmm5;
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3074
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3075
    __ enter(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3076
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3077
    // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3078
    __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3079
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3080
    __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3081
    __ movdqu(xmm_result, Address(from, 0));  // get 16 bytes of input
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3082
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3083
    // For encryption, the java expanded key ordering is just what we need
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3084
    // we don't know if the key is aligned, hence not using load-execute form
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3085
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3086
    load_key(xmm_temp1, key, 0x00, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3087
    __ pxor(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3088
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3089
    load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3090
    load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3091
    load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3092
    load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3093
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3094
    __ aesenc(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3095
    __ aesenc(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3096
    __ aesenc(xmm_result, xmm_temp3);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3097
    __ aesenc(xmm_result, xmm_temp4);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3098
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3099
    load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3100
    load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3101
    load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3102
    load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3103
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3104
    __ aesenc(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3105
    __ aesenc(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3106
    __ aesenc(xmm_result, xmm_temp3);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3107
    __ aesenc(xmm_result, xmm_temp4);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3108
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3109
    load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3110
    load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3111
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3112
    __ cmpl(keylen, 44);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3113
    __ jccb(Assembler::equal, L_doLast);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3114
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3115
    __ aesenc(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3116
    __ aesenc(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3117
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3118
    load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3119
    load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3120
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3121
    __ cmpl(keylen, 52);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3122
    __ jccb(Assembler::equal, L_doLast);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3123
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3124
    __ aesenc(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3125
    __ aesenc(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3126
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3127
    load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3128
    load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3129
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3130
    __ BIND(L_doLast);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3131
    __ aesenc(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3132
    __ aesenclast(xmm_result, xmm_temp2);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3133
    __ movdqu(Address(to, 0), xmm_result);        // store the result
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3134
    __ xorptr(rax, rax); // return 0
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3135
    __ leave(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3136
    __ ret(0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3137
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3138
    return start;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3139
  }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3140
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3141
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3142
  // Arguments:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3143
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3144
  // Inputs:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3145
  //   c_rarg0   - source byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3146
  //   c_rarg1   - destination byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3147
  //   c_rarg2   - K (key) in little endian int array
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3148
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3149
  address generate_aescrypt_decryptBlock() {
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3150
    assert(UseAES, "need AES instructions and misaligned SSE support");
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3151
    __ align(CodeEntryAlignment);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3152
    StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3153
    Label L_doLast;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3154
    address start = __ pc();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3155
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3156
    const Register from        = c_rarg0;  // source array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3157
    const Register to          = c_rarg1;  // destination array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3158
    const Register key         = c_rarg2;  // key array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3159
    const Register keylen      = rax;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3160
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3161
    const XMMRegister xmm_result = xmm0;
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3162
    const XMMRegister xmm_key_shuf_mask = xmm1;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3163
    // On win64 xmm6-xmm15 must be preserved so don't use them.
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3164
    const XMMRegister xmm_temp1  = xmm2;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3165
    const XMMRegister xmm_temp2  = xmm3;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3166
    const XMMRegister xmm_temp3  = xmm4;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3167
    const XMMRegister xmm_temp4  = xmm5;
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3168
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3169
    __ enter(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3170
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3171
    // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3172
    __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3173
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3174
    __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3175
    __ movdqu(xmm_result, Address(from, 0));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3176
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3177
    // for decryption java expanded key ordering is rotated one position from what we want
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3178
    // so we start from 0x10 here and hit 0x00 last
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3179
    // we don't know if the key is aligned, hence not using load-execute form
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3180
    load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3181
    load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3182
    load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3183
    load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3184
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3185
    __ pxor  (xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3186
    __ aesdec(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3187
    __ aesdec(xmm_result, xmm_temp3);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3188
    __ aesdec(xmm_result, xmm_temp4);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3189
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3190
    load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3191
    load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3192
    load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3193
    load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3194
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3195
    __ aesdec(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3196
    __ aesdec(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3197
    __ aesdec(xmm_result, xmm_temp3);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3198
    __ aesdec(xmm_result, xmm_temp4);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3199
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3200
    load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3201
    load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3202
    load_key(xmm_temp3, key, 0x00, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3203
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3204
    __ cmpl(keylen, 44);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3205
    __ jccb(Assembler::equal, L_doLast);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3206
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3207
    __ aesdec(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3208
    __ aesdec(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3209
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3210
    load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3211
    load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3212
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3213
    __ cmpl(keylen, 52);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3214
    __ jccb(Assembler::equal, L_doLast);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3215
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3216
    __ aesdec(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3217
    __ aesdec(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3218
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3219
    load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3220
    load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3221
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3222
    __ BIND(L_doLast);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3223
    __ aesdec(xmm_result, xmm_temp1);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3224
    __ aesdec(xmm_result, xmm_temp2);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3225
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3226
    // for decryption the aesdeclast operation is always on key+0x00
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3227
    __ aesdeclast(xmm_result, xmm_temp3);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3228
    __ movdqu(Address(to, 0), xmm_result);  // store the result
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3229
    __ xorptr(rax, rax); // return 0
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3230
    __ leave(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3231
    __ ret(0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3232
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3233
    return start;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3234
  }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3235
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3236
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3237
  // Arguments:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3238
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3239
  // Inputs:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3240
  //   c_rarg0   - source byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3241
  //   c_rarg1   - destination byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3242
  //   c_rarg2   - K (key) in little endian int array
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3243
  //   c_rarg3   - r vector byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3244
  //   c_rarg4   - input length
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3245
  //
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3246
  // Output:
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3247
  //   rax       - input length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3248
  //
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3249
  address generate_cipherBlockChaining_encryptAESCrypt() {
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3250
    assert(UseAES, "need AES instructions and misaligned SSE support");
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3251
    __ align(CodeEntryAlignment);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3252
    StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3253
    address start = __ pc();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3254
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3255
    Label L_exit, L_key_192_256, L_key_256, L_loopTop_128, L_loopTop_192, L_loopTop_256;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3256
    const Register from        = c_rarg0;  // source array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3257
    const Register to          = c_rarg1;  // destination array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3258
    const Register key         = c_rarg2;  // key array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3259
    const Register rvec        = c_rarg3;  // r byte array initialized from initvector array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3260
                                           // and left with the results of the last encryption block
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3261
#ifndef _WIN64
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3262
    const Register len_reg     = c_rarg4;  // src len (must be multiple of blocksize 16)
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3263
#else
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3264
    const Address  len_mem(rbp, 6 * wordSize);  // length is on stack on Win64
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3265
    const Register len_reg     = r10;      // pick the first volatile windows register
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3266
#endif
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3267
    const Register pos         = rax;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3268
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3269
    // xmm register assignments for the loops below
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3270
    const XMMRegister xmm_result = xmm0;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3271
    const XMMRegister xmm_temp   = xmm1;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3272
    // keys 0-10 preloaded into xmm2-xmm12
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3273
    const int XMM_REG_NUM_KEY_FIRST = 2;
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3274
    const int XMM_REG_NUM_KEY_LAST  = 15;
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3275
    const XMMRegister xmm_key0   = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3276
    const XMMRegister xmm_key10  = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+10);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3277
    const XMMRegister xmm_key11  = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+11);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3278
    const XMMRegister xmm_key12  = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+12);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3279
    const XMMRegister xmm_key13  = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+13);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3280
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3281
    __ enter(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3282
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3283
#ifdef _WIN64
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3284
    // on win64, fill len_reg from stack position
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3285
    __ movl(len_reg, len_mem);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3286
    // save the xmm registers which must be preserved 6-15
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3287
    __ subptr(rsp, -rsp_after_call_off * wordSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3288
    for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3289
      __ movdqu(xmm_save(i), as_XMMRegister(i));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3290
    }
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3291
#else
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3292
    __ push(len_reg); // Save
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3293
#endif
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3294
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3295
    const XMMRegister xmm_key_shuf_mask = xmm_temp;  // used temporarily to swap key bytes up front
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3296
    __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3297
    // load up xmm regs xmm2 thru xmm12 with key 0x00 - 0xa0
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3298
    for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x00; rnum <= XMM_REG_NUM_KEY_FIRST+10; rnum++) {
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3299
      load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3300
      offset += 0x10;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3301
    }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3302
    __ movdqu(xmm_result, Address(rvec, 0x00));   // initialize xmm_result with r vec
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3303
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3304
    // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3305
    __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3306
    __ cmpl(rax, 44);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3307
    __ jcc(Assembler::notEqual, L_key_192_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3308
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3309
    // 128 bit code follows here
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3310
    __ movptr(pos, 0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3311
    __ align(OptoLoopAlignment);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3312
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3313
    __ BIND(L_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3314
    __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0));   // get next 16 bytes of input
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3315
    __ pxor  (xmm_result, xmm_temp);               // xor with the current r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3316
    __ pxor  (xmm_result, xmm_key0);               // do the aes rounds
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3317
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 9; rnum++) {
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3318
      __ aesenc(xmm_result, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3319
    }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3320
    __ aesenclast(xmm_result, xmm_key10);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3321
    __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result);     // store into the next 16 bytes of output
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3322
    // no need to store r to memory until we exit
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3323
    __ addptr(pos, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3324
    __ subptr(len_reg, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3325
    __ jcc(Assembler::notEqual, L_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3326
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3327
    __ BIND(L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3328
    __ movdqu(Address(rvec, 0), xmm_result);     // final value of r stored in rvec of CipherBlockChaining object
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3329
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3330
#ifdef _WIN64
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3331
    // restore xmm regs belonging to calling function
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3332
    for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3333
      __ movdqu(as_XMMRegister(i), xmm_save(i));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3334
    }
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3335
    __ movl(rax, len_mem);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3336
#else
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3337
    __ pop(rax); // return length
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3338
#endif
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3339
    __ leave(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3340
    __ ret(0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3341
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3342
    __ BIND(L_key_192_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3343
    // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3344
    load_key(xmm_key11, key, 0xb0, xmm_key_shuf_mask);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3345
    load_key(xmm_key12, key, 0xc0, xmm_key_shuf_mask);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3346
    __ cmpl(rax, 52);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3347
    __ jcc(Assembler::notEqual, L_key_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3348
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3349
    // 192-bit code follows here (could be changed to use more xmm registers)
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3350
    __ movptr(pos, 0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3351
    __ align(OptoLoopAlignment);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3352
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3353
    __ BIND(L_loopTop_192);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3354
    __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0));   // get next 16 bytes of input
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3355
    __ pxor  (xmm_result, xmm_temp);               // xor with the current r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3356
    __ pxor  (xmm_result, xmm_key0);               // do the aes rounds
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3357
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum  <= XMM_REG_NUM_KEY_FIRST + 11; rnum++) {
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3358
      __ aesenc(xmm_result, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3359
    }
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3360
    __ aesenclast(xmm_result, xmm_key12);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3361
    __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result);     // store into the next 16 bytes of output
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3362
    // no need to store r to memory until we exit
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3363
    __ addptr(pos, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3364
    __ subptr(len_reg, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3365
    __ jcc(Assembler::notEqual, L_loopTop_192);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3366
    __ jmp(L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3367
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3368
    __ BIND(L_key_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3369
    // 256-bit code follows here (could be changed to use more xmm registers)
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3370
    load_key(xmm_key13, key, 0xd0, xmm_key_shuf_mask);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3371
    __ movptr(pos, 0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3372
    __ align(OptoLoopAlignment);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3373
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3374
    __ BIND(L_loopTop_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3375
    __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0));   // get next 16 bytes of input
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3376
    __ pxor  (xmm_result, xmm_temp);               // xor with the current r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3377
    __ pxor  (xmm_result, xmm_key0);               // do the aes rounds
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3378
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum  <= XMM_REG_NUM_KEY_FIRST + 13; rnum++) {
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3379
      __ aesenc(xmm_result, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3380
    }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3381
    load_key(xmm_temp, key, 0xe0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3382
    __ aesenclast(xmm_result, xmm_temp);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3383
    __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result);     // store into the next 16 bytes of output
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3384
    // no need to store r to memory until we exit
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3385
    __ addptr(pos, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3386
    __ subptr(len_reg, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3387
    __ jcc(Assembler::notEqual, L_loopTop_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3388
    __ jmp(L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3389
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3390
    return start;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3391
  }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3392
18740
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3393
  // Safefetch stubs.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3394
  void generate_safefetch(const char* name, int size, address* entry,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3395
                          address* fault_pc, address* continuation_pc) {
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3396
    // safefetch signatures:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3397
    //   int      SafeFetch32(int*      adr, int      errValue);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3398
    //   intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3399
    //
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3400
    // arguments:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3401
    //   c_rarg0 = adr
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3402
    //   c_rarg1 = errValue
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3403
    //
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3404
    // result:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3405
    //   PPC_RET  = *adr or errValue
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3406
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3407
    StubCodeMark mark(this, "StubRoutines", name);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3408
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3409
    // Entry point, pc or function descriptor.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3410
    *entry = __ pc();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3411
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3412
    // Load *adr into c_rarg1, may fault.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3413
    *fault_pc = __ pc();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3414
    switch (size) {
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3415
      case 4:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3416
        // int32_t
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3417
        __ movl(c_rarg1, Address(c_rarg0, 0));
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3418
        break;
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3419
      case 8:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3420
        // int64_t
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3421
        __ movq(c_rarg1, Address(c_rarg0, 0));
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3422
        break;
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3423
      default:
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3424
        ShouldNotReachHere();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3425
    }
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3426
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3427
    // return errValue or *adr
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3428
    *continuation_pc = __ pc();
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3429
    __ movq(rax, c_rarg1);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3430
    __ ret(0);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  3431
  }
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3432
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3433
  // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3434
  // to hide instruction latency
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3435
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3436
  // Arguments:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3437
  //
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3438
  // Inputs:
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3439
  //   c_rarg0   - source byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3440
  //   c_rarg1   - destination byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3441
  //   c_rarg2   - K (key) in little endian int array
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3442
  //   c_rarg3   - r vector byte array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3443
  //   c_rarg4   - input length
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3444
  //
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3445
  // Output:
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3446
  //   rax       - input length
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3447
  //
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3448
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3449
  address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3450
    assert(UseAES, "need AES instructions and misaligned SSE support");
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3451
    __ align(CodeEntryAlignment);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3452
    StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3453
    address start = __ pc();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3454
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3455
    Label L_exit, L_key_192_256, L_key_256;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3456
    Label L_singleBlock_loopTop_128, L_multiBlock_loopTop_128;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3457
    Label L_singleBlock_loopTop_192, L_singleBlock_loopTop_256;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3458
    const Register from        = c_rarg0;  // source array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3459
    const Register to          = c_rarg1;  // destination array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3460
    const Register key         = c_rarg2;  // key array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3461
    const Register rvec        = c_rarg3;  // r byte array initialized from initvector array address
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3462
                                           // and left with the results of the last encryption block
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3463
#ifndef _WIN64
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3464
    const Register len_reg     = c_rarg4;  // src len (must be multiple of blocksize 16)
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3465
#else
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3466
    const Address  len_mem(rbp, 6 * wordSize);  // length is on stack on Win64
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3467
    const Register len_reg     = r10;      // pick the first volatile windows register
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3468
#endif
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3469
    const Register pos         = rax;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3470
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3471
    // keys 0-10 preloaded into xmm2-xmm12
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3472
    const int XMM_REG_NUM_KEY_FIRST = 5;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3473
    const int XMM_REG_NUM_KEY_LAST  = 15;
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3474
    const XMMRegister xmm_key_first = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3475
    const XMMRegister xmm_key_last  = as_XMMRegister(XMM_REG_NUM_KEY_LAST);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3476
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3477
    __ enter(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3478
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3479
#ifdef _WIN64
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3480
    // on win64, fill len_reg from stack position
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3481
    __ movl(len_reg, len_mem);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3482
    // save the xmm registers which must be preserved 6-15
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3483
    __ subptr(rsp, -rsp_after_call_off * wordSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3484
    for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3485
      __ movdqu(xmm_save(i), as_XMMRegister(i));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3486
    }
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3487
#else
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3488
    __ push(len_reg); // Save
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3489
#endif
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3490
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3491
    // the java expanded key ordering is rotated one position from what we want
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3492
    // so we start from 0x10 here and hit 0x00 last
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3493
    const XMMRegister xmm_key_shuf_mask = xmm1;  // used temporarily to swap key bytes up front
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3494
    __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3495
    // load up xmm regs 5 thru 15 with key 0x10 - 0xa0 - 0x00
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3496
    for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x10; rnum < XMM_REG_NUM_KEY_LAST; rnum++) {
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3497
      load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3498
      offset += 0x10;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3499
    }
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3500
    load_key(xmm_key_last, key, 0x00, xmm_key_shuf_mask);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3501
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3502
    const XMMRegister xmm_prev_block_cipher = xmm1;  // holds cipher of previous block
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3503
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3504
    // registers holding the four results in the parallelized loop
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3505
    const XMMRegister xmm_result0 = xmm0;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3506
    const XMMRegister xmm_result1 = xmm2;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3507
    const XMMRegister xmm_result2 = xmm3;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3508
    const XMMRegister xmm_result3 = xmm4;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3509
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3510
    __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00));   // initialize with initial rvec
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3511
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3512
    // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3513
    __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3514
    __ cmpl(rax, 44);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3515
    __ jcc(Assembler::notEqual, L_key_192_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3516
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3517
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3518
    // 128-bit code follows here, parallelized
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3519
    __ movptr(pos, 0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3520
    __ align(OptoLoopAlignment);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3521
    __ BIND(L_multiBlock_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3522
    __ cmpptr(len_reg, 4*AESBlockSize);           // see if at least 4 blocks left
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3523
    __ jcc(Assembler::less, L_singleBlock_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3524
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3525
    __ movdqu(xmm_result0, Address(from, pos, Address::times_1, 0*AESBlockSize));   // get next 4 blocks into xmmresult registers
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3526
    __ movdqu(xmm_result1, Address(from, pos, Address::times_1, 1*AESBlockSize));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3527
    __ movdqu(xmm_result2, Address(from, pos, Address::times_1, 2*AESBlockSize));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3528
    __ movdqu(xmm_result3, Address(from, pos, Address::times_1, 3*AESBlockSize));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3529
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3530
#define DoFour(opc, src_reg)                    \
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3531
    __ opc(xmm_result0, src_reg);               \
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3532
    __ opc(xmm_result1, src_reg);               \
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3533
    __ opc(xmm_result2, src_reg);               \
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3534
    __ opc(xmm_result3, src_reg);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3535
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3536
    DoFour(pxor, xmm_key_first);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3537
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum  <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3538
      DoFour(aesdec, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3539
    }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3540
    DoFour(aesdeclast, xmm_key_last);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3541
    // for each result, xor with the r vector of previous cipher block
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3542
    __ pxor(xmm_result0, xmm_prev_block_cipher);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3543
    __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 0*AESBlockSize));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3544
    __ pxor(xmm_result1, xmm_prev_block_cipher);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3545
    __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 1*AESBlockSize));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3546
    __ pxor(xmm_result2, xmm_prev_block_cipher);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3547
    __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 2*AESBlockSize));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3548
    __ pxor(xmm_result3, xmm_prev_block_cipher);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3549
    __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 3*AESBlockSize));   // this will carry over to next set of blocks
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3550
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3551
    __ movdqu(Address(to, pos, Address::times_1, 0*AESBlockSize), xmm_result0);     // store 4 results into the next 64 bytes of output
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3552
    __ movdqu(Address(to, pos, Address::times_1, 1*AESBlockSize), xmm_result1);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3553
    __ movdqu(Address(to, pos, Address::times_1, 2*AESBlockSize), xmm_result2);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3554
    __ movdqu(Address(to, pos, Address::times_1, 3*AESBlockSize), xmm_result3);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3555
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3556
    __ addptr(pos, 4*AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3557
    __ subptr(len_reg, 4*AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3558
    __ jmp(L_multiBlock_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3559
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3560
    // registers used in the non-parallelized loops
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3561
    // xmm register assignments for the loops below
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3562
    const XMMRegister xmm_result = xmm0;
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3563
    const XMMRegister xmm_prev_block_cipher_save = xmm2;
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3564
    const XMMRegister xmm_key11 = xmm3;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3565
    const XMMRegister xmm_key12 = xmm4;
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3566
    const XMMRegister xmm_temp  = xmm4;
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3567
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3568
    __ align(OptoLoopAlignment);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3569
    __ BIND(L_singleBlock_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3570
    __ cmpptr(len_reg, 0);           // any blocks left??
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3571
    __ jcc(Assembler::equal, L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3572
    __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0));   // get next 16 bytes of cipher input
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3573
    __ movdqa(xmm_prev_block_cipher_save, xmm_result);              // save for next r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3574
    __ pxor  (xmm_result, xmm_key_first);               // do the aes dec rounds
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3575
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum  <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3576
      __ aesdec(xmm_result, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3577
    }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3578
    __ aesdeclast(xmm_result, xmm_key_last);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3579
    __ pxor  (xmm_result, xmm_prev_block_cipher);               // xor with the current r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3580
    __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result);     // store into the next 16 bytes of output
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3581
    // no need to store r to memory until we exit
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3582
    __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save);              // set up next r vector with cipher input from this block
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3583
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3584
    __ addptr(pos, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3585
    __ subptr(len_reg, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3586
    __ jmp(L_singleBlock_loopTop_128);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3587
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3588
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3589
    __ BIND(L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3590
    __ movdqu(Address(rvec, 0), xmm_prev_block_cipher);     // final value of r stored in rvec of CipherBlockChaining object
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3591
#ifdef _WIN64
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3592
    // restore regs belonging to calling function
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3593
    for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3594
      __ movdqu(as_XMMRegister(i), xmm_save(i));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3595
    }
22505
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3596
    __ movl(rax, len_mem);
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3597
#else
4523090c9674 8002074: Support for AES on SPARC
kvn
parents: 19319
diff changeset
  3598
    __ pop(rax); // return length
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3599
#endif
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3600
    __ leave(); // required for proper stackwalking of RuntimeStub frame
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3601
    __ ret(0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3602
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3603
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3604
    __ BIND(L_key_192_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3605
    // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3606
    load_key(xmm_key11, key, 0xb0);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3607
    __ cmpl(rax, 52);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3608
    __ jcc(Assembler::notEqual, L_key_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3609
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3610
    // 192-bit code follows here (could be optimized to use parallelism)
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3611
    load_key(xmm_key12, key, 0xc0);     // 192-bit key goes up to c0
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3612
    __ movptr(pos, 0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3613
    __ align(OptoLoopAlignment);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3614
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3615
    __ BIND(L_singleBlock_loopTop_192);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3616
    __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0));   // get next 16 bytes of cipher input
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3617
    __ movdqa(xmm_prev_block_cipher_save, xmm_result);              // save for next r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3618
    __ pxor  (xmm_result, xmm_key_first);               // do the aes dec rounds
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3619
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3620
      __ aesdec(xmm_result, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3621
    }
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3622
    __ aesdec(xmm_result, xmm_key11);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3623
    __ aesdec(xmm_result, xmm_key12);
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3624
    __ aesdeclast(xmm_result, xmm_key_last);                    // xmm15 always came from key+0
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3625
    __ pxor  (xmm_result, xmm_prev_block_cipher);               // xor with the current r vector
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3626
    __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result);  // store into the next 16 bytes of output
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3627
    // no need to store r to memory until we exit
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3628
    __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save);  // set up next r vector with cipher input from this block
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3629
    __ addptr(pos, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3630
    __ subptr(len_reg, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3631
    __ jcc(Assembler::notEqual,L_singleBlock_loopTop_192);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3632
    __ jmp(L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3633
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3634
    __ BIND(L_key_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3635
    // 256-bit code follows here (could be optimized to use parallelism)
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3636
    __ movptr(pos, 0);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3637
    __ align(OptoLoopAlignment);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3638
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3639
    __ BIND(L_singleBlock_loopTop_256);
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3640
    __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3641
    __ movdqa(xmm_prev_block_cipher_save, xmm_result);              // save for next r vector
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3642
    __ pxor  (xmm_result, xmm_key_first);               // do the aes dec rounds
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3643
    for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3644
      __ aesdec(xmm_result, as_XMMRegister(rnum));
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3645
    }
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3646
    __ aesdec(xmm_result, xmm_key11);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3647
    load_key(xmm_temp, key, 0xc0);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3648
    __ aesdec(xmm_result, xmm_temp);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3649
    load_key(xmm_temp, key, 0xd0);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3650
    __ aesdec(xmm_result, xmm_temp);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3651
    load_key(xmm_temp, key, 0xe0);     // 256-bit key goes up to e0
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3652
    __ aesdec(xmm_result, xmm_temp);
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3653
    __ aesdeclast(xmm_result, xmm_key_last);          // xmm15 came from key+0
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3654
    __ pxor  (xmm_result, xmm_prev_block_cipher);               // xor with the current r vector
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3655
    __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result);  // store into the next 16 bytes of output
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3656
    // no need to store r to memory until we exit
14834
f29c91f2f22b 8004835: Improve AES intrinsics on x86
kvn
parents: 14633
diff changeset
  3657
    __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save);  // set up next r vector with cipher input from this block
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3658
    __ addptr(pos, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3659
    __ subptr(len_reg, AESBlockSize);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3660
    __ jcc(Assembler::notEqual,L_singleBlock_loopTop_256);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3661
    __ jmp(L_exit);
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3662
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3663
    return start;
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3664
  }
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3665
18507
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3666
  /**
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3667
   *  Arguments:
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3668
   *
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3669
   * Inputs:
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3670
   *   c_rarg0   - int crc
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3671
   *   c_rarg1   - byte* buf
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3672
   *   c_rarg2   - int length
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3673
   *
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3674
   * Ouput:
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3675
   *       rax   - int crc result
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3676
   */
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3677
  address generate_updateBytesCRC32() {
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3678
    assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3679
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3680
    __ align(CodeEntryAlignment);
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3681
    StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3682
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3683
    address start = __ pc();
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3684
    // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3685
    // Unix:  rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3686
    // rscratch1: r10
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3687
    const Register crc   = c_rarg0;  // crc
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3688
    const Register buf   = c_rarg1;  // source java byte array address
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3689
    const Register len   = c_rarg2;  // length
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3690
    const Register table = c_rarg3;  // crc_table address (reuse register)
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3691
    const Register tmp   = r11;
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3692
    assert_different_registers(crc, buf, len, table, tmp, rax);
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3693
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3694
    BLOCK_COMMENT("Entry:");
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3695
    __ enter(); // required for proper stackwalking of RuntimeStub frame
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3696
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3697
    __ kernel_crc32(crc, buf, len, table, tmp);
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3698
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3699
    __ movl(rax, crc);
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3700
    __ leave(); // required for proper stackwalking of RuntimeStub frame
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3701
    __ ret(0);
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3702
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3703
    return start;
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3704
  }
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3705
26434
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3706
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3707
  /**
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3708
   *  Arguments:
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3709
   *
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3710
   *  Input:
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3711
   *    c_rarg0   - x address
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3712
   *    c_rarg1   - x length
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3713
   *    c_rarg2   - y address
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3714
   *    c_rarg3   - y lenth
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3715
   * not Win64
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3716
   *    c_rarg4   - z address
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3717
   *    c_rarg5   - z length
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3718
   * Win64
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3719
   *    rsp+40    - z address
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3720
   *    rsp+48    - z length
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3721
   */
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3722
  address generate_multiplyToLen() {
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3723
    __ align(CodeEntryAlignment);
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3724
    StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3725
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3726
    address start = __ pc();
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3727
    // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3728
    // Unix:  rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3729
    const Register x     = rdi;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3730
    const Register xlen  = rax;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3731
    const Register y     = rsi;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3732
    const Register ylen  = rcx;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3733
    const Register z     = r8;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3734
    const Register zlen  = r11;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3735
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3736
    // Next registers will be saved on stack in multiply_to_len().
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3737
    const Register tmp1  = r12;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3738
    const Register tmp2  = r13;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3739
    const Register tmp3  = r14;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3740
    const Register tmp4  = r15;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3741
    const Register tmp5  = rbx;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3742
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3743
    BLOCK_COMMENT("Entry:");
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3744
    __ enter(); // required for proper stackwalking of RuntimeStub frame
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3745
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3746
#ifndef _WIN64
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3747
    __ movptr(zlen, r9); // Save r9 in r11 - zlen
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3748
#endif
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3749
    setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3750
                       // ylen => rcx, z => r8, zlen => r11
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3751
                       // r9 and r10 may be used to save non-volatile registers
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3752
#ifdef _WIN64
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3753
    // last 2 arguments (#4, #5) are on stack on Win64
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3754
    __ movptr(z, Address(rsp, 6 * wordSize));
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3755
    __ movptr(zlen, Address(rsp, 7 * wordSize));
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3756
#endif
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3757
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3758
    __ movptr(xlen, rsi);
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3759
    __ movptr(y,    rdx);
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3760
    __ multiply_to_len(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5);
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3761
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3762
    restore_arg_regs();
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3763
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3764
    __ leave(); // required for proper stackwalking of RuntimeStub frame
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3765
    __ ret(0);
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3766
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3767
    return start;
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3768
  }
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  3769
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3770
#undef __
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3771
#define __ masm->
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3772
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3773
  // Continuation point for throwing of implicit exceptions that are
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3774
  // not handled in the current activation. Fabricates an exception
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3775
  // oop and initiates normal exception dispatching in this
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3776
  // frame. Since we need to preserve callee-saved values (currently
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3777
  // only for C2, but done for C1 as well) we need a callee-saved oop
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3778
  // map and therefore have to make these stubs into RuntimeStubs
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3779
  // rather than BufferBlobs.  If the compiler needs all registers to
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3780
  // be preserved between the fault point and the exception handler
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3781
  // then it must assume responsibility for that in
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3782
  // AbstractCompiler::continuation_for_implicit_null_exception or
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3783
  // continuation_for_implicit_division_by_zero_exception. All other
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3784
  // implicit exceptions (e.g., NullPointerException or
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3785
  // AbstractMethodError on entry) are either at call sites or
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3786
  // otherwise assume that stack unwinding will be initiated, so
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3787
  // caller saved registers were assumed volatile in the compiler.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3788
  address generate_throw_exception(const char* name,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3789
                                   address runtime_entry,
10004
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3790
                                   Register arg1 = noreg,
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3791
                                   Register arg2 = noreg) {
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3792
    // Information about frame layout at time of blocking runtime call.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3793
    // Note that we only have to preserve callee-saved registers since
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3794
    // the compilers are responsible for supplying a continuation point
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3795
    // if they expect all registers to be preserved.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3796
    enum layout {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3797
      rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3798
      rbp_off2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3799
      return_off,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3800
      return_off2,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3801
      framesize // inclusive of return address
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3802
    };
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3803
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3804
    int insts_size = 512;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3805
    int locs_size  = 64;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3806
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3807
    CodeBuffer code(name, insts_size, locs_size);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3808
    OopMapSet* oop_maps  = new OopMapSet();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3809
    MacroAssembler* masm = new MacroAssembler(&code);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3810
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3811
    address start = __ pc();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3812
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3813
    // This is an inlined and slightly modified version of call_VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3814
    // which has the ability to fetch the return PC out of
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3815
    // thread-local storage and also sets up last_Java_sp slightly
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3816
    // differently than the real call_VM
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3817
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3818
    __ enter(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3819
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3820
    assert(is_even(framesize/2), "sp not 16-byte aligned");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3821
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3822
    // return address and rbp are already in place
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3823
    __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3824
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3825
    int frame_complete = __ pc() - start;
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3826
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3827
    // Set up last_Java_sp and last_Java_fp
11725
49d1e1f1421f 7119286: JSR292: SIGSEGV in JNIHandleBlock::release_block(JNIHandleBlock*, Thread*)+0x3c
roland
parents: 11439
diff changeset
  3828
    address the_pc = __ pc();
49d1e1f1421f 7119286: JSR292: SIGSEGV in JNIHandleBlock::release_block(JNIHandleBlock*, Thread*)+0x3c
roland
parents: 11439
diff changeset
  3829
    __ set_last_Java_frame(rsp, rbp, the_pc);
49d1e1f1421f 7119286: JSR292: SIGSEGV in JNIHandleBlock::release_block(JNIHandleBlock*, Thread*)+0x3c
roland
parents: 11439
diff changeset
  3830
    __ andptr(rsp, -(StackAlignmentInBytes));    // Align stack
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3831
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3832
    // Call runtime
10004
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3833
    if (arg1 != noreg) {
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3834
      assert(arg2 != c_rarg1, "clobbered");
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3835
      __ movptr(c_rarg1, arg1);
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3836
    }
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3837
    if (arg2 != noreg) {
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3838
      __ movptr(c_rarg2, arg2);
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3839
    }
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3840
    __ movptr(c_rarg0, r15_thread);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3841
    BLOCK_COMMENT("call runtime_entry");
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3842
    __ call(RuntimeAddress(runtime_entry));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3843
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3844
    // Generate oop map
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3845
    OopMap* map = new OopMap(framesize, 0);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3846
11785
10f778832544 7144405: JumbleGC002 assert(m->offset() == pc_offset) failed: oopmap not found
roland
parents: 11725
diff changeset
  3847
    oop_maps->add_gc_map(the_pc - start, map);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3848
11725
49d1e1f1421f 7119286: JSR292: SIGSEGV in JNIHandleBlock::release_block(JNIHandleBlock*, Thread*)+0x3c
roland
parents: 11439
diff changeset
  3849
    __ reset_last_Java_frame(true, true);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3850
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3851
    __ leave(); // required for proper stackwalking of RuntimeStub frame
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3852
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3853
    // check for pending exceptions
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3854
#ifdef ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3855
    Label L;
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3856
    __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()),
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3857
            (int32_t) NULL_WORD);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3858
    __ jcc(Assembler::notEqual, L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3859
    __ should_not_reach_here();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3860
    __ bind(L);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3861
#endif // ASSERT
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3862
    __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3863
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3864
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3865
    // codeBlob framesize is in words (not VMRegImpl::slot_size)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3866
    RuntimeStub* stub =
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3867
      RuntimeStub::new_runtime_stub(name,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3868
                                    &code,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3869
                                    frame_complete,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3870
                                    (framesize >> (LogBytesPerWord - LogBytesPerInt)),
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3871
                                    oop_maps, false);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3872
    return stub->entry_point();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3873
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3874
18958
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3875
  void create_control_words() {
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3876
    // Round to nearest, 53-bit mode, exceptions masked
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3877
    StubRoutines::_fpu_cntrl_wrd_std   = 0x027F;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3878
    // Round to zero, 53-bit mode, exception mased
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3879
    StubRoutines::_fpu_cntrl_wrd_trunc = 0x0D7F;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3880
    // Round to nearest, 24-bit mode, exceptions masked
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3881
    StubRoutines::_fpu_cntrl_wrd_24    = 0x007F;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3882
    // Round to nearest, 64-bit mode, exceptions masked
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3883
    StubRoutines::_fpu_cntrl_wrd_64    = 0x037F;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3884
    // Round to nearest, 64-bit mode, exceptions masked
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3885
    StubRoutines::_mxcsr_std           = 0x1F80;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3886
    // Note: the following two constants are 80-bit values
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3887
    //       layout is critical for correct loading by FPU.
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3888
    // Bias for strict fp multiply/divide
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3889
    StubRoutines::_fpu_subnormal_bias1[0]= 0x00000000; // 2^(-15360) == 0x03ff 8000 0000 0000 0000
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3890
    StubRoutines::_fpu_subnormal_bias1[1]= 0x80000000;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3891
    StubRoutines::_fpu_subnormal_bias1[2]= 0x03ff;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3892
    // Un-Bias for strict fp multiply/divide
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3893
    StubRoutines::_fpu_subnormal_bias2[0]= 0x00000000; // 2^(+15360) == 0x7bff 8000 0000 0000 0000
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3894
    StubRoutines::_fpu_subnormal_bias2[1]= 0x80000000;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3895
    StubRoutines::_fpu_subnormal_bias2[2]= 0x7bff;
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3896
  }
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3897
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3898
  // Initialization
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3899
  void generate_initial() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3900
    // Generates all stubs and initializes the entry points
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3901
18958
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3902
    // This platform-specific settings are needed by generate_call_stub()
7fdf6d23667d 8020433: Crash when using -XX:+RestoreMXCSROnJNICalls
kvn
parents: 18507
diff changeset
  3903
    create_control_words();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3904
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3905
    // entry points that exist in all platforms Note: This is code
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3906
    // that could be shared among different platforms - however the
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3907
    // benefit seems to be smaller than the disadvantage of having a
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3908
    // much more complicated generator structure. See also comment in
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3909
    // stubRoutines.hpp.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3910
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3911
    StubRoutines::_forward_exception_entry = generate_forward_exception();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3912
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3913
    StubRoutines::_call_stub_entry =
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3914
      generate_call_stub(StubRoutines::_call_stub_return_address);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3915
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3916
    // is referenced by megamorphic call
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3917
    StubRoutines::_catch_exception_entry = generate_catch_exception();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3918
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3919
    // atomic calls
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3920
    StubRoutines::_atomic_xchg_entry         = generate_atomic_xchg();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3921
    StubRoutines::_atomic_xchg_ptr_entry     = generate_atomic_xchg_ptr();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3922
    StubRoutines::_atomic_cmpxchg_entry      = generate_atomic_cmpxchg();
27691
733f189ad1f7 8058255: Native jbyte Atomic::cmpxchg for supported x86 platforms
jwilhelm
parents: 26434
diff changeset
  3923
    StubRoutines::_atomic_cmpxchg_byte_entry = generate_atomic_cmpxchg_byte();
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3924
    StubRoutines::_atomic_cmpxchg_long_entry = generate_atomic_cmpxchg_long();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3925
    StubRoutines::_atomic_add_entry          = generate_atomic_add();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3926
    StubRoutines::_atomic_add_ptr_entry      = generate_atomic_add_ptr();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3927
    StubRoutines::_fence_entry               = generate_orderaccess_fence();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3928
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3929
    StubRoutines::_handler_for_unsafe_access_entry =
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3930
      generate_handler_for_unsafe_access();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3931
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3932
    // platform dependent
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3933
    StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp();
11961
0abd4cd26e5a 7147740: add assertions to check stack alignment on VM entry from generated code (x64)
roland
parents: 11785
diff changeset
  3934
    StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3935
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3936
    StubRoutines::x86::_verify_mxcsr_entry    = generate_verify_mxcsr();
10004
190e88f7edd1 7055355: JSR 292: crash while throwing WrongMethodTypeException
never
parents: 8876
diff changeset
  3937
11411
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 11194
diff changeset
  3938
    // Build this early so it's available for the interpreter.
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 11194
diff changeset
  3939
    StubRoutines::_throw_StackOverflowError_entry =
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 11194
diff changeset
  3940
      generate_throw_exception("StackOverflowError throw_exception",
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 11194
diff changeset
  3941
                               CAST_FROM_FN_PTR(address,
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 11194
diff changeset
  3942
                                                SharedRuntime::
16b151e1e088 7116216: StackOverflow GC crash
bdelsart
parents: 11194
diff changeset
  3943
                                                throw_StackOverflowError));
18507
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3944
    if (UseCRC32Intrinsics) {
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3945
      // set table address before stub generation which use it
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3946
      StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3947
      StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
61bfc8995bb3 7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
drchase
parents: 18073
diff changeset
  3948
    }
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3949
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3950
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3951
  void generate_all() {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3952
    // Generates all stubs and initializes the entry points
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3953
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3954
    // These entry points require SharedInfo::stack0 to be set up in
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3955
    // non-core builds and need to be relocatable, so they each
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3956
    // fabricate a RuntimeStub internally.
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3957
    StubRoutines::_throw_AbstractMethodError_entry =
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3958
      generate_throw_exception("AbstractMethodError throw_exception",
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3959
                               CAST_FROM_FN_PTR(address,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3960
                                                SharedRuntime::
10545
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10004
diff changeset
  3961
                                                throw_AbstractMethodError));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3962
189
4248c8e21063 6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents: 1
diff changeset
  3963
    StubRoutines::_throw_IncompatibleClassChangeError_entry =
4248c8e21063 6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents: 1
diff changeset
  3964
      generate_throw_exception("IncompatibleClassChangeError throw_exception",
4248c8e21063 6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents: 1
diff changeset
  3965
                               CAST_FROM_FN_PTR(address,
4248c8e21063 6664627: Merge changes made only in hotspot 11 forward to jdk 7
dcubed
parents: 1
diff changeset
  3966
                                                SharedRuntime::
10545
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10004
diff changeset
  3967
                                                throw_IncompatibleClassChangeError));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3968
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3969
    StubRoutines::_throw_NullPointerException_at_call_entry =
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3970
      generate_throw_exception("NullPointerException at call throw_exception",
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3971
                               CAST_FROM_FN_PTR(address,
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3972
                                                SharedRuntime::
10545
fec876499aae 7088020: SEGV in JNIHandleBlock::release_block
never
parents: 10004
diff changeset
  3973
                                                throw_NullPointerException_at_call));
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3974
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3975
    // entry points that are platform specific
1066
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3976
    StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3977
    StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3978
    StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3979
    StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3980
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3981
    StubRoutines::x86::_float_sign_mask  = generate_fp_mask("float_sign_mask",  0x7FFFFFFF7FFFFFFF);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3982
    StubRoutines::x86::_float_sign_flip  = generate_fp_mask("float_sign_flip",  0x8000000080000000);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3983
    StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
717c3345024f 5108146: Merge i486 and amd64 cpu directories
never
parents: 957
diff changeset
  3984
    StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3985
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3986
    // support for verify_oop (must happen after universe_init)
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3987
    StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3988
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3989
    // arraycopy stubs used by compilers
489c9b5090e2 Initial load
duke
parents:
diff changeset
  3990
    generate_arraycopy_stubs();
4478
c3a8af0fc6b0 6829192: JSR 292 needs to support 64-bit x86
twisti
parents: 3262
diff changeset
  3991
4645
0c5f5b94e93a 6849984: Value methods for platform dependent math functions constant fold incorrectly
never
parents: 4478
diff changeset
  3992
    generate_math_stubs();
14132
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3993
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3994
    // don't bother generating these AES intrinsic stubs unless global flag is set
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3995
    if (UseAESIntrinsics) {
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3996
      StubRoutines::x86::_key_shuffle_mask_addr = generate_key_shuffle_mask();  // needed by the others
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3997
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3998
      StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  3999
      StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  4000
      StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  4001
      StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
3c1437abcefd 7184394: add intrinsics to use AES instructions
kvn
parents: 13952
diff changeset
  4002
    }
18740
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4003
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4004
    // Safefetch stubs.
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4005
    generate_safefetch("SafeFetch32", sizeof(int),     &StubRoutines::_safefetch32_entry,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4006
                                                       &StubRoutines::_safefetch32_fault_pc,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4007
                                                       &StubRoutines::_safefetch32_continuation_pc);
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4008
    generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4009
                                                       &StubRoutines::_safefetchN_fault_pc,
db44b1599483 8016697: Use stubs to implement safefetch
goetz
parents: 18507
diff changeset
  4010
                                                       &StubRoutines::_safefetchN_continuation_pc);
26434
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  4011
#ifdef COMPILER2
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  4012
    if (UseMultiplyToLenIntrinsic) {
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  4013
      StubRoutines::_multiplyToLen = generate_multiplyToLen();
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  4014
    }
09ad55e5f486 8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
kvn
parents: 22505
diff changeset
  4015
#endif
1
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4016
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4017
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4018
 public:
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4019
  StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4020
    if (all) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4021
      generate_all();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4022
    } else {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4023
      generate_initial();
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4024
    }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4025
  }
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4026
}; // end class declaration
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4027
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4028
void StubGenerator_generate(CodeBuffer* code, bool all) {
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4029
  StubGenerator g(code, all);
489c9b5090e2 Initial load
duke
parents:
diff changeset
  4030
}