8035283: Second phase of branch shortening doesn't account for loop alignment
Summary: added missing check for loop padding case.
Reviewed-by: kvn, jrose
/*
* Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "asm/assembler.inline.hpp"
#include "code/compiledIC.hpp"
#include "code/debugInfo.hpp"
#include "code/debugInfoRec.hpp"
#include "compiler/compileBroker.hpp"
#include "compiler/oopMap.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/locknode.hpp"
#include "opto/machnode.hpp"
#include "opto/output.hpp"
#include "opto/regalloc.hpp"
#include "opto/runtime.hpp"
#include "opto/subnode.hpp"
#include "opto/type.hpp"
#include "runtime/handles.inline.hpp"
#include "utilities/xmlstream.hpp"
extern uint size_exception_handler();
extern uint size_deopt_handler();
#ifndef PRODUCT
#define DEBUG_ARG(x) , x
#else
#define DEBUG_ARG(x)
#endif
extern int emit_exception_handler(CodeBuffer &cbuf);
extern int emit_deopt_handler(CodeBuffer &cbuf);
// Convert Nodes to instruction bits and pass off to the VM
void Compile::Output() {
// RootNode goes
assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
// The number of new nodes (mostly MachNop) is proportional to
// the number of java calls and inner loops which are aligned.
if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
C->inner_loops()*(OptoLoopAlignment-1)),
"out of nodes before code generation" ) ) {
return;
}
// Make sure I can find the Start Node
Block *entry = _cfg->get_block(1);
Block *broot = _cfg->get_root_block();
const StartNode *start = entry->head()->as_Start();
// Replace StartNode with prolog
MachPrologNode *prolog = new (this) MachPrologNode();
entry->map_node(prolog, 0);
_cfg->map_node_to_block(prolog, entry);
_cfg->unmap_node_from_block(start); // start is no longer in any block
// Virtual methods need an unverified entry point
if( is_osr_compilation() ) {
if( PoisonOSREntry ) {
// TODO: Should use a ShouldNotReachHereNode...
_cfg->insert( broot, 0, new (this) MachBreakpointNode() );
}
} else {
if( _method && !_method->flags().is_static() ) {
// Insert unvalidated entry point
_cfg->insert( broot, 0, new (this) MachUEPNode() );
}
}
// Break before main entry point
if( (_method && _method->break_at_execute())
#ifndef PRODUCT
||(OptoBreakpoint && is_method_compilation())
||(OptoBreakpointOSR && is_osr_compilation())
||(OptoBreakpointC2R && !_method)
#endif
) {
// checking for _method means that OptoBreakpoint does not apply to
// runtime stubs or frame converters
_cfg->insert( entry, 1, new (this) MachBreakpointNode() );
}
// Insert epilogs before every return
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
Block* block = _cfg->get_block(i);
if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
Node* m = block->end();
if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
block->add_inst(epilog);
_cfg->map_node_to_block(epilog, block);
}
}
}
# ifdef ENABLE_ZAP_DEAD_LOCALS
if (ZapDeadCompiledLocals) {
Insert_zap_nodes();
}
# endif
uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
blk_starts[0] = 0;
// Initialize code buffer and process short branches.
CodeBuffer* cb = init_buffer(blk_starts);
if (cb == NULL || failing()) {
return;
}
ScheduleAndBundle();
#ifndef PRODUCT
if (trace_opto_output()) {
tty->print("\n---- After ScheduleAndBundle ----\n");
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
tty->print("\nBB#%03d:\n", i);
Block* block = _cfg->get_block(i);
for (uint j = 0; j < block->number_of_nodes(); j++) {
Node* n = block->get_node(j);
OptoReg::Name reg = _regalloc->get_reg_first(n);
tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
n->dump();
}
}
}
#endif
if (failing()) {
return;
}
BuildOopMaps();
if (failing()) {
return;
}
fill_buffer(cb, blk_starts);
}
bool Compile::need_stack_bang(int frame_size_in_bytes) const {
// Determine if we need to generate a stack overflow check.
// Do it if the method is not a stub function and
// has java calls or has frame size > vm_page_size/8.
return (UseStackBanging && stub_function() == NULL &&
(has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
}
bool Compile::need_register_stack_bang() const {
// Determine if we need to generate a register stack overflow check.
// This is only used on architectures which have split register
// and memory stacks (ie. IA64).
// Bang if the method is not a stub function and has java calls
return (stub_function() == NULL && has_java_calls());
}
# ifdef ENABLE_ZAP_DEAD_LOCALS
// In order to catch compiler oop-map bugs, we have implemented
// a debugging mode called ZapDeadCompilerLocals.
// This mode causes the compiler to insert a call to a runtime routine,
// "zap_dead_locals", right before each place in compiled code
// that could potentially be a gc-point (i.e., a safepoint or oop map point).
// The runtime routine checks that locations mapped as oops are really
// oops, that locations mapped as values do not look like oops,
// and that locations mapped as dead are not used later
// (by zapping them to an invalid address).
int Compile::_CompiledZap_count = 0;
void Compile::Insert_zap_nodes() {
bool skip = false;
// Dink with static counts because code code without the extra
// runtime calls is MUCH faster for debugging purposes
if ( CompileZapFirst == 0 ) ; // nothing special
else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
else if ( CompileZapFirst == CompiledZap_count() )
warning("starting zap compilation after skipping");
if ( CompileZapLast == -1 ) ; // nothing special
else if ( CompileZapLast < CompiledZap_count() ) skip = true;
else if ( CompileZapLast == CompiledZap_count() )
warning("about to compile last zap");
++_CompiledZap_count; // counts skipped zaps, too
if ( skip ) return;
if ( _method == NULL )
return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
// Insert call to zap runtime stub before every node with an oop map
for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
Block *b = _cfg->get_block(i);
for ( uint j = 0; j < b->number_of_nodes(); ++j ) {
Node *n = b->get_node(j);
// Determining if we should insert a zap-a-lot node in output.
// We do that for all nodes that has oopmap info, except for calls
// to allocation. Calls to allocation passes in the old top-of-eden pointer
// and expect the C code to reset it. Hence, there can be no safepoints between
// the inlined-allocation and the call to new_Java, etc.
// We also cannot zap monitor calls, as they must hold the microlock
// during the call to Zap, which also wants to grab the microlock.
bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
if ( insert ) { // it is MachSafePoint
if ( !n->is_MachCall() ) {
insert = false;
} else if ( n->is_MachCall() ) {
MachCallNode* call = n->as_MachCall();
if (call->entry_point() == OptoRuntime::new_instance_Java() ||
call->entry_point() == OptoRuntime::new_array_Java() ||
call->entry_point() == OptoRuntime::multianewarray2_Java() ||
call->entry_point() == OptoRuntime::multianewarray3_Java() ||
call->entry_point() == OptoRuntime::multianewarray4_Java() ||
call->entry_point() == OptoRuntime::multianewarray5_Java() ||
call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
) {
insert = false;
}
}
if (insert) {
Node *zap = call_zap_node(n->as_MachSafePoint(), i);
b->insert_node(zap, j);
_cfg->map_node_to_block(zap, b);
++j;
}
}
}
}
}
Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
CallStaticJavaNode* ideal_node =
new (this) CallStaticJavaNode( tf,
OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
"call zap dead locals stub", 0, TypePtr::BOTTOM);
// We need to copy the OopMap from the site we're zapping at.
// We have to make a copy, because the zap site might not be
// a call site, and zap_dead is a call site.
OopMap* clone = node_to_check->oop_map()->deep_copy();
// Add the cloned OopMap to the zap node
ideal_node->set_oop_map(clone);
return _matcher->match_sfpt(ideal_node);
}
bool Compile::is_node_getting_a_safepoint( Node* n) {
// This code duplicates the logic prior to the call of add_safepoint
// below in this file.
if( n->is_MachSafePoint() ) return true;
return false;
}
# endif // ENABLE_ZAP_DEAD_LOCALS
// Compute the size of first NumberOfLoopInstrToAlign instructions at the top
// of a loop. When aligning a loop we need to provide enough instructions
// in cpu's fetch buffer to feed decoders. The loop alignment could be
// avoided if we have enough instructions in fetch buffer at the head of a loop.
// By default, the size is set to 999999 by Block's constructor so that
// a loop will be aligned if the size is not reset here.
//
// Note: Mach instructions could contain several HW instructions
// so the size is estimated only.
//
void Compile::compute_loop_first_inst_sizes() {
// The next condition is used to gate the loop alignment optimization.
// Don't aligned a loop if there are enough instructions at the head of a loop
// or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
// is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
// equal to 11 bytes which is the largest address NOP instruction.
if (MaxLoopPad < OptoLoopAlignment - 1) {
uint last_block = _cfg->number_of_blocks() - 1;
for (uint i = 1; i <= last_block; i++) {
Block* block = _cfg->get_block(i);
// Check the first loop's block which requires an alignment.
if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
uint sum_size = 0;
uint inst_cnt = NumberOfLoopInstrToAlign;
inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
// Check subsequent fallthrough blocks if the loop's first
// block(s) does not have enough instructions.
Block *nb = block;
while(inst_cnt > 0 &&
i < last_block &&
!_cfg->get_block(i + 1)->has_loop_alignment() &&
!nb->has_successor(block)) {
i++;
nb = _cfg->get_block(i);
inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
} // while( inst_cnt > 0 && i < last_block )
block->set_first_inst_size(sum_size);
} // f( b->head()->is_Loop() )
} // for( i <= last_block )
} // if( MaxLoopPad < OptoLoopAlignment-1 )
}
// The architecture description provides short branch variants for some long
// branch instructions. Replace eligible long branches with short branches.
void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
// Compute size of each block, method size, and relocation information size
uint nblocks = _cfg->number_of_blocks();
uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
// Collect worst case block paddings
int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
memset(block_worst_case_pad, 0, nblocks * sizeof(int));
DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
bool has_short_branch_candidate = false;
// Initialize the sizes to 0
code_size = 0; // Size in bytes of generated code
stub_size = 0; // Size in bytes of all stub entries
// Size in bytes of all relocation entries, including those in local stubs.
// Start with 2-bytes of reloc info for the unvalidated entry point
reloc_size = 1; // Number of relocation entries
// Make three passes. The first computes pessimistic blk_starts,
// relative jmp_offset and reloc_size information. The second performs
// short branch substitution using the pessimistic sizing. The
// third inserts nops where needed.
// Step one, perform a pessimistic sizing pass.
uint last_call_adr = max_uint;
uint last_avoid_back_to_back_adr = max_uint;
uint nop_size = (new (this) MachNopNode())->size(_regalloc);
for (uint i = 0; i < nblocks; i++) { // For all blocks
Block* block = _cfg->get_block(i);
// During short branch replacement, we store the relative (to blk_starts)
// offset of jump in jmp_offset, rather than the absolute offset of jump.
// This is so that we do not need to recompute sizes of all nodes when
// we compute correct blk_starts in our next sizing pass.
jmp_offset[i] = 0;
jmp_size[i] = 0;
jmp_nidx[i] = -1;
DEBUG_ONLY( jmp_target[i] = 0; )
DEBUG_ONLY( jmp_rule[i] = 0; )
// Sum all instruction sizes to compute block size
uint last_inst = block->number_of_nodes();
uint blk_size = 0;
for (uint j = 0; j < last_inst; j++) {
Node* nj = block->get_node(j);
// Handle machine instruction nodes
if (nj->is_Mach()) {
MachNode *mach = nj->as_Mach();
blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
reloc_size += mach->reloc();
if (mach->is_MachCall()) {
MachCallNode *mcall = mach->as_MachCall();
// This destination address is NOT PC-relative
mcall->method_set((intptr_t)mcall->entry_point());
if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
stub_size += CompiledStaticCall::to_interp_stub_size();
reloc_size += CompiledStaticCall::reloc_to_interp_stub();
}
} else if (mach->is_MachSafePoint()) {
// If call/safepoint are adjacent, account for possible
// nop to disambiguate the two safepoints.
// ScheduleAndBundle() can rearrange nodes in a block,
// check for all offsets inside this block.
if (last_call_adr >= blk_starts[i]) {
blk_size += nop_size;
}
}
if (mach->avoid_back_to_back()) {
// Nop is inserted between "avoid back to back" instructions.
// ScheduleAndBundle() can rearrange nodes in a block,
// check for all offsets inside this block.
if (last_avoid_back_to_back_adr >= blk_starts[i]) {
blk_size += nop_size;
}
}
if (mach->may_be_short_branch()) {
if (!nj->is_MachBranch()) {
#ifndef PRODUCT
nj->dump(3);
#endif
Unimplemented();
}
assert(jmp_nidx[i] == -1, "block should have only one branch");
jmp_offset[i] = blk_size;
jmp_size[i] = nj->size(_regalloc);
jmp_nidx[i] = j;
has_short_branch_candidate = true;
}
}
blk_size += nj->size(_regalloc);
// Remember end of call offset
if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
last_call_adr = blk_starts[i]+blk_size;
}
// Remember end of avoid_back_to_back offset
if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
}
}
// When the next block starts a loop, we may insert pad NOP
// instructions. Since we cannot know our future alignment,
// assume the worst.
if (i < nblocks - 1) {
Block* nb = _cfg->get_block(i + 1);
int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
if (max_loop_pad > 0) {
assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
// Adjust last_call_adr and/or last_avoid_back_to_back_adr.
// If either is the last instruction in this block, bump by
// max_loop_pad in lock-step with blk_size, so sizing
// calculations in subsequent blocks still can conservatively
// detect that it may the last instruction in this block.
if (last_call_adr == blk_starts[i]+blk_size) {
last_call_adr += max_loop_pad;
}
if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
last_avoid_back_to_back_adr += max_loop_pad;
}
blk_size += max_loop_pad;
block_worst_case_pad[i + 1] = max_loop_pad;
}
}
// Save block size; update total method size
blk_starts[i+1] = blk_starts[i]+blk_size;
}
// Step two, replace eligible long jumps.
bool progress = true;
uint last_may_be_short_branch_adr = max_uint;
while (has_short_branch_candidate && progress) {
progress = false;
has_short_branch_candidate = false;
int adjust_block_start = 0;
for (uint i = 0; i < nblocks; i++) {
Block* block = _cfg->get_block(i);
int idx = jmp_nidx[i];
MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
if (mach != NULL && mach->may_be_short_branch()) {
#ifdef ASSERT
assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
int j;
// Find the branch; ignore trailing NOPs.
for (j = block->number_of_nodes()-1; j>=0; j--) {
Node* n = block->get_node(j);
if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
break;
}
assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
#endif
int br_size = jmp_size[i];
int br_offs = blk_starts[i] + jmp_offset[i];
// This requires the TRUE branch target be in succs[0]
uint bnum = block->non_connector_successor(0)->_pre_order;
int offset = blk_starts[bnum] - br_offs;
if (bnum > i) { // adjust following block's offset
offset -= adjust_block_start;
}
// This block can be a loop header, account for the padding
// in the previous block.
int block_padding = block_worst_case_pad[i];
assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
// In the following code a nop could be inserted before
// the branch which will increase the backward distance.
bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
if (needs_padding && offset <= 0)
offset -= nop_size;
if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
// We've got a winner. Replace this branch.
MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
// Update the jmp_size.
int new_size = replacement->size(_regalloc);
int diff = br_size - new_size;
assert(diff >= (int)nop_size, "short_branch size should be smaller");
// Conservatively take into accound padding between
// avoid_back_to_back branches. Previous branch could be
// converted into avoid_back_to_back branch during next
// rounds.
if (needs_padding && replacement->avoid_back_to_back()) {
jmp_offset[i] += nop_size;
diff -= nop_size;
}
adjust_block_start += diff;
block->map_node(replacement, idx);
mach->subsume_by(replacement, C);
mach = replacement;
progress = true;
jmp_size[i] = new_size;
DEBUG_ONLY( jmp_target[i] = bnum; );
DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
} else {
// The jump distance is not short, try again during next iteration.
has_short_branch_candidate = true;
}
} // (mach->may_be_short_branch())
if (mach != NULL && (mach->may_be_short_branch() ||
mach->avoid_back_to_back())) {
last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
}
blk_starts[i+1] -= adjust_block_start;
}
}
#ifdef ASSERT
for (uint i = 0; i < nblocks; i++) { // For all blocks
if (jmp_target[i] != 0) {
int br_size = jmp_size[i];
int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
}
assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
}
}
#endif
// Step 3, compute the offsets of all blocks, will be done in fill_buffer()
// after ScheduleAndBundle().
// ------------------
// Compute size for code buffer
code_size = blk_starts[nblocks];
// Relocation records
reloc_size += 1; // Relo entry for exception handler
// Adjust reloc_size to number of record of relocation info
// Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
// a relocation index.
// The CodeBuffer will expand the locs array if this estimate is too low.
reloc_size *= 10 / sizeof(relocInfo);
}
//------------------------------FillLocArray-----------------------------------
// Create a bit of debug info and append it to the array. The mapping is from
// Java local or expression stack to constant, register or stack-slot. For
// doubles, insert 2 mappings and return 1 (to tell the caller that the next
// entry has been taken care of and caller should skip it).
static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
// This should never have accepted Bad before
assert(OptoReg::is_valid(regnum), "location must be valid");
return (OptoReg::is_reg(regnum))
? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
: new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
}
ObjectValue*
Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
for (int i = 0; i < objs->length(); i++) {
assert(objs->at(i)->is_object(), "corrupt object cache");
ObjectValue* sv = (ObjectValue*) objs->at(i);
if (sv->id() == id) {
return sv;
}
}
// Otherwise..
return NULL;
}
void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
ObjectValue* sv ) {
assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
objs->append(sv);
}
void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
GrowableArray<ScopeValue*> *array,
GrowableArray<ScopeValue*> *objs ) {
assert( local, "use _top instead of null" );
if (array->length() != idx) {
assert(array->length() == idx + 1, "Unexpected array count");
// Old functionality:
// return
// New functionality:
// Assert if the local is not top. In product mode let the new node
// override the old entry.
assert(local == top(), "LocArray collision");
if (local == top()) {
return;
}
array->pop();
}
const Type *t = local->bottom_type();
// Is it a safepoint scalar object node?
if (local->is_SafePointScalarObject()) {
SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
if (sv == NULL) {
ciKlass* cik = t->is_oopptr()->klass();
assert(cik->is_instance_klass() ||
cik->is_array_klass(), "Not supported allocation.");
sv = new ObjectValue(spobj->_idx,
new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
Compile::set_sv_for_object_node(objs, sv);
uint first_ind = spobj->first_index(sfpt->jvms());
for (uint i = 0; i < spobj->n_fields(); i++) {
Node* fld_node = sfpt->in(first_ind+i);
(void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
}
}
array->append(sv);
return;
}
// Grab the register number for the local
OptoReg::Name regnum = _regalloc->get_reg_first(local);
if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
// Record the double as two float registers.
// The register mask for such a value always specifies two adjacent
// float registers, with the lower register number even.
// Normally, the allocation of high and low words to these registers
// is irrelevant, because nearly all operations on register pairs
// (e.g., StoreD) treat them as a single unit.
// Here, we assume in addition that the words in these two registers
// stored "naturally" (by operations like StoreD and double stores
// within the interpreter) such that the lower-numbered register
// is written to the lower memory address. This may seem like
// a machine dependency, but it is not--it is a requirement on
// the author of the <arch>.ad file to ensure that, for every
// even/odd double-register pair to which a double may be allocated,
// the word in the even single-register is stored to the first
// memory word. (Note that register numbers are completely
// arbitrary, and are not tied to any machine-level encodings.)
#ifdef _LP64
if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
array->append(new ConstantIntValue(0));
array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
} else if ( t->base() == Type::Long ) {
array->append(new ConstantIntValue(0));
array->append(new_loc_value( _regalloc, regnum, Location::lng ));
} else if ( t->base() == Type::RawPtr ) {
// jsr/ret return address which must be restored into a the full
// width 64-bit stack slot.
array->append(new_loc_value( _regalloc, regnum, Location::lng ));
}
#else //_LP64
#ifdef SPARC
if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
// For SPARC we have to swap high and low words for
// long values stored in a single-register (g0-g7).
array->append(new_loc_value( _regalloc, regnum , Location::normal ));
array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
} else
#endif //SPARC
if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
// Repack the double/long as two jints.
// The convention the interpreter uses is that the second local
// holds the first raw word of the native double representation.
// This is actually reasonable, since locals and stack arrays
// grow downwards in all implementations.
// (If, on some machine, the interpreter's Java locals or stack
// were to grow upwards, the embedded doubles would be word-swapped.)
array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
array->append(new_loc_value( _regalloc, regnum , Location::normal ));
}
#endif //_LP64
else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
OptoReg::is_reg(regnum) ) {
array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
? Location::float_in_dbl : Location::normal ));
} else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
? Location::int_in_long : Location::normal ));
} else if( t->base() == Type::NarrowOop ) {
array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
} else {
array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
}
return;
}
// No register. It must be constant data.
switch (t->base()) {
case Type::Half: // Second half of a double
ShouldNotReachHere(); // Caller should skip 2nd halves
break;
case Type::AnyPtr:
array->append(new ConstantOopWriteValue(NULL));
break;
case Type::AryPtr:
case Type::InstPtr: // fall through
array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
break;
case Type::NarrowOop:
if (t == TypeNarrowOop::NULL_PTR) {
array->append(new ConstantOopWriteValue(NULL));
} else {
array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
}
break;
case Type::Int:
array->append(new ConstantIntValue(t->is_int()->get_con()));
break;
case Type::RawPtr:
// A return address (T_ADDRESS).
assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
#ifdef _LP64
// Must be restored to the full-width 64-bit stack slot.
array->append(new ConstantLongValue(t->is_ptr()->get_con()));
#else
array->append(new ConstantIntValue(t->is_ptr()->get_con()));
#endif
break;
case Type::FloatCon: {
float f = t->is_float_constant()->getf();
array->append(new ConstantIntValue(jint_cast(f)));
break;
}
case Type::DoubleCon: {
jdouble d = t->is_double_constant()->getd();
#ifdef _LP64
array->append(new ConstantIntValue(0));
array->append(new ConstantDoubleValue(d));
#else
// Repack the double as two jints.
// The convention the interpreter uses is that the second local
// holds the first raw word of the native double representation.
// This is actually reasonable, since locals and stack arrays
// grow downwards in all implementations.
// (If, on some machine, the interpreter's Java locals or stack
// were to grow upwards, the embedded doubles would be word-swapped.)
jint *dp = (jint*)&d;
array->append(new ConstantIntValue(dp[1]));
array->append(new ConstantIntValue(dp[0]));
#endif
break;
}
case Type::Long: {
jlong d = t->is_long()->get_con();
#ifdef _LP64
array->append(new ConstantIntValue(0));
array->append(new ConstantLongValue(d));
#else
// Repack the long as two jints.
// The convention the interpreter uses is that the second local
// holds the first raw word of the native double representation.
// This is actually reasonable, since locals and stack arrays
// grow downwards in all implementations.
// (If, on some machine, the interpreter's Java locals or stack
// were to grow upwards, the embedded doubles would be word-swapped.)
jint *dp = (jint*)&d;
array->append(new ConstantIntValue(dp[1]));
array->append(new ConstantIntValue(dp[0]));
#endif
break;
}
case Type::Top: // Add an illegal value here
array->append(new LocationValue(Location()));
break;
default:
ShouldNotReachHere();
break;
}
}
// Determine if this node starts a bundle
bool Compile::starts_bundle(const Node *n) const {
return (_node_bundling_limit > n->_idx &&
_node_bundling_base[n->_idx].starts_bundle());
}
//--------------------------Process_OopMap_Node--------------------------------
void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
// Handle special safepoint nodes for synchronization
MachSafePointNode *sfn = mach->as_MachSafePoint();
MachCallNode *mcall;
#ifdef ENABLE_ZAP_DEAD_LOCALS
assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
#endif
int safepoint_pc_offset = current_offset;
bool is_method_handle_invoke = false;
bool return_oop = false;
// Add the safepoint in the DebugInfoRecorder
if( !mach->is_MachCall() ) {
mcall = NULL;
debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
} else {
mcall = mach->as_MachCall();
// Is the call a MethodHandle call?
if (mcall->is_MachCallJava()) {
if (mcall->as_MachCallJava()->_method_handle_invoke) {
assert(has_method_handle_invokes(), "must have been set during call generation");
is_method_handle_invoke = true;
}
}
// Check if a call returns an object.
if (mcall->return_value_is_used() &&
mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
return_oop = true;
}
safepoint_pc_offset += mcall->ret_addr_offset();
debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
}
// Loop over the JVMState list to add scope information
// Do not skip safepoints with a NULL method, they need monitor info
JVMState* youngest_jvms = sfn->jvms();
int max_depth = youngest_jvms->depth();
// Allocate the object pool for scalar-replaced objects -- the map from
// small-integer keys (which can be recorded in the local and ostack
// arrays) to descriptions of the object state.
GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
// Visit scopes from oldest to youngest.
for (int depth = 1; depth <= max_depth; depth++) {
JVMState* jvms = youngest_jvms->of_depth(depth);
int idx;
ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
// Safepoints that do not have method() set only provide oop-map and monitor info
// to support GC; these do not support deoptimization.
int num_locs = (method == NULL) ? 0 : jvms->loc_size();
int num_exps = (method == NULL) ? 0 : jvms->stk_size();
int num_mon = jvms->nof_monitors();
assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
"JVMS local count must match that of the method");
// Add Local and Expression Stack Information
// Insert locals into the locarray
GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
for( idx = 0; idx < num_locs; idx++ ) {
FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
}
// Insert expression stack entries into the exparray
GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
for( idx = 0; idx < num_exps; idx++ ) {
FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
}
// Add in mappings of the monitors
assert( !method ||
!method->is_synchronized() ||
method->is_native() ||
num_mon > 0 ||
!GenerateSynchronizationCode,
"monitors must always exist for synchronized methods");
// Build the growable array of ScopeValues for exp stack
GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
// Loop over monitors and insert into array
for (idx = 0; idx < num_mon; idx++) {
// Grab the node that defines this monitor
Node* box_node = sfn->monitor_box(jvms, idx);
Node* obj_node = sfn->monitor_obj(jvms, idx);
// Create ScopeValue for object
ScopeValue *scval = NULL;
if (obj_node->is_SafePointScalarObject()) {
SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
scval = Compile::sv_for_node_id(objs, spobj->_idx);
if (scval == NULL) {
const Type *t = spobj->bottom_type();
ciKlass* cik = t->is_oopptr()->klass();
assert(cik->is_instance_klass() ||
cik->is_array_klass(), "Not supported allocation.");
ObjectValue* sv = new ObjectValue(spobj->_idx,
new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
Compile::set_sv_for_object_node(objs, sv);
uint first_ind = spobj->first_index(youngest_jvms);
for (uint i = 0; i < spobj->n_fields(); i++) {
Node* fld_node = sfn->in(first_ind+i);
(void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
}
scval = sv;
}
} else if (!obj_node->is_Con()) {
OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
} else {
scval = new_loc_value( _regalloc, obj_reg, Location::oop );
}
} else {
const TypePtr *tp = obj_node->get_ptr_type();
scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
}
OptoReg::Name box_reg = BoxLockNode::reg(box_node);
Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
monarray->append(new MonitorValue(scval, basic_lock, eliminated));
}
// We dump the object pool first, since deoptimization reads it in first.
debug_info()->dump_object_pool(objs);
// Build first class objects to pass to scope
DebugToken *locvals = debug_info()->create_scope_values(locarray);
DebugToken *expvals = debug_info()->create_scope_values(exparray);
DebugToken *monvals = debug_info()->create_monitor_values(monarray);
// Make method available for all Safepoints
ciMethod* scope_method = method ? method : _method;
// Describe the scope here
assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
// Now we can describe the scope.
debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
} // End jvms loop
// Mark the end of the scope set.
debug_info()->end_safepoint(safepoint_pc_offset);
}
// A simplified version of Process_OopMap_Node, to handle non-safepoints.
class NonSafepointEmitter {
Compile* C;
JVMState* _pending_jvms;
int _pending_offset;
void emit_non_safepoint();
public:
NonSafepointEmitter(Compile* compile) {
this->C = compile;
_pending_jvms = NULL;
_pending_offset = 0;
}
void observe_instruction(Node* n, int pc_offset) {
if (!C->debug_info()->recording_non_safepoints()) return;
Node_Notes* nn = C->node_notes_at(n->_idx);
if (nn == NULL || nn->jvms() == NULL) return;
if (_pending_jvms != NULL &&
_pending_jvms->same_calls_as(nn->jvms())) {
// Repeated JVMS? Stretch it up here.
_pending_offset = pc_offset;
} else {
if (_pending_jvms != NULL &&
_pending_offset < pc_offset) {
emit_non_safepoint();
}
_pending_jvms = NULL;
if (pc_offset > C->debug_info()->last_pc_offset()) {
// This is the only way _pending_jvms can become non-NULL:
_pending_jvms = nn->jvms();
_pending_offset = pc_offset;
}
}
}
// Stay out of the way of real safepoints:
void observe_safepoint(JVMState* jvms, int pc_offset) {
if (_pending_jvms != NULL &&
!_pending_jvms->same_calls_as(jvms) &&
_pending_offset < pc_offset) {
emit_non_safepoint();
}
_pending_jvms = NULL;
}
void flush_at_end() {
if (_pending_jvms != NULL) {
emit_non_safepoint();
}
_pending_jvms = NULL;
}
};
void NonSafepointEmitter::emit_non_safepoint() {
JVMState* youngest_jvms = _pending_jvms;
int pc_offset = _pending_offset;
// Clear it now:
_pending_jvms = NULL;
DebugInformationRecorder* debug_info = C->debug_info();
assert(debug_info->recording_non_safepoints(), "sanity");
debug_info->add_non_safepoint(pc_offset);
int max_depth = youngest_jvms->depth();
// Visit scopes from oldest to youngest.
for (int depth = 1; depth <= max_depth; depth++) {
JVMState* jvms = youngest_jvms->of_depth(depth);
ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
}
// Mark the end of the scope set.
debug_info->end_non_safepoint(pc_offset);
}
//------------------------------init_buffer------------------------------------
CodeBuffer* Compile::init_buffer(uint* blk_starts) {
// Set the initially allocated size
int code_req = initial_code_capacity;
int locs_req = initial_locs_capacity;
int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
int const_req = initial_const_capacity;
int pad_req = NativeCall::instruction_size;
// The extra spacing after the code is necessary on some platforms.
// Sometimes we need to patch in a jump after the last instruction,
// if the nmethod has been deoptimized. (See 4932387, 4894843.)
// Compute the byte offset where we can store the deopt pc.
if (fixed_slots() != 0) {
_orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
}
// Compute prolog code size
_method_size = 0;
_frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
#if defined(IA64) && !defined(AIX)
if (save_argument_registers()) {
// 4815101: this is a stub with implicit and unknown precision fp args.
// The usual spill mechanism can only generate stfd's in this case, which
// doesn't work if the fp reg to spill contains a single-precision denorm.
// Instead, we hack around the normal spill mechanism using stfspill's and
// ldffill's in the MachProlog and MachEpilog emit methods. We allocate
// space here for the fp arg regs (f8-f15) we're going to thusly spill.
//
// If we ever implement 16-byte 'registers' == stack slots, we can
// get rid of this hack and have SpillCopy generate stfspill/ldffill
// instead of stfd/stfs/ldfd/ldfs.
_frame_slots += 8*(16/BytesPerInt);
}
#endif
assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
if (has_mach_constant_base_node()) {
uint add_size = 0;
// Fill the constant table.
// Note: This must happen before shorten_branches.
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
Block* b = _cfg->get_block(i);
for (uint j = 0; j < b->number_of_nodes(); j++) {
Node* n = b->get_node(j);
// If the node is a MachConstantNode evaluate the constant
// value section.
if (n->is_MachConstant()) {
MachConstantNode* machcon = n->as_MachConstant();
machcon->eval_constant(C);
} else if (n->is_Mach()) {
// On Power there are more nodes that issue constants.
add_size += (n->as_Mach()->ins_num_consts() * 8);
}
}
}
// Calculate the offsets of the constants and the size of the
// constant table (including the padding to the next section).
constant_table().calculate_offsets_and_size();
const_req = constant_table().size() + add_size;
}
// Initialize the space for the BufferBlob used to find and verify
// instruction size in MachNode::emit_size()
init_scratch_buffer_blob(const_req);
if (failing()) return NULL; // Out of memory
// Pre-compute the length of blocks and replace
// long branches with short if machine supports it.
shorten_branches(blk_starts, code_req, locs_req, stub_req);
// nmethod and CodeBuffer count stubs & constants as part of method's code.
int exception_handler_req = size_exception_handler();
int deopt_handler_req = size_deopt_handler();
exception_handler_req += MAX_stubs_size; // add marginal slop for handler
deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
stub_req += MAX_stubs_size; // ensure per-stub margin
code_req += MAX_inst_size; // ensure per-instruction margin
if (StressCodeBuffers)
code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
int total_req =
const_req +
code_req +
pad_req +
stub_req +
exception_handler_req +
deopt_handler_req; // deopt handler
if (has_method_handle_invokes())
total_req += deopt_handler_req; // deopt MH handler
CodeBuffer* cb = code_buffer();
cb->initialize(total_req, locs_req);
// Have we run out of code space?
if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
C->record_failure("CodeCache is full");
return NULL;
}
// Configure the code buffer.
cb->initialize_consts_size(const_req);
cb->initialize_stubs_size(stub_req);
cb->initialize_oop_recorder(env()->oop_recorder());
// fill in the nop array for bundling computations
MachNode *_nop_list[Bundle::_nop_count];
Bundle::initialize_nops(_nop_list, this);
return cb;
}
//------------------------------fill_buffer------------------------------------
void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
// blk_starts[] contains offsets calculated during short branches processing,
// offsets should not be increased during following steps.
// Compute the size of first NumberOfLoopInstrToAlign instructions at head
// of a loop. It is used to determine the padding for loop alignment.
compute_loop_first_inst_sizes();
// Create oopmap set.
_oop_map_set = new OopMapSet();
// !!!!! This preserves old handling of oopmaps for now
debug_info()->set_oopmaps(_oop_map_set);
uint nblocks = _cfg->number_of_blocks();
// Count and start of implicit null check instructions
uint inct_cnt = 0;
uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
// Count and start of calls
uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
uint return_offset = 0;
int nop_size = (new (this) MachNopNode())->size(_regalloc);
int previous_offset = 0;
int current_offset = 0;
int last_call_offset = -1;
int last_avoid_back_to_back_offset = -1;
#ifdef ASSERT
uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
#endif
// Create an array of unused labels, one for each basic block, if printing is enabled
#ifndef PRODUCT
int *node_offsets = NULL;
uint node_offset_limit = unique();
if (print_assembly())
node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
#endif
NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
// Emit the constant table.
if (has_mach_constant_base_node()) {
constant_table().emit(*cb);
}
// Create an array of labels, one for each basic block
Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
for (uint i=0; i <= nblocks; i++) {
blk_labels[i].init();
}
// ------------------
// Now fill in the code buffer
Node *delay_slot = NULL;
for (uint i = 0; i < nblocks; i++) {
Block* block = _cfg->get_block(i);
Node* head = block->head();
// If this block needs to start aligned (i.e, can be reached other
// than by falling-thru from the previous block), then force the
// start of a new bundle.
if (Pipeline::requires_bundling() && starts_bundle(head)) {
cb->flush_bundle(true);
}
#ifdef ASSERT
if (!block->is_connector()) {
stringStream st;
block->dump_head(_cfg, &st);
MacroAssembler(cb).block_comment(st.as_string());
}
jmp_target[i] = 0;
jmp_offset[i] = 0;
jmp_size[i] = 0;
jmp_rule[i] = 0;
#endif
int blk_offset = current_offset;
// Define the label at the beginning of the basic block
MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
uint last_inst = block->number_of_nodes();
// Emit block normally, except for last instruction.
// Emit means "dump code bits into code buffer".
for (uint j = 0; j<last_inst; j++) {
// Get the node
Node* n = block->get_node(j);
// See if delay slots are supported
if (valid_bundle_info(n) &&
node_bundling(n)->used_in_unconditional_delay()) {
assert(delay_slot == NULL, "no use of delay slot node");
assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
delay_slot = n;
continue;
}
// If this starts a new instruction group, then flush the current one
// (but allow split bundles)
if (Pipeline::requires_bundling() && starts_bundle(n))
cb->flush_bundle(false);
// The following logic is duplicated in the code ifdeffed for
// ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
// should be factored out. Or maybe dispersed to the nodes?
// Special handling for SafePoint/Call Nodes
bool is_mcall = false;
if (n->is_Mach()) {
MachNode *mach = n->as_Mach();
is_mcall = n->is_MachCall();
bool is_sfn = n->is_MachSafePoint();
// If this requires all previous instructions be flushed, then do so
if (is_sfn || is_mcall || mach->alignment_required() != 1) {
cb->flush_bundle(true);
current_offset = cb->insts_size();
}
// A padding may be needed again since a previous instruction
// could be moved to delay slot.
// align the instruction if necessary
int padding = mach->compute_padding(current_offset);
// Make sure safepoint node for polling is distinct from a call's
// return by adding a nop if needed.
if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
padding = nop_size;
}
if (padding == 0 && mach->avoid_back_to_back() &&
current_offset == last_avoid_back_to_back_offset) {
// Avoid back to back some instructions.
padding = nop_size;
}
if(padding > 0) {
assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
int nops_cnt = padding / nop_size;
MachNode *nop = new (this) MachNopNode(nops_cnt);
block->insert_node(nop, j++);
last_inst++;
_cfg->map_node_to_block(nop, block);
nop->emit(*cb, _regalloc);
cb->flush_bundle(true);
current_offset = cb->insts_size();
}
// Remember the start of the last call in a basic block
if (is_mcall) {
MachCallNode *mcall = mach->as_MachCall();
// This destination address is NOT PC-relative
mcall->method_set((intptr_t)mcall->entry_point());
// Save the return address
call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
if (mcall->is_MachCallLeaf()) {
is_mcall = false;
is_sfn = false;
}
}
// sfn will be valid whenever mcall is valid now because of inheritance
if (is_sfn || is_mcall) {
// Handle special safepoint nodes for synchronization
if (!is_mcall) {
MachSafePointNode *sfn = mach->as_MachSafePoint();
// !!!!! Stubs only need an oopmap right now, so bail out
if (sfn->jvms()->method() == NULL) {
// Write the oopmap directly to the code blob??!!
# ifdef ENABLE_ZAP_DEAD_LOCALS
assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
# endif
continue;
}
} // End synchronization
non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
current_offset);
Process_OopMap_Node(mach, current_offset);
} // End if safepoint
// If this is a null check, then add the start of the previous instruction to the list
else if( mach->is_MachNullCheck() ) {
inct_starts[inct_cnt++] = previous_offset;
}
// If this is a branch, then fill in the label with the target BB's label
else if (mach->is_MachBranch()) {
// This requires the TRUE branch target be in succs[0]
uint block_num = block->non_connector_successor(0)->_pre_order;
// Try to replace long branch if delay slot is not used,
// it is mostly for back branches since forward branch's
// distance is not updated yet.
bool delay_slot_is_used = valid_bundle_info(n) &&
node_bundling(n)->use_unconditional_delay();
if (!delay_slot_is_used && mach->may_be_short_branch()) {
assert(delay_slot == NULL, "not expecting delay slot node");
int br_size = n->size(_regalloc);
int offset = blk_starts[block_num] - current_offset;
if (block_num >= i) {
// Current and following block's offset are not
// finalized yet, adjust distance by the difference
// between calculated and final offsets of current block.
offset -= (blk_starts[i] - blk_offset);
}
// In the following code a nop could be inserted before
// the branch which will increase the backward distance.
bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
if (needs_padding && offset <= 0)
offset -= nop_size;
if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
// We've got a winner. Replace this branch.
MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
// Update the jmp_size.
int new_size = replacement->size(_regalloc);
assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
// Insert padding between avoid_back_to_back branches.
if (needs_padding && replacement->avoid_back_to_back()) {
MachNode *nop = new (this) MachNopNode();
block->insert_node(nop, j++);
_cfg->map_node_to_block(nop, block);
last_inst++;
nop->emit(*cb, _regalloc);
cb->flush_bundle(true);
current_offset = cb->insts_size();
}
#ifdef ASSERT
jmp_target[i] = block_num;
jmp_offset[i] = current_offset - blk_offset;
jmp_size[i] = new_size;
jmp_rule[i] = mach->rule();
#endif
block->map_node(replacement, j);
mach->subsume_by(replacement, C);
n = replacement;
mach = replacement;
}
}
mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
} else if (mach->ideal_Opcode() == Op_Jump) {
for (uint h = 0; h < block->_num_succs; h++) {
Block* succs_block = block->_succs[h];
for (uint j = 1; j < succs_block->num_preds(); j++) {
Node* jpn = succs_block->pred(j);
if (jpn->is_JumpProj() && jpn->in(0) == mach) {
uint block_num = succs_block->non_connector()->_pre_order;
Label *blkLabel = &blk_labels[block_num];
mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
}
}
}
}
#ifdef ASSERT
// Check that oop-store precedes the card-mark
else if (mach->ideal_Opcode() == Op_StoreCM) {
uint storeCM_idx = j;
int count = 0;
for (uint prec = mach->req(); prec < mach->len(); prec++) {
Node *oop_store = mach->in(prec); // Precedence edge
if (oop_store == NULL) continue;
count++;
uint i4;
for (i4 = 0; i4 < last_inst; ++i4) {
if (block->get_node(i4) == oop_store) {
break;
}
}
// Note: This test can provide a false failure if other precedence
// edges have been added to the storeCMNode.
assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
}
assert(count > 0, "storeCM expects at least one precedence edge");
}
#endif
else if (!n->is_Proj()) {
// Remember the beginning of the previous instruction, in case
// it's followed by a flag-kill and a null-check. Happens on
// Intel all the time, with add-to-memory kind of opcodes.
previous_offset = current_offset;
}
// Not an else-if!
// If this is a trap based cmp then add its offset to the list.
if (mach->is_TrapBasedCheckNode()) {
inct_starts[inct_cnt++] = current_offset;
}
}
// Verify that there is sufficient space remaining
cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
C->record_failure("CodeCache is full");
return;
}
// Save the offset for the listing
#ifndef PRODUCT
if (node_offsets && n->_idx < node_offset_limit)
node_offsets[n->_idx] = cb->insts_size();
#endif
// "Normal" instruction case
DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
n->emit(*cb, _regalloc);
current_offset = cb->insts_size();
#ifdef ASSERT
if (n->size(_regalloc) < (current_offset-instr_offset)) {
n->dump();
assert(false, "wrong size of mach node");
}
#endif
non_safepoints.observe_instruction(n, current_offset);
// mcall is last "call" that can be a safepoint
// record it so we can see if a poll will directly follow it
// in which case we'll need a pad to make the PcDesc sites unique
// see 5010568. This can be slightly inaccurate but conservative
// in the case that return address is not actually at current_offset.
// This is a small price to pay.
if (is_mcall) {
last_call_offset = current_offset;
}
if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
// Avoid back to back some instructions.
last_avoid_back_to_back_offset = current_offset;
}
// See if this instruction has a delay slot
if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
assert(delay_slot != NULL, "expecting delay slot node");
// Back up 1 instruction
cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
// Save the offset for the listing
#ifndef PRODUCT
if (node_offsets && delay_slot->_idx < node_offset_limit)
node_offsets[delay_slot->_idx] = cb->insts_size();
#endif
// Support a SafePoint in the delay slot
if (delay_slot->is_MachSafePoint()) {
MachNode *mach = delay_slot->as_Mach();
// !!!!! Stubs only need an oopmap right now, so bail out
if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
// Write the oopmap directly to the code blob??!!
# ifdef ENABLE_ZAP_DEAD_LOCALS
assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
# endif
delay_slot = NULL;
continue;
}
int adjusted_offset = current_offset - Pipeline::instr_unit_size();
non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
adjusted_offset);
// Generate an OopMap entry
Process_OopMap_Node(mach, adjusted_offset);
}
// Insert the delay slot instruction
delay_slot->emit(*cb, _regalloc);
// Don't reuse it
delay_slot = NULL;
}
} // End for all instructions in block
// If the next block is the top of a loop, pad this block out to align
// the loop top a little. Helps prevent pipe stalls at loop back branches.
if (i < nblocks-1) {
Block *nb = _cfg->get_block(i + 1);
int padding = nb->alignment_padding(current_offset);
if( padding > 0 ) {
MachNode *nop = new (this) MachNopNode(padding / nop_size);
block->insert_node(nop, block->number_of_nodes());
_cfg->map_node_to_block(nop, block);
nop->emit(*cb, _regalloc);
current_offset = cb->insts_size();
}
}
// Verify that the distance for generated before forward
// short branches is still valid.
guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
// Save new block start offset
blk_starts[i] = blk_offset;
} // End of for all blocks
blk_starts[nblocks] = current_offset;
non_safepoints.flush_at_end();
// Offset too large?
if (failing()) return;
// Define a pseudo-label at the end of the code
MacroAssembler(cb).bind( blk_labels[nblocks] );
// Compute the size of the first block
_first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
assert(cb->insts_size() < 500000, "method is unreasonably large");
#ifdef ASSERT
for (uint i = 0; i < nblocks; i++) { // For all blocks
if (jmp_target[i] != 0) {
int br_size = jmp_size[i];
int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
assert(false, "Displacement too large for short jmp");
}
}
}
#endif
#ifndef PRODUCT
// Information on the size of the method, without the extraneous code
Scheduling::increment_method_size(cb->insts_size());
#endif
// ------------------
// Fill in exception table entries.
FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
// Only java methods have exception handlers and deopt handlers
if (_method) {
// Emit the exception handler code.
_code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
// Emit the deopt handler code.
_code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
// Emit the MethodHandle deopt handler code (if required).
if (has_method_handle_invokes()) {
// We can use the same code as for the normal deopt handler, we
// just need a different entry point address.
_code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
}
}
// One last check for failed CodeBuffer::expand:
if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
C->record_failure("CodeCache is full");
return;
}
#ifndef PRODUCT
// Dump the assembly code, including basic-block numbers
if (print_assembly()) {
ttyLocker ttyl; // keep the following output all in one block
if (!VMThread::should_terminate()) { // test this under the tty lock
// This output goes directly to the tty, not the compiler log.
// To enable tools to match it up with the compilation activity,
// be sure to tag this tty output with the compile ID.
if (xtty != NULL) {
xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
is_osr_compilation() ? " compile_kind='osr'" :
"");
}
if (method() != NULL) {
method()->print_metadata();
}
dump_asm(node_offsets, node_offset_limit);
if (xtty != NULL) {
xtty->tail("opto_assembly");
}
}
}
#endif
}
void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
_inc_table.set_size(cnt);
uint inct_cnt = 0;
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
Block* block = _cfg->get_block(i);
Node *n = NULL;
int j;
// Find the branch; ignore trailing NOPs.
for (j = block->number_of_nodes() - 1; j >= 0; j--) {
n = block->get_node(j);
if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
break;
}
}
// If we didn't find anything, continue
if (j < 0) {
continue;
}
// Compute ExceptionHandlerTable subtable entry and add it
// (skip empty blocks)
if (n->is_Catch()) {
// Get the offset of the return from the call
uint call_return = call_returns[block->_pre_order];
#ifdef ASSERT
assert( call_return > 0, "no call seen for this basic block" );
while (block->get_node(--j)->is_MachProj()) ;
assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
#endif
// last instruction is a CatchNode, find it's CatchProjNodes
int nof_succs = block->_num_succs;
// allocate space
GrowableArray<intptr_t> handler_bcis(nof_succs);
GrowableArray<intptr_t> handler_pcos(nof_succs);
// iterate through all successors
for (int j = 0; j < nof_succs; j++) {
Block* s = block->_succs[j];
bool found_p = false;
for (uint k = 1; k < s->num_preds(); k++) {
Node* pk = s->pred(k);
if (pk->is_CatchProj() && pk->in(0) == n) {
const CatchProjNode* p = pk->as_CatchProj();
found_p = true;
// add the corresponding handler bci & pco information
if (p->_con != CatchProjNode::fall_through_index) {
// p leads to an exception handler (and is not fall through)
assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
// no duplicates, please
if (!handler_bcis.contains(p->handler_bci())) {
uint block_num = s->non_connector()->_pre_order;
handler_bcis.append(p->handler_bci());
handler_pcos.append(blk_labels[block_num].loc_pos());
}
}
}
}
assert(found_p, "no matching predecessor found");
// Note: Due to empty block removal, one block may have
// several CatchProj inputs, from the same Catch.
}
// Set the offset of the return from the call
_handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
continue;
}
// Handle implicit null exception table updates
if (n->is_MachNullCheck()) {
uint block_num = block->non_connector_successor(0)->_pre_order;
_inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
continue;
}
// Handle implicit exception table updates: trap instructions.
if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
uint block_num = block->non_connector_successor(0)->_pre_order;
_inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
continue;
}
} // End of for all blocks fill in exception table entries
}
// Static Variables
#ifndef PRODUCT
uint Scheduling::_total_nop_size = 0;
uint Scheduling::_total_method_size = 0;
uint Scheduling::_total_branches = 0;
uint Scheduling::_total_unconditional_delays = 0;
uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
#endif
// Initializer for class Scheduling
Scheduling::Scheduling(Arena *arena, Compile &compile)
: _arena(arena),
_cfg(compile.cfg()),
_regalloc(compile.regalloc()),
_reg_node(arena),
_bundle_instr_count(0),
_bundle_cycle_number(0),
_scheduled(arena),
_available(arena),
_next_node(NULL),
_bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
_pinch_free_list(arena)
#ifndef PRODUCT
, _branches(0)
, _unconditional_delays(0)
#endif
{
// Create a MachNopNode
_nop = new (&compile) MachNopNode();
// Now that the nops are in the array, save the count
// (but allow entries for the nops)
_node_bundling_limit = compile.unique();
uint node_max = _regalloc->node_regs_max_index();
compile.set_node_bundling_limit(_node_bundling_limit);
// This one is persistent within the Compile class
_node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
// Allocate space for fixed-size arrays
_node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
_uses = NEW_ARENA_ARRAY(arena, short, node_max);
_current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
// Clear the arrays
memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
memset(_node_latency, 0, node_max * sizeof(unsigned short));
memset(_uses, 0, node_max * sizeof(short));
memset(_current_latency, 0, node_max * sizeof(unsigned short));
// Clear the bundling information
memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
// Get the last node
Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
_next_node = block->get_node(block->number_of_nodes() - 1);
}
#ifndef PRODUCT
// Scheduling destructor
Scheduling::~Scheduling() {
_total_branches += _branches;
_total_unconditional_delays += _unconditional_delays;
}
#endif
// Step ahead "i" cycles
void Scheduling::step(uint i) {
Bundle *bundle = node_bundling(_next_node);
bundle->set_starts_bundle();
// Update the bundle record, but leave the flags information alone
if (_bundle_instr_count > 0) {
bundle->set_instr_count(_bundle_instr_count);
bundle->set_resources_used(_bundle_use.resourcesUsed());
}
// Update the state information
_bundle_instr_count = 0;
_bundle_cycle_number += i;
_bundle_use.step(i);
}
void Scheduling::step_and_clear() {
Bundle *bundle = node_bundling(_next_node);
bundle->set_starts_bundle();
// Update the bundle record
if (_bundle_instr_count > 0) {
bundle->set_instr_count(_bundle_instr_count);
bundle->set_resources_used(_bundle_use.resourcesUsed());
_bundle_cycle_number += 1;
}
// Clear the bundling information
_bundle_instr_count = 0;
_bundle_use.reset();
memcpy(_bundle_use_elements,
Pipeline_Use::elaborated_elements,
sizeof(Pipeline_Use::elaborated_elements));
}
// Perform instruction scheduling and bundling over the sequence of
// instructions in backwards order.
void Compile::ScheduleAndBundle() {
// Don't optimize this if it isn't a method
if (!_method)
return;
// Don't optimize this if scheduling is disabled
if (!do_scheduling())
return;
// Scheduling code works only with pairs (8 bytes) maximum.
if (max_vector_size() > 8)
return;
NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
// Create a data structure for all the scheduling information
Scheduling scheduling(Thread::current()->resource_area(), *this);
// Walk backwards over each basic block, computing the needed alignment
// Walk over all the basic blocks
scheduling.DoScheduling();
}
// Compute the latency of all the instructions. This is fairly simple,
// because we already have a legal ordering. Walk over the instructions
// from first to last, and compute the latency of the instruction based
// on the latency of the preceding instruction(s).
void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# -> ComputeLocalLatenciesForward\n");
#endif
// Walk over all the schedulable instructions
for( uint j=_bb_start; j < _bb_end; j++ ) {
// This is a kludge, forcing all latency calculations to start at 1.
// Used to allow latency 0 to force an instruction to the beginning
// of the bb
uint latency = 1;
Node *use = bb->get_node(j);
uint nlen = use->len();
// Walk over all the inputs
for ( uint k=0; k < nlen; k++ ) {
Node *def = use->in(k);
if (!def)
continue;
uint l = _node_latency[def->_idx] + use->latency(k);
if (latency < l)
latency = l;
}
_node_latency[use->_idx] = latency;
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# latency %4d: ", latency);
use->dump();
}
#endif
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# <- ComputeLocalLatenciesForward\n");
#endif
} // end ComputeLocalLatenciesForward
// See if this node fits into the present instruction bundle
bool Scheduling::NodeFitsInBundle(Node *n) {
uint n_idx = n->_idx;
// If this is the unconditional delay instruction, then it fits
if (n == _unconditional_delay_slot) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
#endif
return (true);
}
// If the node cannot be scheduled this cycle, skip it
if (_current_latency[n_idx] > _bundle_cycle_number) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
n->_idx, _current_latency[n_idx], _bundle_cycle_number);
#endif
return (false);
}
const Pipeline *node_pipeline = n->pipeline();
uint instruction_count = node_pipeline->instructionCount();
if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
instruction_count = 0;
else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
instruction_count++;
if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
#endif
return (false);
}
// Don't allow non-machine nodes to be handled this way
if (!n->is_Mach() && instruction_count == 0)
return (false);
// See if there is any overlap
uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
if (delay > 0) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
#endif
return false;
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
#endif
return true;
}
Node * Scheduling::ChooseNodeToBundle() {
uint siz = _available.size();
if (siz == 0) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# ChooseNodeToBundle: NULL\n");
#endif
return (NULL);
}
// Fast path, if only 1 instruction in the bundle
if (siz == 1) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# ChooseNodeToBundle (only 1): ");
_available[0]->dump();
}
#endif
return (_available[0]);
}
// Don't bother, if the bundle is already full
if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
for ( uint i = 0; i < siz; i++ ) {
Node *n = _available[i];
// Skip projections, we'll handle them another way
if (n->is_Proj())
continue;
// This presupposed that instructions are inserted into the
// available list in a legality order; i.e. instructions that
// must be inserted first are at the head of the list
if (NodeFitsInBundle(n)) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# ChooseNodeToBundle: ");
n->dump();
}
#endif
return (n);
}
}
}
// Nothing fits in this bundle, choose the highest priority
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# ChooseNodeToBundle: ");
_available[0]->dump();
}
#endif
return _available[0];
}
void Scheduling::AddNodeToAvailableList(Node *n) {
assert( !n->is_Proj(), "projections never directly made available" );
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# AddNodeToAvailableList: ");
n->dump();
}
#endif
int latency = _current_latency[n->_idx];
// Insert in latency order (insertion sort)
uint i;
for ( i=0; i < _available.size(); i++ )
if (_current_latency[_available[i]->_idx] > latency)
break;
// Special Check for compares following branches
if( n->is_Mach() && _scheduled.size() > 0 ) {
int op = n->as_Mach()->ideal_Opcode();
Node *last = _scheduled[0];
if( last->is_MachIf() && last->in(1) == n &&
( op == Op_CmpI ||
op == Op_CmpU ||
op == Op_CmpP ||
op == Op_CmpF ||
op == Op_CmpD ||
op == Op_CmpL ) ) {
// Recalculate position, moving to front of same latency
for ( i=0 ; i < _available.size(); i++ )
if (_current_latency[_available[i]->_idx] >= latency)
break;
}
}
// Insert the node in the available list
_available.insert(i, n);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
dump_available();
#endif
}
void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
for ( uint i=0; i < n->len(); i++ ) {
Node *def = n->in(i);
if (!def) continue;
if( def->is_Proj() ) // If this is a machine projection, then
def = def->in(0); // propagate usage thru to the base instruction
if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
continue;
}
// Compute the latency
uint l = _bundle_cycle_number + n->latency(i);
if (_current_latency[def->_idx] < l)
_current_latency[def->_idx] = l;
// If this does not have uses then schedule it
if ((--_uses[def->_idx]) == 0)
AddNodeToAvailableList(def);
}
}
void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# AddNodeToBundle: ");
n->dump();
}
#endif
// Remove this from the available list
uint i;
for (i = 0; i < _available.size(); i++)
if (_available[i] == n)
break;
assert(i < _available.size(), "entry in _available list not found");
_available.remove(i);
// See if this fits in the current bundle
const Pipeline *node_pipeline = n->pipeline();
const Pipeline_Use& node_usage = node_pipeline->resourceUse();
// Check for instructions to be placed in the delay slot. We
// do this before we actually schedule the current instruction,
// because the delay slot follows the current instruction.
if (Pipeline::_branch_has_delay_slot &&
node_pipeline->hasBranchDelay() &&
!_unconditional_delay_slot) {
uint siz = _available.size();
// Conditional branches can support an instruction that
// is unconditionally executed and not dependent by the
// branch, OR a conditionally executed instruction if
// the branch is taken. In practice, this means that
// the first instruction at the branch target is
// copied to the delay slot, and the branch goes to
// the instruction after that at the branch target
if ( n->is_MachBranch() ) {
assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
assert( !n->is_Catch(), "should not look for delay slot for Catch" );
#ifndef PRODUCT
_branches++;
#endif
// At least 1 instruction is on the available list
// that is not dependent on the branch
for (uint i = 0; i < siz; i++) {
Node *d = _available[i];
const Pipeline *avail_pipeline = d->pipeline();
// Don't allow safepoints in the branch shadow, that will
// cause a number of difficulties
if ( avail_pipeline->instructionCount() == 1 &&
!avail_pipeline->hasMultipleBundles() &&
!avail_pipeline->hasBranchDelay() &&
Pipeline::instr_has_unit_size() &&
d->size(_regalloc) == Pipeline::instr_unit_size() &&
NodeFitsInBundle(d) &&
!node_bundling(d)->used_in_delay()) {
if (d->is_Mach() && !d->is_MachSafePoint()) {
// A node that fits in the delay slot was found, so we need to
// set the appropriate bits in the bundle pipeline information so
// that it correctly indicates resource usage. Later, when we
// attempt to add this instruction to the bundle, we will skip
// setting the resource usage.
_unconditional_delay_slot = d;
node_bundling(n)->set_use_unconditional_delay();
node_bundling(d)->set_used_in_unconditional_delay();
_bundle_use.add_usage(avail_pipeline->resourceUse());
_current_latency[d->_idx] = _bundle_cycle_number;
_next_node = d;
++_bundle_instr_count;
#ifndef PRODUCT
_unconditional_delays++;
#endif
break;
}
}
}
}
// No delay slot, add a nop to the usage
if (!_unconditional_delay_slot) {
// See if adding an instruction in the delay slot will overflow
// the bundle.
if (!NodeFitsInBundle(_nop)) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(1 instruction for delay slot) ***\n");
#endif
step(1);
}
_bundle_use.add_usage(_nop->pipeline()->resourceUse());
_next_node = _nop;
++_bundle_instr_count;
}
// See if the instruction in the delay slot requires a
// step of the bundles
if (!NodeFitsInBundle(n)) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(branch won't fit) ***\n");
#endif
// Update the state information
_bundle_instr_count = 0;
_bundle_cycle_number += 1;
_bundle_use.step(1);
}
}
// Get the number of instructions
uint instruction_count = node_pipeline->instructionCount();
if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
instruction_count = 0;
// Compute the latency information
uint delay = 0;
if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
if (relative_latency < 0)
relative_latency = 0;
delay = _bundle_use.full_latency(relative_latency, node_usage);
// Does not fit in this bundle, start a new one
if (delay > 0) {
step(delay);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(%d) ***\n", delay);
#endif
}
}
// If this was placed in the delay slot, ignore it
if (n != _unconditional_delay_slot) {
if (delay == 0) {
if (node_pipeline->hasMultipleBundles()) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(multiple instructions) ***\n");
#endif
step(1);
}
else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(%d >= %d instructions) ***\n",
instruction_count + _bundle_instr_count,
Pipeline::_max_instrs_per_cycle);
#endif
step(1);
}
}
if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
_bundle_instr_count++;
// Set the node's latency
_current_latency[n->_idx] = _bundle_cycle_number;
// Now merge the functional unit information
if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
_bundle_use.add_usage(node_usage);
// Increment the number of instructions in this bundle
_bundle_instr_count += instruction_count;
// Remember this node for later
if (n->is_Mach())
_next_node = n;
}
// It's possible to have a BoxLock in the graph and in the _bbs mapping but
// not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
// 'Schedule' them (basically ignore in the schedule) but do not insert them
// into the block. All other scheduled nodes get put in the schedule here.
int op = n->Opcode();
if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
(op != Op_Node && // Not an unused antidepedence node and
// not an unallocated boxlock
(OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
// Push any trailing projections
if( bb->get_node(bb->number_of_nodes()-1) != n ) {
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node *foi = n->fast_out(i);
if( foi->is_Proj() )
_scheduled.push(foi);
}
}
// Put the instruction in the schedule list
_scheduled.push(n);
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
dump_available();
#endif
// Walk all the definitions, decrementing use counts, and
// if a definition has a 0 use count, place it in the available list.
DecrementUseCounts(n,bb);
}
// This method sets the use count within a basic block. We will ignore all
// uses outside the current basic block. As we are doing a backwards walk,
// any node we reach that has a use count of 0 may be scheduled. This also
// avoids the problem of cyclic references from phi nodes, as long as phi
// nodes are at the front of the basic block. This method also initializes
// the available list to the set of instructions that have no uses within this
// basic block.
void Scheduling::ComputeUseCount(const Block *bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# -> ComputeUseCount\n");
#endif
// Clear the list of available and scheduled instructions, just in case
_available.clear();
_scheduled.clear();
// No delay slot specified
_unconditional_delay_slot = NULL;
#ifdef ASSERT
for( uint i=0; i < bb->number_of_nodes(); i++ )
assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
#endif
// Force the _uses count to never go to zero for unscheduable pieces
// of the block
for( uint k = 0; k < _bb_start; k++ )
_uses[bb->get_node(k)->_idx] = 1;
for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
_uses[bb->get_node(l)->_idx] = 1;
// Iterate backwards over the instructions in the block. Don't count the
// branch projections at end or the block header instructions.
for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
Node *n = bb->get_node(j);
if( n->is_Proj() ) continue; // Projections handled another way
// Account for all uses
for ( uint k = 0; k < n->len(); k++ ) {
Node *inp = n->in(k);
if (!inp) continue;
assert(inp != n, "no cycles allowed" );
if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
if (inp->is_Proj()) { // Skip through Proj's
inp = inp->in(0);
}
++_uses[inp->_idx]; // Count 1 block-local use
}
}
// If this instruction has a 0 use count, then it is available
if (!_uses[n->_idx]) {
_current_latency[n->_idx] = _bundle_cycle_number;
AddNodeToAvailableList(n);
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# uses: %3d: ", _uses[n->_idx]);
n->dump();
}
#endif
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# <- ComputeUseCount\n");
#endif
}
// This routine performs scheduling on each basic block in reverse order,
// using instruction latencies and taking into account function unit
// availability.
void Scheduling::DoScheduling() {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# -> DoScheduling\n");
#endif
Block *succ_bb = NULL;
Block *bb;
// Walk over all the basic blocks in reverse order
for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
bb = _cfg->get_block(i);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# Schedule BB#%03d (initial)\n", i);
for (uint j = 0; j < bb->number_of_nodes(); j++) {
bb->get_node(j)->dump();
}
}
#endif
// On the head node, skip processing
if (bb == _cfg->get_root_block()) {
continue;
}
// Skip empty, connector blocks
if (bb->is_connector())
continue;
// If the following block is not the sole successor of
// this one, then reset the pipeline information
if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
_next_node->_idx, _bundle_instr_count);
}
#endif
step_and_clear();
}
// Leave untouched the starting instruction, any Phis, a CreateEx node
// or Top. bb->get_node(_bb_start) is the first schedulable instruction.
_bb_end = bb->number_of_nodes()-1;
for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
Node *n = bb->get_node(_bb_start);
// Things not matched, like Phinodes and ProjNodes don't get scheduled.
// Also, MachIdealNodes do not get scheduled
if( !n->is_Mach() ) continue; // Skip non-machine nodes
MachNode *mach = n->as_Mach();
int iop = mach->ideal_Opcode();
if( iop == Op_CreateEx ) continue; // CreateEx is pinned
if( iop == Op_Con ) continue; // Do not schedule Top
if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
mach->pipeline() == MachNode::pipeline_class() &&
!n->is_SpillCopy() ) // Breakpoints, Prolog, etc
continue;
break; // Funny loop structure to be sure...
}
// Compute last "interesting" instruction in block - last instruction we
// might schedule. _bb_end points just after last schedulable inst. We
// normally schedule conditional branches (despite them being forced last
// in the block), because they have delay slots we can fill. Calls all
// have their delay slots filled in the template expansions, so we don't
// bother scheduling them.
Node *last = bb->get_node(_bb_end);
// Ignore trailing NOPs.
while (_bb_end > 0 && last->is_Mach() &&
last->as_Mach()->ideal_Opcode() == Op_Con) {
last = bb->get_node(--_bb_end);
}
assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
if( last->is_Catch() ||
// Exclude unreachable path case when Halt node is in a separate block.
(_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
// There must be a prior call. Skip it.
while( !bb->get_node(--_bb_end)->is_MachCall() ) {
assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
}
} else if( last->is_MachNullCheck() ) {
// Backup so the last null-checked memory instruction is
// outside the schedulable range. Skip over the nullcheck,
// projection, and the memory nodes.
Node *mem = last->in(1);
do {
_bb_end--;
} while (mem != bb->get_node(_bb_end));
} else {
// Set _bb_end to point after last schedulable inst.
_bb_end++;
}
assert( _bb_start <= _bb_end, "inverted block ends" );
// Compute the register antidependencies for the basic block
ComputeRegisterAntidependencies(bb);
if (_cfg->C->failing()) return; // too many D-U pinch points
// Compute intra-bb latencies for the nodes
ComputeLocalLatenciesForward(bb);
// Compute the usage within the block, and set the list of all nodes
// in the block that have no uses within the block.
ComputeUseCount(bb);
// Schedule the remaining instructions in the block
while ( _available.size() > 0 ) {
Node *n = ChooseNodeToBundle();
guarantee(n != NULL, "no nodes available");
AddNodeToBundle(n,bb);
}
assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
#ifdef ASSERT
for( uint l = _bb_start; l < _bb_end; l++ ) {
Node *n = bb->get_node(l);
uint m;
for( m = 0; m < _bb_end-_bb_start; m++ )
if( _scheduled[m] == n )
break;
assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
}
#endif
// Now copy the instructions (in reverse order) back to the block
for ( uint k = _bb_start; k < _bb_end; k++ )
bb->map_node(_scheduled[_bb_end-k-1], k);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# Schedule BB#%03d (final)\n", i);
uint current = 0;
for (uint j = 0; j < bb->number_of_nodes(); j++) {
Node *n = bb->get_node(j);
if( valid_bundle_info(n) ) {
Bundle *bundle = node_bundling(n);
if (bundle->instr_count() > 0 || bundle->flags() > 0) {
tty->print("*** Bundle: ");
bundle->dump();
}
n->dump();
}
}
}
#endif
#ifdef ASSERT
verify_good_schedule(bb,"after block local scheduling");
#endif
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# <- DoScheduling\n");
#endif
// Record final node-bundling array location
_regalloc->C->set_node_bundling_base(_node_bundling_base);
} // end DoScheduling
// Verify that no live-range used in the block is killed in the block by a
// wrong DEF. This doesn't verify live-ranges that span blocks.
// Check for edge existence. Used to avoid adding redundant precedence edges.
static bool edge_from_to( Node *from, Node *to ) {
for( uint i=0; i<from->len(); i++ )
if( from->in(i) == to )
return true;
return false;
}
#ifdef ASSERT
void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
// Check for bad kills
if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
Node *prior_use = _reg_node[def];
if( prior_use && !edge_from_to(prior_use,n) ) {
tty->print("%s = ",OptoReg::as_VMReg(def)->name());
n->dump();
tty->print_cr("...");
prior_use->dump();
assert(edge_from_to(prior_use,n),msg);
}
_reg_node.map(def,NULL); // Kill live USEs
}
}
void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
// Zap to something reasonable for the verify code
_reg_node.clear();
// Walk over the block backwards. Check to make sure each DEF doesn't
// kill a live value (other than the one it's supposed to). Add each
// USE to the live set.
for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
Node *n = b->get_node(i);
int n_op = n->Opcode();
if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
// Fat-proj kills a slew of registers
RegMask rm = n->out_RegMask();// Make local copy
while( rm.is_NotEmpty() ) {
OptoReg::Name kill = rm.find_first_elem();
rm.Remove(kill);
verify_do_def( n, kill, msg );
}
} else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
// Get DEF'd registers the normal way
verify_do_def( n, _regalloc->get_reg_first(n), msg );
verify_do_def( n, _regalloc->get_reg_second(n), msg );
}
// Now make all USEs live
for( uint i=1; i<n->req(); i++ ) {
Node *def = n->in(i);
assert(def != 0, "input edge required");
OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
if( OptoReg::is_valid(reg_lo) ) {
assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
_reg_node.map(reg_lo,n);
}
if( OptoReg::is_valid(reg_hi) ) {
assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
_reg_node.map(reg_hi,n);
}
}
}
// Zap to something reasonable for the Antidependence code
_reg_node.clear();
}
#endif
// Conditionally add precedence edges. Avoid putting edges on Projs.
static void add_prec_edge_from_to( Node *from, Node *to ) {
if( from->is_Proj() ) { // Put precedence edge on Proj's input
assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
from = from->in(0);
}
if( from != to && // No cycles (for things like LD L0,[L0+4] )
!edge_from_to( from, to ) ) // Avoid duplicate edge
from->add_prec(to);
}
void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
return;
Node *pinch = _reg_node[def_reg]; // Get pinch point
if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
is_def ) { // Check for a true def (not a kill)
_reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
return;
}
Node *kill = def; // Rename 'def' to more descriptive 'kill'
debug_only( def = (Node*)0xdeadbeef; )
// After some number of kills there _may_ be a later def
Node *later_def = NULL;
// Finding a kill requires a real pinch-point.
// Check for not already having a pinch-point.
// Pinch points are Op_Node's.
if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
later_def = pinch; // Must be def/kill as optimistic pinch-point
if ( _pinch_free_list.size() > 0) {
pinch = _pinch_free_list.pop();
} else {
pinch = new (_cfg->C) Node(1); // Pinch point to-be
}
if (pinch->_idx >= _regalloc->node_regs_max_index()) {
_cfg->C->record_method_not_compilable("too many D-U pinch points");
return;
}
_cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
_reg_node.map(def_reg,pinch); // Record pinch-point
//_regalloc->set_bad(pinch->_idx); // Already initialized this way.
if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
later_def = NULL; // and no later def
}
pinch->set_req(0,later_def); // Hook later def so we can find it
} else { // Else have valid pinch point
if( pinch->in(0) ) // If there is a later-def
later_def = pinch->in(0); // Get it
}
// Add output-dependence edge from later def to kill
if( later_def ) // If there is some original def
add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
// See if current kill is also a use, and so is forced to be the pinch-point.
if( pinch->Opcode() == Op_Node ) {
Node *uses = kill->is_Proj() ? kill->in(0) : kill;
for( uint i=1; i<uses->req(); i++ ) {
if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
_regalloc->get_reg_second(uses->in(i)) == def_reg ) {
// Yes, found a use/kill pinch-point
pinch->set_req(0,NULL); //
pinch->replace_by(kill); // Move anti-dep edges up
pinch = kill;
_reg_node.map(def_reg,pinch);
return;
}
}
}
// Add edge from kill to pinch-point
add_prec_edge_from_to(kill,pinch);
}
void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
return;
Node *pinch = _reg_node[use_reg]; // Get pinch point
// Check for no later def_reg/kill in block
if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
// Use has to be block-local as well
_cfg->get_block_for_node(use) == b) {
if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
pinch->req() == 1 ) { // pinch not yet in block?
pinch->del_req(0); // yank pointer to later-def, also set flag
// Insert the pinch-point in the block just after the last use
b->insert_node(pinch, b->find_node(use) + 1);
_bb_end++; // Increase size scheduled region in block
}
add_prec_edge_from_to(pinch,use);
}
}
// We insert antidependences between the reads and following write of
// allocated registers to prevent illegal code motion. Hopefully, the
// number of added references should be fairly small, especially as we
// are only adding references within the current basic block.
void Scheduling::ComputeRegisterAntidependencies(Block *b) {
#ifdef ASSERT
verify_good_schedule(b,"before block local scheduling");
#endif
// A valid schedule, for each register independently, is an endless cycle
// of: a def, then some uses (connected to the def by true dependencies),
// then some kills (defs with no uses), finally the cycle repeats with a new
// def. The uses are allowed to float relative to each other, as are the
// kills. No use is allowed to slide past a kill (or def). This requires
// antidependencies between all uses of a single def and all kills that
// follow, up to the next def. More edges are redundant, because later defs
// & kills are already serialized with true or antidependencies. To keep
// the edge count down, we add a 'pinch point' node if there's more than
// one use or more than one kill/def.
// We add dependencies in one bottom-up pass.
// For each instruction we handle it's DEFs/KILLs, then it's USEs.
// For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
// register. If not, we record the DEF/KILL in _reg_node, the
// register-to-def mapping. If there is a prior DEF/KILL, we insert a
// "pinch point", a new Node that's in the graph but not in the block.
// We put edges from the prior and current DEF/KILLs to the pinch point.
// We put the pinch point in _reg_node. If there's already a pinch point
// we merely add an edge from the current DEF/KILL to the pinch point.
// After doing the DEF/KILLs, we handle USEs. For each used register, we
// put an edge from the pinch point to the USE.
// To be expedient, the _reg_node array is pre-allocated for the whole
// compilation. _reg_node is lazily initialized; it either contains a NULL,
// or a valid def/kill/pinch-point, or a leftover node from some prior
// block. Leftover node from some prior block is treated like a NULL (no
// prior def, so no anti-dependence needed). Valid def is distinguished by
// it being in the current block.
bool fat_proj_seen = false;
uint last_safept = _bb_end-1;
Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
Node* last_safept_node = end_node;
for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
Node *n = b->get_node(i);
int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
// Fat-proj kills a slew of registers
// This can add edges to 'n' and obscure whether or not it was a def,
// hence the is_def flag.
fat_proj_seen = true;
RegMask rm = n->out_RegMask();// Make local copy
while( rm.is_NotEmpty() ) {
OptoReg::Name kill = rm.find_first_elem();
rm.Remove(kill);
anti_do_def( b, n, kill, is_def );
}
} else {
// Get DEF'd registers the normal way
anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
}
// Kill projections on a branch should appear to occur on the
// branch, not afterwards, so grab the masks from the projections
// and process them.
if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* use = n->fast_out(i);
if (use->is_Proj()) {
RegMask rm = use->out_RegMask();// Make local copy
while( rm.is_NotEmpty() ) {
OptoReg::Name kill = rm.find_first_elem();
rm.Remove(kill);
anti_do_def( b, n, kill, false );
}
}
}
}
// Check each register used by this instruction for a following DEF/KILL
// that must occur afterward and requires an anti-dependence edge.
for( uint j=0; j<n->req(); j++ ) {
Node *def = n->in(j);
if( def ) {
assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
anti_do_use( b, n, _regalloc->get_reg_first(def) );
anti_do_use( b, n, _regalloc->get_reg_second(def) );
}
}
// Do not allow defs of new derived values to float above GC
// points unless the base is definitely available at the GC point.
Node *m = b->get_node(i);
// Add precedence edge from following safepoint to use of derived pointer
if( last_safept_node != end_node &&
m != last_safept_node) {
for (uint k = 1; k < m->req(); k++) {
const Type *t = m->in(k)->bottom_type();
if( t->isa_oop_ptr() &&
t->is_ptr()->offset() != 0 ) {
last_safept_node->add_prec( m );
break;
}
}
}
if( n->jvms() ) { // Precedence edge from derived to safept
// Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
if( b->get_node(last_safept) != last_safept_node ) {
last_safept = b->find_node(last_safept_node);
}
for( uint j=last_safept; j > i; j-- ) {
Node *mach = b->get_node(j);
if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
mach->add_prec( n );
}
last_safept = i;
last_safept_node = m;
}
}
if (fat_proj_seen) {
// Garbage collect pinch nodes that were not consumed.
// They are usually created by a fat kill MachProj for a call.
garbage_collect_pinch_nodes();
}
}
// Garbage collect pinch nodes for reuse by other blocks.
//
// The block scheduler's insertion of anti-dependence
// edges creates many pinch nodes when the block contains
// 2 or more Calls. A pinch node is used to prevent a
// combinatorial explosion of edges. If a set of kills for a
// register is anti-dependent on a set of uses (or defs), rather
// than adding an edge in the graph between each pair of kill
// and use (or def), a pinch is inserted between them:
//
// use1 use2 use3
// \ | /
// \ | /
// pinch
// / | \
// / | \
// kill1 kill2 kill3
//
// One pinch node is created per register killed when
// the second call is encountered during a backwards pass
// over the block. Most of these pinch nodes are never
// wired into the graph because the register is never
// used or def'ed in the block.
//
void Scheduling::garbage_collect_pinch_nodes() {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
#endif
int trace_cnt = 0;
for (uint k = 0; k < _reg_node.Size(); k++) {
Node* pinch = _reg_node[k];
if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
// no predecence input edges
(pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
cleanup_pinch(pinch);
_pinch_free_list.push(pinch);
_reg_node.map(k, NULL);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
trace_cnt++;
if (trace_cnt > 40) {
tty->print("\n");
trace_cnt = 0;
}
tty->print(" %d", pinch->_idx);
}
#endif
}
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) tty->print("\n");
#endif
}
// Clean up a pinch node for reuse.
void Scheduling::cleanup_pinch( Node *pinch ) {
assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
Node* use = pinch->last_out(i);
uint uses_found = 0;
for (uint j = use->req(); j < use->len(); j++) {
if (use->in(j) == pinch) {
use->rm_prec(j);
uses_found++;
}
}
assert(uses_found > 0, "must be a precedence edge");
i -= uses_found; // we deleted 1 or more copies of this edge
}
// May have a later_def entry
pinch->set_req(0, NULL);
}
#ifndef PRODUCT
void Scheduling::dump_available() const {
tty->print("#Availist ");
for (uint i = 0; i < _available.size(); i++)
tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
tty->cr();
}
// Print Scheduling Statistics
void Scheduling::print_statistics() {
// Print the size added by nops for bundling
tty->print("Nops added %d bytes to total of %d bytes",
_total_nop_size, _total_method_size);
if (_total_method_size > 0)
tty->print(", for %.2f%%",
((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
tty->print("\n");
// Print the number of branch shadows filled
if (Pipeline::_branch_has_delay_slot) {
tty->print("Of %d branches, %d had unconditional delay slots filled",
_total_branches, _total_unconditional_delays);
if (_total_branches > 0)
tty->print(", for %.2f%%",
((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
tty->print("\n");
}
uint total_instructions = 0, total_bundles = 0;
for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
uint bundle_count = _total_instructions_per_bundle[i];
total_instructions += bundle_count * i;
total_bundles += bundle_count;
}
if (total_bundles > 0)
tty->print("Average ILP (excluding nops) is %.2f\n",
((double)total_instructions) / ((double)total_bundles));
}
#endif