2
* Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5
* This code is free software; you can redistribute it and/or modify it
6
* under the terms of the GNU General Public License version 2 only, as
7
* published by the Free Software Foundation.
9
* This code is distributed in the hope that it will be useful, but WITHOUT
10
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12
* version 2 for more details (a copy is included in the LICENSE file that
13
* accompanied this code).
15
* You should have received a copy of the GNU General Public License version
16
* 2 along with this work; if not, write to the Free Software Foundation,
17
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20
* or visit www.oracle.com if you need additional information or have any
25
#include "precompiled.hpp"
26
#include "asm/assembler.inline.hpp"
27
#include "asm/macroAssembler.inline.hpp"
28
#include "code/compiledIC.hpp"
29
#include "code/debugInfo.hpp"
30
#include "code/debugInfoRec.hpp"
31
#include "compiler/compileBroker.hpp"
32
#include "compiler/compilerDirectives.hpp"
33
#include "compiler/disassembler.hpp"
34
#include "compiler/oopMap.hpp"
35
#include "gc/shared/barrierSet.hpp"
36
#include "gc/shared/c2/barrierSetC2.hpp"
37
#include "memory/allocation.inline.hpp"
38
#include "memory/allocation.hpp"
40
#include "opto/block.hpp"
41
#include "opto/c2compiler.hpp"
42
#include "opto/c2_MacroAssembler.hpp"
43
#include "opto/callnode.hpp"
44
#include "opto/cfgnode.hpp"
45
#include "opto/locknode.hpp"
46
#include "opto/machnode.hpp"
47
#include "opto/node.hpp"
48
#include "opto/optoreg.hpp"
49
#include "opto/output.hpp"
50
#include "opto/regalloc.hpp"
51
#include "opto/runtime.hpp"
52
#include "opto/subnode.hpp"
53
#include "opto/type.hpp"
54
#include "runtime/handles.inline.hpp"
55
#include "runtime/sharedRuntime.hpp"
56
#include "utilities/macros.hpp"
57
#include "utilities/powerOfTwo.hpp"
58
#include "utilities/xmlstream.hpp"
61
#define DEBUG_ARG(x) , x
66
//------------------------------Scheduling----------------------------------
67
// This class contains all the information necessary to implement instruction
68
// scheduling and bundling.
75
// Control-Flow Graph info
78
// Register Allocation info
79
PhaseRegAlloc *_regalloc;
81
// Number of nodes in the method
82
uint _node_bundling_limit;
84
// List of scheduled nodes. Generated in reverse order
87
// List of nodes currently available for choosing for scheduling
90
// For each instruction beginning a bundle, the number of following
91
// nodes to be bundled with it.
92
Bundle *_node_bundling_base;
94
// Mapping from register to Node
97
// Free list for pinch nodes.
98
Node_List _pinch_free_list;
100
// Number of uses of this node within the containing basic block.
103
// Schedulable portion of current block. Skips Region/Phi/CreateEx up
104
// front, branch+proj at end. Also skips Catch/CProj (same as
105
// branch-at-end), plus just-prior exception-throwing call.
106
uint _bb_start, _bb_end;
108
// Latency from the end of the basic block as scheduled
109
unsigned short *_current_latency;
111
// Remember the next node
114
// Use this for an unconditional branch delay slot
115
Node *_unconditional_delay_slot;
120
// Length of the current bundle, in instructions
121
uint _bundle_instr_count;
123
// Current Cycle number, for computing latencies and bundling
124
uint _bundle_cycle_number;
126
// Bundle information
127
Pipeline_Use_Element _bundle_use_elements[resource_count];
128
Pipeline_Use _bundle_use;
130
// Dump the available list
131
void dump_available() const;
134
Scheduling(Arena *arena, Compile &compile);
137
NOT_PRODUCT( ~Scheduling(); )
139
// Step ahead "i" cycles
142
// Step ahead 1 cycle, and clear the bundle state (for example,
143
// at a branch target)
144
void step_and_clear();
146
Bundle* node_bundling(const Node *n) {
147
assert(valid_bundle_info(n), "oob");
148
return (&_node_bundling_base[n->_idx]);
151
bool valid_bundle_info(const Node *n) const {
152
return (_node_bundling_limit > n->_idx);
155
bool starts_bundle(const Node *n) const {
156
return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
162
// Compute the register antidependencies within a basic block
163
void ComputeRegisterAntidependencies(Block *bb);
164
void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
165
void verify_good_schedule( Block *b, const char *msg );
166
void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
167
void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
169
// Add a node to the current bundle
170
void AddNodeToBundle(Node *n, const Block *bb);
172
// Return an integer less than, equal to, or greater than zero
173
// if the stack offset of the first argument is respectively
174
// less than, equal to, or greater than the second.
175
int compare_two_spill_nodes(Node* first, Node* second);
177
// Add a node to the list of available nodes
178
void AddNodeToAvailableList(Node *n);
180
// Compute the local use count for the nodes in a block, and compute
181
// the list of instructions with no uses in the block as available
182
void ComputeUseCount(const Block *bb);
184
// Choose an instruction from the available list to add to the bundle
185
Node * ChooseNodeToBundle();
187
// See if this Node fits into the currently accumulating bundle
188
bool NodeFitsInBundle(Node *n);
190
// Decrement the use count for a node
191
void DecrementUseCounts(Node *n, const Block *bb);
193
// Garbage collect pinch nodes for reuse by other blocks.
194
void garbage_collect_pinch_nodes();
195
// Clean up a pinch node for reuse (helper for above).
196
void cleanup_pinch( Node *pinch );
198
// Information for statistics gathering
201
// Gather information on size of nops relative to total
202
uint _branches, _unconditional_delays;
204
static uint _total_nop_size, _total_method_size;
205
static uint _total_branches, _total_unconditional_delays;
206
static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
209
static void print_statistics();
211
static void increment_instructions_per_bundle(uint i) {
212
_total_instructions_per_bundle[i]++;
215
static void increment_nop_size(uint s) {
216
_total_nop_size += s;
219
static void increment_method_size(uint s) {
220
_total_method_size += s;
226
PhaseOutput::PhaseOutput()
227
: Phase(Phase::Output),
228
_code_buffer("Compile::Fill_buffer"),
229
_first_block_size(0),
233
_oop_map_set(nullptr),
234
_scratch_buffer_blob(nullptr),
235
_scratch_locs_memory(nullptr),
236
_scratch_const_size(-1),
237
_in_scratch_emit_size(false),
240
_node_bundling_limit(0),
241
_node_bundling_base(nullptr),
243
_orig_pc_slot_offset_in_bytes(0),
248
if (C->stub_name() == nullptr) {
249
_orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
253
PhaseOutput::~PhaseOutput() {
254
C->set_output(nullptr);
255
if (_scratch_buffer_blob != nullptr) {
256
BufferBlob::free(_scratch_buffer_blob);
260
void PhaseOutput::perform_mach_node_analysis() {
261
// Late barrier analysis must be done after schedule and bundle
262
// Otherwise liveness based spilling will fail
263
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
264
bs->late_barrier_analysis();
266
pd_perform_mach_node_analysis();
268
C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
271
// Convert Nodes to instruction bits and pass off to the VM
272
void PhaseOutput::Output() {
274
assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
276
// The number of new nodes (mostly MachNop) is proportional to
277
// the number of java calls and inner loops which are aligned.
278
if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
279
C->inner_loops()*(OptoLoopAlignment-1)),
280
"out of nodes before code generation" ) ) {
283
// Make sure I can find the Start Node
284
Block *entry = C->cfg()->get_block(1);
285
Block *broot = C->cfg()->get_root_block();
287
const StartNode *start = entry->head()->as_Start();
289
// Replace StartNode with prolog
290
MachPrologNode *prolog = new MachPrologNode();
291
entry->map_node(prolog, 0);
292
C->cfg()->map_node_to_block(prolog, entry);
293
C->cfg()->unmap_node_from_block(start); // start is no longer in any block
295
// Virtual methods need an unverified entry point
297
if( C->is_osr_compilation() ) {
298
if( PoisonOSREntry ) {
299
// TODO: Should use a ShouldNotReachHereNode...
300
C->cfg()->insert( broot, 0, new MachBreakpointNode() );
303
if( C->method() && !C->method()->flags().is_static() ) {
304
// Insert unvalidated entry point
305
C->cfg()->insert( broot, 0, new MachUEPNode() );
310
// Break before main entry point
311
if ((C->method() && C->directive()->BreakAtExecuteOption) ||
312
(OptoBreakpoint && C->is_method_compilation()) ||
313
(OptoBreakpointOSR && C->is_osr_compilation()) ||
314
(OptoBreakpointC2R && !C->method()) ) {
315
// checking for C->method() means that OptoBreakpoint does not apply to
316
// runtime stubs or frame converters
317
C->cfg()->insert( entry, 1, new MachBreakpointNode() );
320
// Insert epilogs before every return
321
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
322
Block* block = C->cfg()->get_block(i);
323
if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
324
Node* m = block->end();
325
if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
326
MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
327
block->add_inst(epilog);
328
C->cfg()->map_node_to_block(epilog, block);
333
// Keeper of sizing aspects
334
_buf_sizes = BufferSizingData();
336
// Initialize code buffer
337
estimate_buffer_size(_buf_sizes._const);
338
if (C->failing()) return;
340
// Pre-compute the length of blocks and replace
341
// long branches with short if machine supports it.
342
// Must be done before ScheduleAndBundle due to SPARC delay slots
343
uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
345
shorten_branches(blk_starts);
352
perform_mach_node_analysis();
354
// Complete sizing of codebuffer
355
CodeBuffer* cb = init_buffer();
356
if (cb == nullptr || C->failing()) {
366
C2_MacroAssembler masm(cb);
367
fill_buffer(&masm, blk_starts);
370
bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
371
// Determine if we need to generate a stack overflow check.
372
// Do it if the method is not a stub function and
373
// has java calls or has frame size > vm_page_size/8.
374
// The debug VM checks that deoptimization doesn't trigger an
375
// unexpected stack overflow (compiled method stack banging should
376
// guarantee it doesn't happen) so we always need the stack bang in
378
return (C->stub_function() == nullptr &&
379
(C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
380
DEBUG_ONLY(|| true)));
383
bool PhaseOutput::need_register_stack_bang() const {
384
// Determine if we need to generate a register stack overflow check.
385
// This is only used on architectures which have split register
386
// and memory stacks (ie. IA64).
387
// Bang if the method is not a stub function and has java calls
388
return (C->stub_function() == nullptr && C->has_java_calls());
392
// Compute the size of first NumberOfLoopInstrToAlign instructions at the top
393
// of a loop. When aligning a loop we need to provide enough instructions
394
// in cpu's fetch buffer to feed decoders. The loop alignment could be
395
// avoided if we have enough instructions in fetch buffer at the head of a loop.
396
// By default, the size is set to 999999 by Block's constructor so that
397
// a loop will be aligned if the size is not reset here.
399
// Note: Mach instructions could contain several HW instructions
400
// so the size is estimated only.
402
void PhaseOutput::compute_loop_first_inst_sizes() {
403
// The next condition is used to gate the loop alignment optimization.
404
// Don't aligned a loop if there are enough instructions at the head of a loop
405
// or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
406
// is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
407
// equal to 11 bytes which is the largest address NOP instruction.
408
if (MaxLoopPad < OptoLoopAlignment - 1) {
409
uint last_block = C->cfg()->number_of_blocks() - 1;
410
for (uint i = 1; i <= last_block; i++) {
411
Block* block = C->cfg()->get_block(i);
412
// Check the first loop's block which requires an alignment.
413
if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
415
uint inst_cnt = NumberOfLoopInstrToAlign;
416
inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
418
// Check subsequent fallthrough blocks if the loop's first
419
// block(s) does not have enough instructions.
421
while(inst_cnt > 0 &&
423
!C->cfg()->get_block(i + 1)->has_loop_alignment() &&
424
!nb->has_successor(block)) {
426
nb = C->cfg()->get_block(i);
427
inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
428
} // while( inst_cnt > 0 && i < last_block )
430
block->set_first_inst_size(sum_size);
431
} // f( b->head()->is_Loop() )
432
} // for( i <= last_block )
433
} // if( MaxLoopPad < OptoLoopAlignment-1 )
436
// The architecture description provides short branch variants for some long
437
// branch instructions. Replace eligible long branches with short branches.
438
void PhaseOutput::shorten_branches(uint* blk_starts) {
440
Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
442
// Compute size of each block, method size, and relocation information size
443
uint nblocks = C->cfg()->number_of_blocks();
445
uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
446
uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
447
int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
449
// Collect worst case block paddings
450
int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
451
memset(block_worst_case_pad, 0, nblocks * sizeof(int));
453
DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
454
DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
456
bool has_short_branch_candidate = false;
458
// Initialize the sizes to 0
459
int code_size = 0; // Size in bytes of generated code
460
int stub_size = 0; // Size in bytes of all stub entries
461
// Size in bytes of all relocation entries, including those in local stubs.
462
// Start with 2-bytes of reloc info for the unvalidated entry point
463
int reloc_size = 1; // Number of relocation entries
465
// Make three passes. The first computes pessimistic blk_starts,
466
// relative jmp_offset and reloc_size information. The second performs
467
// short branch substitution using the pessimistic sizing. The
468
// third inserts nops where needed.
470
// Step one, perform a pessimistic sizing pass.
471
uint last_call_adr = max_juint;
472
uint last_avoid_back_to_back_adr = max_juint;
473
uint nop_size = (new MachNopNode())->size(C->regalloc());
474
for (uint i = 0; i < nblocks; i++) { // For all blocks
475
Block* block = C->cfg()->get_block(i);
478
// During short branch replacement, we store the relative (to blk_starts)
479
// offset of jump in jmp_offset, rather than the absolute offset of jump.
480
// This is so that we do not need to recompute sizes of all nodes when
481
// we compute correct blk_starts in our next sizing pass.
485
DEBUG_ONLY( jmp_target[i] = 0; )
486
DEBUG_ONLY( jmp_rule[i] = 0; )
488
// Sum all instruction sizes to compute block size
489
uint last_inst = block->number_of_nodes();
491
for (uint j = 0; j < last_inst; j++) {
493
Node* nj = block->get_node(_index);
494
// Handle machine instruction nodes
496
MachNode* mach = nj->as_Mach();
497
blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
498
reloc_size += mach->reloc();
499
if (mach->is_MachCall()) {
500
// add size information for trampoline stub
501
// class CallStubImpl is platform-specific and defined in the *.ad files.
502
stub_size += CallStubImpl::size_call_trampoline();
503
reloc_size += CallStubImpl::reloc_call_trampoline();
505
MachCallNode *mcall = mach->as_MachCall();
506
// This destination address is NOT PC-relative
508
mcall->method_set((intptr_t)mcall->entry_point());
510
if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
511
stub_size += CompiledDirectCall::to_interp_stub_size();
512
reloc_size += CompiledDirectCall::reloc_to_interp_stub();
514
} else if (mach->is_MachSafePoint()) {
515
// If call/safepoint are adjacent, account for possible
516
// nop to disambiguate the two safepoints.
517
// ScheduleAndBundle() can rearrange nodes in a block,
518
// check for all offsets inside this block.
519
if (last_call_adr >= blk_starts[i]) {
520
blk_size += nop_size;
523
if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
524
// Nop is inserted between "avoid back to back" instructions.
525
// ScheduleAndBundle() can rearrange nodes in a block,
526
// check for all offsets inside this block.
527
if (last_avoid_back_to_back_adr >= blk_starts[i]) {
528
blk_size += nop_size;
531
if (mach->may_be_short_branch()) {
532
if (!nj->is_MachBranch()) {
538
assert(jmp_nidx[i] == -1, "block should have only one branch");
539
jmp_offset[i] = blk_size;
540
jmp_size[i] = nj->size(C->regalloc());
542
has_short_branch_candidate = true;
545
blk_size += nj->size(C->regalloc());
546
// Remember end of call offset
547
if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
548
last_call_adr = blk_starts[i]+blk_size;
550
// Remember end of avoid_back_to_back offset
551
if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
552
last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
556
// When the next block starts a loop, we may insert pad NOP
557
// instructions. Since we cannot know our future alignment,
559
if (i < nblocks - 1) {
560
Block* nb = C->cfg()->get_block(i + 1);
561
int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
562
if (max_loop_pad > 0) {
563
assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
564
// Adjust last_call_adr and/or last_avoid_back_to_back_adr.
565
// If either is the last instruction in this block, bump by
566
// max_loop_pad in lock-step with blk_size, so sizing
567
// calculations in subsequent blocks still can conservatively
568
// detect that it may the last instruction in this block.
569
if (last_call_adr == blk_starts[i]+blk_size) {
570
last_call_adr += max_loop_pad;
572
if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
573
last_avoid_back_to_back_adr += max_loop_pad;
575
blk_size += max_loop_pad;
576
block_worst_case_pad[i + 1] = max_loop_pad;
580
// Save block size; update total method size
581
blk_starts[i+1] = blk_starts[i]+blk_size;
584
// Step two, replace eligible long jumps.
585
bool progress = true;
586
uint last_may_be_short_branch_adr = max_juint;
587
while (has_short_branch_candidate && progress) {
589
has_short_branch_candidate = false;
590
int adjust_block_start = 0;
591
for (uint i = 0; i < nblocks; i++) {
592
Block* block = C->cfg()->get_block(i);
593
int idx = jmp_nidx[i];
594
MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
595
if (mach != nullptr && mach->may_be_short_branch()) {
597
assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
599
// Find the branch; ignore trailing NOPs.
600
for (j = block->number_of_nodes()-1; j>=0; j--) {
601
Node* n = block->get_node(j);
602
if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
605
assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
607
int br_size = jmp_size[i];
608
int br_offs = blk_starts[i] + jmp_offset[i];
610
// This requires the TRUE branch target be in succs[0]
611
uint bnum = block->non_connector_successor(0)->_pre_order;
612
int offset = blk_starts[bnum] - br_offs;
613
if (bnum > i) { // adjust following block's offset
614
offset -= adjust_block_start;
617
// This block can be a loop header, account for the padding
618
// in the previous block.
619
int block_padding = block_worst_case_pad[i];
620
assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
621
// In the following code a nop could be inserted before
622
// the branch which will increase the backward distance.
623
bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
624
assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
626
if (needs_padding && offset <= 0)
629
if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
630
// We've got a winner. Replace this branch.
631
MachNode* replacement = mach->as_MachBranch()->short_branch_version();
633
// Update the jmp_size.
634
int new_size = replacement->size(C->regalloc());
635
int diff = br_size - new_size;
636
assert(diff >= (int)nop_size, "short_branch size should be smaller");
637
// Conservatively take into account padding between
638
// avoid_back_to_back branches. Previous branch could be
639
// converted into avoid_back_to_back branch during next
641
if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
642
jmp_offset[i] += nop_size;
645
adjust_block_start += diff;
646
block->map_node(replacement, idx);
647
mach->subsume_by(replacement, C);
651
jmp_size[i] = new_size;
652
DEBUG_ONLY( jmp_target[i] = bnum; );
653
DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
655
// The jump distance is not short, try again during next iteration.
656
has_short_branch_candidate = true;
658
} // (mach->may_be_short_branch())
659
if (mach != nullptr && (mach->may_be_short_branch() ||
660
mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
661
last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
663
blk_starts[i+1] -= adjust_block_start;
668
for (uint i = 0; i < nblocks; i++) { // For all blocks
669
if (jmp_target[i] != 0) {
670
int br_size = jmp_size[i];
671
int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
672
if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
673
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]);
675
assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
680
// Step 3, compute the offsets of all blocks, will be done in fill_buffer()
681
// after ScheduleAndBundle().
683
// ------------------
684
// Compute size for code buffer
685
code_size = blk_starts[nblocks];
687
// Relocation records
688
reloc_size += 1; // Relo entry for exception handler
690
// Adjust reloc_size to number of record of relocation info
691
// Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
692
// a relocation index.
693
// The CodeBuffer will expand the locs array if this estimate is too low.
694
reloc_size *= 10 / sizeof(relocInfo);
696
_buf_sizes._reloc = reloc_size;
697
_buf_sizes._code = code_size;
698
_buf_sizes._stub = stub_size;
701
//------------------------------FillLocArray-----------------------------------
702
// Create a bit of debug info and append it to the array. The mapping is from
703
// Java local or expression stack to constant, register or stack-slot. For
704
// doubles, insert 2 mappings and return 1 (to tell the caller that the next
705
// entry has been taken care of and caller should skip it).
706
static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
707
// This should never have accepted Bad before
708
assert(OptoReg::is_valid(regnum), "location must be valid");
709
return (OptoReg::is_reg(regnum))
710
? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
711
: new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
716
PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
717
for (int i = 0; i < objs->length(); i++) {
718
assert(objs->at(i)->is_object(), "corrupt object cache");
719
ObjectValue* sv = (ObjectValue*) objs->at(i);
720
if (sv->id() == id) {
728
void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
730
assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
735
void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
736
GrowableArray<ScopeValue*> *array,
737
GrowableArray<ScopeValue*> *objs ) {
738
assert( local, "use _top instead of null" );
739
if (array->length() != idx) {
740
assert(array->length() == idx + 1, "Unexpected array count");
741
// Old functionality:
743
// New functionality:
744
// Assert if the local is not top. In product mode let the new node
745
// override the old entry.
746
assert(local == C->top(), "LocArray collision");
747
if (local == C->top()) {
752
const Type *t = local->bottom_type();
754
// Is it a safepoint scalar object node?
755
if (local->is_SafePointScalarObject()) {
756
SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
758
ObjectValue* sv = (ObjectValue*) sv_for_node_id(objs, spobj->_idx);
760
ciKlass* cik = t->is_oopptr()->exact_klass();
761
assert(cik->is_instance_klass() ||
762
cik->is_array_klass(), "Not supported allocation.");
763
sv = new ObjectValue(spobj->_idx,
764
new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
765
set_sv_for_object_node(objs, sv);
767
uint first_ind = spobj->first_index(sfpt->jvms());
768
for (uint i = 0; i < spobj->n_fields(); i++) {
769
Node* fld_node = sfpt->in(first_ind+i);
770
(void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
775
} else if (local->is_SafePointScalarMerge()) {
776
SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
777
ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
780
GrowableArray<ScopeValue*> deps;
782
int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
783
(void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
784
assert(deps.length() == 1, "missing value");
786
int selector_idx = smerge->selector_idx(sfpt->jvms());
787
(void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
788
assert(deps.length() == 2, "missing value");
790
mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
791
set_sv_for_object_node(objs, mv);
793
for (uint i = 1; i < smerge->req(); i++) {
794
Node* obj_node = smerge->in(i);
795
(void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs);
802
// Grab the register number for the local
803
OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
804
if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
805
// Record the double as two float registers.
806
// The register mask for such a value always specifies two adjacent
807
// float registers, with the lower register number even.
808
// Normally, the allocation of high and low words to these registers
809
// is irrelevant, because nearly all operations on register pairs
810
// (e.g., StoreD) treat them as a single unit.
811
// Here, we assume in addition that the words in these two registers
812
// stored "naturally" (by operations like StoreD and double stores
813
// within the interpreter) such that the lower-numbered register
814
// is written to the lower memory address. This may seem like
815
// a machine dependency, but it is not--it is a requirement on
816
// the author of the <arch>.ad file to ensure that, for every
817
// even/odd double-register pair to which a double may be allocated,
818
// the word in the even single-register is stored to the first
819
// memory word. (Note that register numbers are completely
820
// arbitrary, and are not tied to any machine-level encodings.)
822
if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
823
array->append(new ConstantIntValue((jint)0));
824
array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
825
} else if ( t->base() == Type::Long ) {
826
array->append(new ConstantIntValue((jint)0));
827
array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
828
} else if ( t->base() == Type::RawPtr ) {
829
// jsr/ret return address which must be restored into the full
830
// width 64-bit stack slot.
831
array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
834
if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
835
// Repack the double/long as two jints.
836
// The convention the interpreter uses is that the second local
837
// holds the first raw word of the native double representation.
838
// This is actually reasonable, since locals and stack arrays
839
// grow downwards in all implementations.
840
// (If, on some machine, the interpreter's Java locals or stack
841
// were to grow upwards, the embedded doubles would be word-swapped.)
842
array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
843
array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
846
else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
847
OptoReg::is_reg(regnum) ) {
848
array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
849
? Location::float_in_dbl : Location::normal ));
850
} else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
851
array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
852
? Location::int_in_long : Location::normal ));
853
} else if( t->base() == Type::NarrowOop ) {
854
array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
855
} else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
856
t->base() == Type::VectorD || t->base() == Type::VectorX ||
857
t->base() == Type::VectorY || t->base() == Type::VectorZ) {
858
array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
859
} else if (C->regalloc()->is_oop(local)) {
860
assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
861
t->base() == Type::AryPtr,
862
"Unexpected type: %s", t->msg());
863
array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
865
assert(t->base() == Type::Int || t->base() == Type::Half ||
866
t->base() == Type::FloatCon || t->base() == Type::FloatBot,
867
"Unexpected type: %s", t->msg());
868
array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
873
// No register. It must be constant data.
875
case Type::Half: // Second half of a double
876
ShouldNotReachHere(); // Caller should skip 2nd halves
879
array->append(new ConstantOopWriteValue(nullptr));
882
case Type::InstPtr: // fall through
883
array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
885
case Type::NarrowOop:
886
if (t == TypeNarrowOop::NULL_PTR) {
887
array->append(new ConstantOopWriteValue(nullptr));
889
array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
893
array->append(new ConstantIntValue(t->is_int()->get_con()));
896
// A return address (T_ADDRESS).
897
assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
899
// Must be restored to the full-width 64-bit stack slot.
900
array->append(new ConstantLongValue(t->is_ptr()->get_con()));
902
array->append(new ConstantIntValue(t->is_ptr()->get_con()));
905
case Type::FloatCon: {
906
float f = t->is_float_constant()->getf();
907
array->append(new ConstantIntValue(jint_cast(f)));
910
case Type::DoubleCon: {
911
jdouble d = t->is_double_constant()->getd();
913
array->append(new ConstantIntValue((jint)0));
914
array->append(new ConstantDoubleValue(d));
916
// Repack the double as two jints.
917
// The convention the interpreter uses is that the second local
918
// holds the first raw word of the native double representation.
919
// This is actually reasonable, since locals and stack arrays
920
// grow downwards in all implementations.
921
// (If, on some machine, the interpreter's Java locals or stack
922
// were to grow upwards, the embedded doubles would be word-swapped.)
924
acc.long_value = jlong_cast(d);
925
array->append(new ConstantIntValue(acc.words[1]));
926
array->append(new ConstantIntValue(acc.words[0]));
931
jlong d = t->is_long()->get_con();
933
array->append(new ConstantIntValue((jint)0));
934
array->append(new ConstantLongValue(d));
936
// Repack the long as two jints.
937
// The convention the interpreter uses is that the second local
938
// holds the first raw word of the native double representation.
939
// This is actually reasonable, since locals and stack arrays
940
// grow downwards in all implementations.
941
// (If, on some machine, the interpreter's Java locals or stack
942
// were to grow upwards, the embedded doubles would be word-swapped.)
945
array->append(new ConstantIntValue(acc.words[1]));
946
array->append(new ConstantIntValue(acc.words[0]));
950
case Type::Top: // Add an illegal value here
951
array->append(new LocationValue(Location()));
954
ShouldNotReachHere();
959
// Determine if this node starts a bundle
960
bool PhaseOutput::starts_bundle(const Node *n) const {
961
return (_node_bundling_limit > n->_idx &&
962
_node_bundling_base[n->_idx].starts_bundle());
965
// Determine if there is a monitor that has 'ov' as its owner.
966
bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
967
for (int k = 0; k < monarray->length(); k++) {
968
MonitorValue* mv = monarray->at(k);
969
if (mv->owner() == ov) {
977
// Determine if there is a scalar replaced object description represented by 'ov'.
978
bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
979
GrowableArray<ScopeValue*>* objs,
980
ObjectValue* ov) const {
981
for (int i = 0; i < jvms->scl_size(); i++) {
982
Node* n = sfn->scalarized_obj(jvms, i);
983
// Other kinds of nodes that we may encounter here, for instance constants
984
// representing values of fields of objects scalarized, aren't relevant for
985
// us, since they don't map to ObjectValue.
986
if (!n->is_SafePointScalarObject()) {
990
ObjectValue* other = (ObjectValue*) sv_for_node_id(objs, n->_idx);
998
//--------------------------Process_OopMap_Node--------------------------------
999
void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1000
// Handle special safepoint nodes for synchronization
1001
MachSafePointNode *sfn = mach->as_MachSafePoint();
1002
MachCallNode *mcall;
1004
int safepoint_pc_offset = current_offset;
1005
bool is_method_handle_invoke = false;
1006
bool return_oop = false;
1007
bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1008
bool arg_escape = false;
1010
// Add the safepoint in the DebugInfoRecorder
1011
if( !mach->is_MachCall() ) {
1013
C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1015
mcall = mach->as_MachCall();
1017
// Is the call a MethodHandle call?
1018
if (mcall->is_MachCallJava()) {
1019
if (mcall->as_MachCallJava()->_method_handle_invoke) {
1020
assert(C->has_method_handle_invokes(), "must have been set during call generation");
1021
is_method_handle_invoke = true;
1023
arg_escape = mcall->as_MachCallJava()->_arg_escape;
1026
// Check if a call returns an object.
1027
if (mcall->returns_pointer()) {
1030
safepoint_pc_offset += mcall->ret_addr_offset();
1031
C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1034
// Loop over the JVMState list to add scope information
1035
// Do not skip safepoints with a null method, they need monitor info
1036
JVMState* youngest_jvms = sfn->jvms();
1037
int max_depth = youngest_jvms->depth();
1039
// Allocate the object pool for scalar-replaced objects -- the map from
1040
// small-integer keys (which can be recorded in the local and ostack
1041
// arrays) to descriptions of the object state.
1042
GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1044
// Visit scopes from oldest to youngest.
1045
for (int depth = 1; depth <= max_depth; depth++) {
1046
JVMState* jvms = youngest_jvms->of_depth(depth);
1048
ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1049
// Safepoints that do not have method() set only provide oop-map and monitor info
1050
// to support GC; these do not support deoptimization.
1051
int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1052
int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1053
int num_mon = jvms->nof_monitors();
1054
assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1055
"JVMS local count must match that of the method");
1057
// Add Local and Expression Stack Information
1059
// Insert locals into the locarray
1060
GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1061
for( idx = 0; idx < num_locs; idx++ ) {
1062
FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1065
// Insert expression stack entries into the exparray
1066
GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1067
for( idx = 0; idx < num_exps; idx++ ) {
1068
FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1071
// Add in mappings of the monitors
1073
!method->is_synchronized() ||
1074
method->is_native() ||
1076
!GenerateSynchronizationCode,
1077
"monitors must always exist for synchronized methods");
1079
// Build the growable array of ScopeValues for exp stack
1080
GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1082
// Loop over monitors and insert into array
1083
for (idx = 0; idx < num_mon; idx++) {
1084
// Grab the node that defines this monitor
1085
Node* box_node = sfn->monitor_box(jvms, idx);
1086
Node* obj_node = sfn->monitor_obj(jvms, idx);
1088
// Create ScopeValue for object
1089
ScopeValue *scval = nullptr;
1091
if (obj_node->is_SafePointScalarObject()) {
1092
SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1093
scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1094
if (scval == nullptr) {
1095
const Type *t = spobj->bottom_type();
1096
ciKlass* cik = t->is_oopptr()->exact_klass();
1097
assert(cik->is_instance_klass() ||
1098
cik->is_array_klass(), "Not supported allocation.");
1099
ObjectValue* sv = new ObjectValue(spobj->_idx,
1100
new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1101
PhaseOutput::set_sv_for_object_node(objs, sv);
1103
uint first_ind = spobj->first_index(youngest_jvms);
1104
for (uint i = 0; i < spobj->n_fields(); i++) {
1105
Node* fld_node = sfn->in(first_ind+i);
1106
(void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1110
} else if (obj_node->is_SafePointScalarMerge()) {
1111
SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1112
ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1114
if (mv == nullptr) {
1115
GrowableArray<ScopeValue*> deps;
1117
int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1118
FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1119
assert(deps.length() == 1, "missing value");
1121
int selector_idx = smerge->selector_idx(youngest_jvms);
1122
FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1123
assert(deps.length() == 2, "missing value");
1125
mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1126
set_sv_for_object_node(objs, mv);
1128
for (uint i = 1; i < smerge->req(); i++) {
1129
Node* obj_node = smerge->in(i);
1130
FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs);
1134
} else if (!obj_node->is_Con()) {
1135
OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1136
if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1137
scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1139
scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1142
const TypePtr *tp = obj_node->get_ptr_type();
1143
scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1146
OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1147
Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1148
bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1149
monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1152
// Mark ObjectValue nodes as root nodes if they are directly
1153
// referenced in the JVMS.
1154
for (int i = 0; i < objs->length(); i++) {
1155
ScopeValue* sv = objs->at(i);
1156
if (sv->is_object_merge()) {
1157
ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1159
for (int j = 0; j< merge->possible_objects()->length(); j++) {
1160
ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1161
bool is_root = locarray->contains(ov) ||
1162
exparray->contains(ov) ||
1163
contains_as_owner(monarray, ov) ||
1164
contains_as_scalarized_obj(jvms, sfn, objs, ov);
1165
ov->set_root(is_root);
1170
// We dump the object pool first, since deoptimization reads it in first.
1171
C->debug_info()->dump_object_pool(objs);
1173
// Build first class objects to pass to scope
1174
DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1175
DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1176
DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1178
// Make method available for all Safepoints
1179
ciMethod* scope_method = method ? method : C->method();
1180
// Describe the scope here
1181
assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1182
assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1183
// Now we can describe the scope.
1184
methodHandle null_mh;
1185
bool rethrow_exception = false;
1186
C->debug_info()->describe_scope(
1187
safepoint_pc_offset,
1191
jvms->should_reexecute(),
1193
is_method_handle_invoke,
1195
has_ea_local_in_scope,
1203
// Mark the end of the scope set.
1204
C->debug_info()->end_safepoint(safepoint_pc_offset);
1209
// A simplified version of Process_OopMap_Node, to handle non-safepoints.
1210
class NonSafepointEmitter {
1212
JVMState* _pending_jvms;
1213
int _pending_offset;
1215
void emit_non_safepoint();
1218
NonSafepointEmitter(Compile* compile) {
1220
_pending_jvms = nullptr;
1221
_pending_offset = 0;
1224
void observe_instruction(Node* n, int pc_offset) {
1225
if (!C->debug_info()->recording_non_safepoints()) return;
1227
Node_Notes* nn = C->node_notes_at(n->_idx);
1228
if (nn == nullptr || nn->jvms() == nullptr) return;
1229
if (_pending_jvms != nullptr &&
1230
_pending_jvms->same_calls_as(nn->jvms())) {
1231
// Repeated JVMS? Stretch it up here.
1232
_pending_offset = pc_offset;
1234
if (_pending_jvms != nullptr &&
1235
_pending_offset < pc_offset) {
1236
emit_non_safepoint();
1238
_pending_jvms = nullptr;
1239
if (pc_offset > C->debug_info()->last_pc_offset()) {
1240
// This is the only way _pending_jvms can become non-null:
1241
_pending_jvms = nn->jvms();
1242
_pending_offset = pc_offset;
1247
// Stay out of the way of real safepoints:
1248
void observe_safepoint(JVMState* jvms, int pc_offset) {
1249
if (_pending_jvms != nullptr &&
1250
!_pending_jvms->same_calls_as(jvms) &&
1251
_pending_offset < pc_offset) {
1252
emit_non_safepoint();
1254
_pending_jvms = nullptr;
1257
void flush_at_end() {
1258
if (_pending_jvms != nullptr) {
1259
emit_non_safepoint();
1261
_pending_jvms = nullptr;
1265
void NonSafepointEmitter::emit_non_safepoint() {
1266
JVMState* youngest_jvms = _pending_jvms;
1267
int pc_offset = _pending_offset;
1270
_pending_jvms = nullptr;
1272
DebugInformationRecorder* debug_info = C->debug_info();
1273
assert(debug_info->recording_non_safepoints(), "sanity");
1275
debug_info->add_non_safepoint(pc_offset);
1276
int max_depth = youngest_jvms->depth();
1278
// Visit scopes from oldest to youngest.
1279
for (int depth = 1; depth <= max_depth; depth++) {
1280
JVMState* jvms = youngest_jvms->of_depth(depth);
1281
ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1282
assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1283
methodHandle null_mh;
1284
debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1287
// Mark the end of the scope set.
1288
debug_info->end_non_safepoint(pc_offset);
1291
//------------------------------init_buffer------------------------------------
1292
void PhaseOutput::estimate_buffer_size(int& const_req) {
1294
// Set the initially allocated size
1295
const_req = initial_const_capacity;
1297
// The extra spacing after the code is necessary on some platforms.
1298
// Sometimes we need to patch in a jump after the last instruction,
1299
// if the nmethod has been deoptimized. (See 4932387, 4894843.)
1301
// Compute the byte offset where we can store the deopt pc.
1302
if (C->fixed_slots() != 0) {
1303
_orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1306
// Compute prolog code size
1307
_frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1308
assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1310
if (C->has_mach_constant_base_node()) {
1312
// Fill the constant table.
1313
// Note: This must happen before shorten_branches.
1314
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1315
Block* b = C->cfg()->get_block(i);
1317
for (uint j = 0; j < b->number_of_nodes(); j++) {
1318
Node* n = b->get_node(j);
1320
// If the node is a MachConstantNode evaluate the constant
1322
if (n->is_MachConstant()) {
1323
MachConstantNode* machcon = n->as_MachConstant();
1324
machcon->eval_constant(C);
1325
} else if (n->is_Mach()) {
1326
// On Power there are more nodes that issue constants.
1327
add_size += (n->as_Mach()->ins_num_consts() * 8);
1332
// Calculate the offsets of the constants and the size of the
1333
// constant table (including the padding to the next section).
1334
constant_table().calculate_offsets_and_size();
1335
const_req = constant_table().size() + add_size;
1338
// Initialize the space for the BufferBlob used to find and verify
1339
// instruction size in MachNode::emit_size()
1340
init_scratch_buffer_blob(const_req);
1343
CodeBuffer* PhaseOutput::init_buffer() {
1344
int stub_req = _buf_sizes._stub;
1345
int code_req = _buf_sizes._code;
1346
int const_req = _buf_sizes._const;
1348
int pad_req = NativeCall::byte_size();
1350
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1351
stub_req += bs->estimate_stub_size();
1353
// nmethod and CodeBuffer count stubs & constants as part of method's code.
1354
// class HandlerImpl is platform-specific and defined in the *.ad files.
1355
int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1356
int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1357
stub_req += MAX_stubs_size; // ensure per-stub margin
1358
code_req += MAX_inst_size; // ensure per-instruction margin
1360
if (StressCodeBuffers)
1361
code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1368
exception_handler_req +
1369
deopt_handler_req; // deopt handler
1371
if (C->has_method_handle_invokes())
1372
total_req += deopt_handler_req; // deopt MH handler
1374
CodeBuffer* cb = code_buffer();
1375
cb->initialize(total_req, _buf_sizes._reloc);
1377
// Have we run out of code space?
1378
if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1379
C->record_failure("CodeCache is full");
1382
// Configure the code buffer.
1383
cb->initialize_consts_size(const_req);
1384
cb->initialize_stubs_size(stub_req);
1385
cb->initialize_oop_recorder(C->env()->oop_recorder());
1387
// fill in the nop array for bundling computations
1388
MachNode *_nop_list[Bundle::_nop_count];
1389
Bundle::initialize_nops(_nop_list);
1394
//------------------------------fill_buffer------------------------------------
1395
void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1396
// blk_starts[] contains offsets calculated during short branches processing,
1397
// offsets should not be increased during following steps.
1399
// Compute the size of first NumberOfLoopInstrToAlign instructions at head
1400
// of a loop. It is used to determine the padding for loop alignment.
1401
Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1403
compute_loop_first_inst_sizes();
1405
// Create oopmap set.
1406
_oop_map_set = new OopMapSet();
1408
// !!!!! This preserves old handling of oopmaps for now
1409
C->debug_info()->set_oopmaps(_oop_map_set);
1411
uint nblocks = C->cfg()->number_of_blocks();
1412
// Count and start of implicit null check instructions
1414
uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1416
// Count and start of calls
1417
uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1419
uint return_offset = 0;
1420
int nop_size = (new MachNopNode())->size(C->regalloc());
1422
int previous_offset = 0;
1423
int current_offset = 0;
1424
int last_call_offset = -1;
1425
int last_avoid_back_to_back_offset = -1;
1427
uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1428
uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1429
uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1430
uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1433
// Create an array of unused labels, one for each basic block, if printing is enabled
1434
#if defined(SUPPORT_OPTO_ASSEMBLY)
1435
int* node_offsets = nullptr;
1436
uint node_offset_limit = C->unique();
1438
if (C->print_assembly()) {
1439
node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1441
if (node_offsets != nullptr) {
1442
// We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1443
memset(node_offsets, 0, node_offset_limit*sizeof(int));
1447
NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1449
// Emit the constant table.
1450
if (C->has_mach_constant_base_node()) {
1451
if (!constant_table().emit(masm)) {
1452
C->record_failure("consts section overflow");
1457
// Create an array of labels, one for each basic block
1458
Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1459
for (uint i = 0; i <= nblocks; i++) {
1460
blk_labels[i].init();
1463
// Now fill in the code buffer
1464
Node* delay_slot = nullptr;
1465
for (uint i = 0; i < nblocks; i++) {
1466
Block* block = C->cfg()->get_block(i);
1468
Node* head = block->head();
1470
// If this block needs to start aligned (i.e, can be reached other
1471
// than by falling-thru from the previous block), then force the
1472
// start of a new bundle.
1473
if (Pipeline::requires_bundling() && starts_bundle(head)) {
1474
masm->code()->flush_bundle(true);
1478
if (!block->is_connector()) {
1480
block->dump_head(C->cfg(), &st);
1481
masm->block_comment(st.freeze());
1488
int blk_offset = current_offset;
1490
// Define the label at the beginning of the basic block
1491
masm->bind(blk_labels[block->_pre_order]);
1493
uint last_inst = block->number_of_nodes();
1495
// Emit block normally, except for last instruction.
1496
// Emit means "dump code bits into code buffer".
1497
for (uint j = 0; j<last_inst; j++) {
1501
Node* n = block->get_node(j);
1503
// See if delay slots are supported
1504
if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1505
assert(delay_slot == nullptr, "no use of delay slot node");
1506
assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1512
// If this starts a new instruction group, then flush the current one
1513
// (but allow split bundles)
1514
if (Pipeline::requires_bundling() && starts_bundle(n))
1515
masm->code()->flush_bundle(false);
1517
// Special handling for SafePoint/Call Nodes
1518
bool is_mcall = false;
1520
MachNode *mach = n->as_Mach();
1521
is_mcall = n->is_MachCall();
1522
bool is_sfn = n->is_MachSafePoint();
1524
// If this requires all previous instructions be flushed, then do so
1525
if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1526
masm->code()->flush_bundle(true);
1527
current_offset = masm->offset();
1530
// A padding may be needed again since a previous instruction
1531
// could be moved to delay slot.
1533
// align the instruction if necessary
1534
int padding = mach->compute_padding(current_offset);
1535
// Make sure safepoint node for polling is distinct from a call's
1536
// return by adding a nop if needed.
1537
if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1540
if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1541
current_offset == last_avoid_back_to_back_offset) {
1542
// Avoid back to back some instructions.
1547
assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1548
int nops_cnt = padding / nop_size;
1549
MachNode *nop = new MachNopNode(nops_cnt);
1550
block->insert_node(nop, j++);
1552
C->cfg()->map_node_to_block(nop, block);
1553
// Ensure enough space.
1554
masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1555
if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1556
C->record_failure("CodeCache is full");
1559
nop->emit(masm, C->regalloc());
1560
masm->code()->flush_bundle(true);
1561
current_offset = masm->offset();
1564
bool observe_safepoint = is_sfn;
1565
// Remember the start of the last call in a basic block
1567
MachCallNode *mcall = mach->as_MachCall();
1569
// This destination address is NOT PC-relative
1570
mcall->method_set((intptr_t)mcall->entry_point());
1572
// Save the return address
1573
call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1575
observe_safepoint = mcall->guaranteed_safepoint();
1578
// sfn will be valid whenever mcall is valid now because of inheritance
1579
if (observe_safepoint) {
1580
// Handle special safepoint nodes for synchronization
1582
MachSafePointNode *sfn = mach->as_MachSafePoint();
1583
// !!!!! Stubs only need an oopmap right now, so bail out
1584
if (sfn->jvms()->method() == nullptr) {
1585
// Write the oopmap directly to the code blob??!!
1588
} // End synchronization
1590
non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1592
Process_OopMap_Node(mach, current_offset);
1593
} // End if safepoint
1595
// If this is a null check, then add the start of the previous instruction to the list
1596
else if( mach->is_MachNullCheck() ) {
1597
inct_starts[inct_cnt++] = previous_offset;
1600
// If this is a branch, then fill in the label with the target BB's label
1601
else if (mach->is_MachBranch()) {
1602
// This requires the TRUE branch target be in succs[0]
1603
uint block_num = block->non_connector_successor(0)->_pre_order;
1605
// Try to replace long branch if delay slot is not used,
1606
// it is mostly for back branches since forward branch's
1607
// distance is not updated yet.
1608
bool delay_slot_is_used = valid_bundle_info(n) &&
1609
C->output()->node_bundling(n)->use_unconditional_delay();
1610
if (!delay_slot_is_used && mach->may_be_short_branch()) {
1611
assert(delay_slot == nullptr, "not expecting delay slot node");
1612
int br_size = n->size(C->regalloc());
1613
int offset = blk_starts[block_num] - current_offset;
1614
if (block_num >= i) {
1615
// Current and following block's offset are not
1616
// finalized yet, adjust distance by the difference
1617
// between calculated and final offsets of current block.
1618
offset -= (blk_starts[i] - blk_offset);
1620
// In the following code a nop could be inserted before
1621
// the branch which will increase the backward distance.
1622
bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1623
if (needs_padding && offset <= 0)
1626
if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1627
// We've got a winner. Replace this branch.
1628
MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1630
// Update the jmp_size.
1631
int new_size = replacement->size(C->regalloc());
1632
assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1633
// Insert padding between avoid_back_to_back branches.
1634
if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1635
MachNode *nop = new MachNopNode();
1636
block->insert_node(nop, j++);
1637
C->cfg()->map_node_to_block(nop, block);
1639
nop->emit(masm, C->regalloc());
1640
masm->code()->flush_bundle(true);
1641
current_offset = masm->offset();
1644
jmp_target[i] = block_num;
1645
jmp_offset[i] = current_offset - blk_offset;
1646
jmp_size[i] = new_size;
1647
jmp_rule[i] = mach->rule();
1649
block->map_node(replacement, j);
1650
mach->subsume_by(replacement, C);
1655
mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1656
} else if (mach->ideal_Opcode() == Op_Jump) {
1657
for (uint h = 0; h < block->_num_succs; h++) {
1658
Block* succs_block = block->_succs[h];
1659
for (uint j = 1; j < succs_block->num_preds(); j++) {
1660
Node* jpn = succs_block->pred(j);
1661
if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1662
uint block_num = succs_block->non_connector()->_pre_order;
1663
Label *blkLabel = &blk_labels[block_num];
1664
mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1670
// Check that oop-store precedes the card-mark
1671
else if (mach->ideal_Opcode() == Op_StoreCM) {
1672
uint storeCM_idx = j;
1674
for (uint prec = mach->req(); prec < mach->len(); prec++) {
1675
Node *oop_store = mach->in(prec); // Precedence edge
1676
if (oop_store == nullptr) continue;
1679
for (i4 = 0; i4 < last_inst; ++i4) {
1680
if (block->get_node(i4) == oop_store) {
1684
// Note: This test can provide a false failure if other precedence
1685
// edges have been added to the storeCMNode.
1686
assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1688
assert(count > 0, "storeCM expects at least one precedence edge");
1691
else if (!n->is_Proj()) {
1692
// Remember the beginning of the previous instruction, in case
1693
// it's followed by a flag-kill and a null-check. Happens on
1694
// Intel all the time, with add-to-memory kind of opcodes.
1695
previous_offset = current_offset;
1699
// If this is a trap based cmp then add its offset to the list.
1700
if (mach->is_TrapBasedCheckNode()) {
1701
inct_starts[inct_cnt++] = current_offset;
1705
// Verify that there is sufficient space remaining
1706
masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1707
if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1708
C->record_failure("CodeCache is full");
1712
// Save the offset for the listing
1713
#if defined(SUPPORT_OPTO_ASSEMBLY)
1714
if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1715
node_offsets[n->_idx] = masm->offset();
1718
assert(!C->failing(), "Should not reach here if failing.");
1720
// "Normal" instruction case
1721
DEBUG_ONLY(uint instr_offset = masm->offset());
1722
n->emit(masm, C->regalloc());
1723
current_offset = masm->offset();
1725
// Above we only verified that there is enough space in the instruction section.
1726
// However, the instruction may emit stubs that cause code buffer expansion.
1727
// Bail out here if expansion failed due to a lack of code cache space.
1732
assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1733
"ret_addr_offset() not within emitted code");
1736
uint n_size = n->size(C->regalloc());
1737
if (n_size < (current_offset-instr_offset)) {
1738
MachNode* mach = n->as_Mach();
1740
mach->dump_format(C->regalloc(), tty);
1741
tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1742
Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1743
tty->print_cr(" ------------------- ");
1744
BufferBlob* blob = this->scratch_buffer_blob();
1745
address blob_begin = blob->content_begin();
1746
Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1747
assert(false, "wrong size of mach node");
1750
non_safepoints.observe_instruction(n, current_offset);
1752
// mcall is last "call" that can be a safepoint
1753
// record it so we can see if a poll will directly follow it
1754
// in which case we'll need a pad to make the PcDesc sites unique
1755
// see 5010568. This can be slightly inaccurate but conservative
1756
// in the case that return address is not actually at current_offset.
1757
// This is a small price to pay.
1760
last_call_offset = current_offset;
1763
if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1764
// Avoid back to back some instructions.
1765
last_avoid_back_to_back_offset = current_offset;
1768
// See if this instruction has a delay slot
1769
if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1770
guarantee(delay_slot != nullptr, "expecting delay slot node");
1772
// Back up 1 instruction
1773
masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1775
// Save the offset for the listing
1776
#if defined(SUPPORT_OPTO_ASSEMBLY)
1777
if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1778
node_offsets[delay_slot->_idx] = masm->offset();
1782
// Support a SafePoint in the delay slot
1783
if (delay_slot->is_MachSafePoint()) {
1784
MachNode *mach = delay_slot->as_Mach();
1785
// !!!!! Stubs only need an oopmap right now, so bail out
1786
if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1787
// Write the oopmap directly to the code blob??!!
1788
delay_slot = nullptr;
1792
int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1793
non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1795
// Generate an OopMap entry
1796
Process_OopMap_Node(mach, adjusted_offset);
1799
// Insert the delay slot instruction
1800
delay_slot->emit(masm, C->regalloc());
1803
delay_slot = nullptr;
1806
} // End for all instructions in block
1808
// If the next block is the top of a loop, pad this block out to align
1809
// the loop top a little. Helps prevent pipe stalls at loop back branches.
1810
if (i < nblocks-1) {
1811
Block *nb = C->cfg()->get_block(i + 1);
1812
int padding = nb->alignment_padding(current_offset);
1814
MachNode *nop = new MachNopNode(padding / nop_size);
1815
block->insert_node(nop, block->number_of_nodes());
1816
C->cfg()->map_node_to_block(nop, block);
1817
nop->emit(masm, C->regalloc());
1818
current_offset = masm->offset();
1821
// Verify that the distance for generated before forward
1822
// short branches is still valid.
1823
guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1825
// Save new block start offset
1826
blk_starts[i] = blk_offset;
1827
} // End of for all blocks
1828
blk_starts[nblocks] = current_offset;
1830
non_safepoints.flush_at_end();
1832
// Offset too large?
1833
if (C->failing()) return;
1835
// Define a pseudo-label at the end of the code
1836
masm->bind( blk_labels[nblocks] );
1838
// Compute the size of the first block
1839
_first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1842
for (uint i = 0; i < nblocks; i++) { // For all blocks
1843
if (jmp_target[i] != 0) {
1844
int br_size = jmp_size[i];
1845
int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1846
if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1847
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]);
1848
assert(false, "Displacement too large for short jmp");
1854
if (!masm->code()->finalize_stubs()) {
1855
C->record_failure("CodeCache is full");
1859
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1860
bs->emit_stubs(*masm->code());
1861
if (C->failing()) return;
1864
assert(masm->inst_mark() == nullptr, "should be.");
1865
_stub_list.emit(*masm);
1866
if (C->failing()) return;
1869
// Information on the size of the method, without the extraneous code
1870
Scheduling::increment_method_size(masm->offset());
1873
// ------------------
1874
// Fill in exception table entries.
1875
FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1877
// Only java methods have exception handlers and deopt handlers
1878
// class HandlerImpl is platform-specific and defined in the *.ad files.
1880
// Emit the exception handler code.
1881
_code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1883
return; // CodeBuffer::expand failed
1885
// Emit the deopt handler code.
1886
_code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1888
// Emit the MethodHandle deopt handler code (if required).
1889
if (C->has_method_handle_invokes() && !C->failing()) {
1890
// We can use the same code as for the normal deopt handler, we
1891
// just need a different entry point address.
1892
_code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1896
// One last check for failed CodeBuffer::expand:
1897
if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1898
C->record_failure("CodeCache is full");
1902
#if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1903
if (C->print_assembly()) {
1905
tty->print_cr("============================= C2-compiled nmethod ==============================");
1909
#if defined(SUPPORT_OPTO_ASSEMBLY)
1910
// Dump the assembly code, including basic-block numbers
1911
if (C->print_assembly()) {
1912
ttyLocker ttyl; // keep the following output all in one block
1913
if (!VMThread::should_terminate()) { // test this under the tty lock
1914
// print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1915
// We call them first and write to a stringStream, then we retake the lock to
1916
// make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1918
stringStream method_metadata_str;
1919
if (C->method() != nullptr) {
1920
C->method()->print_metadata(&method_metadata_str);
1922
stringStream dump_asm_str;
1923
dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1925
NoSafepointVerifier nsv;
1927
// This output goes directly to the tty, not the compiler log.
1928
// To enable tools to match it up with the compilation activity,
1929
// be sure to tag this tty output with the compile ID.
1930
if (xtty != nullptr) {
1931
xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1932
C->is_osr_compilation() ? " compile_kind='osr'" : "");
1934
if (C->method() != nullptr) {
1935
tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1936
tty->print_raw(method_metadata_str.freeze());
1937
} else if (C->stub_name() != nullptr) {
1938
tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1941
tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1942
tty->print_raw(dump_asm_str.freeze());
1943
tty->print_cr("--------------------------------------------------------------------------------");
1944
if (xtty != nullptr) {
1945
xtty->tail("opto_assembly");
1952
void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1953
_inc_table.set_size(cnt);
1956
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1957
Block* block = C->cfg()->get_block(i);
1961
// Find the branch; ignore trailing NOPs.
1962
for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1963
n = block->get_node(j);
1964
if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1969
// If we didn't find anything, continue
1974
// Compute ExceptionHandlerTable subtable entry and add it
1975
// (skip empty blocks)
1976
if (n->is_Catch()) {
1978
// Get the offset of the return from the call
1979
uint call_return = call_returns[block->_pre_order];
1981
assert( call_return > 0, "no call seen for this basic block" );
1982
while (block->get_node(--j)->is_MachProj()) ;
1983
assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1985
// last instruction is a CatchNode, find it's CatchProjNodes
1986
int nof_succs = block->_num_succs;
1988
GrowableArray<intptr_t> handler_bcis(nof_succs);
1989
GrowableArray<intptr_t> handler_pcos(nof_succs);
1990
// iterate through all successors
1991
for (int j = 0; j < nof_succs; j++) {
1992
Block* s = block->_succs[j];
1993
bool found_p = false;
1994
for (uint k = 1; k < s->num_preds(); k++) {
1995
Node* pk = s->pred(k);
1996
if (pk->is_CatchProj() && pk->in(0) == n) {
1997
const CatchProjNode* p = pk->as_CatchProj();
1999
// add the corresponding handler bci & pco information
2000
if (p->_con != CatchProjNode::fall_through_index) {
2001
// p leads to an exception handler (and is not fall through)
2002
assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2003
// no duplicates, please
2004
if (!handler_bcis.contains(p->handler_bci())) {
2005
uint block_num = s->non_connector()->_pre_order;
2006
handler_bcis.append(p->handler_bci());
2007
handler_pcos.append(blk_labels[block_num].loc_pos());
2012
assert(found_p, "no matching predecessor found");
2013
// Note: Due to empty block removal, one block may have
2014
// several CatchProj inputs, from the same Catch.
2017
// Set the offset of the return from the call
2018
assert(handler_bcis.find(-1) != -1, "must have default handler");
2019
_handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2023
// Handle implicit null exception table updates
2024
if (n->is_MachNullCheck()) {
2025
uint block_num = block->non_connector_successor(0)->_pre_order;
2026
_inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2029
// Handle implicit exception table updates: trap instructions.
2030
if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2031
uint block_num = block->non_connector_successor(0)->_pre_order;
2032
_inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2035
} // End of for all blocks fill in exception table entries
2040
uint Scheduling::_total_nop_size = 0;
2041
uint Scheduling::_total_method_size = 0;
2042
uint Scheduling::_total_branches = 0;
2043
uint Scheduling::_total_unconditional_delays = 0;
2044
uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2047
// Initializer for class Scheduling
2049
Scheduling::Scheduling(Arena *arena, Compile &compile)
2051
_cfg(compile.cfg()),
2052
_regalloc(compile.regalloc()),
2056
_pinch_free_list(arena),
2057
_next_node(nullptr),
2058
_bundle_instr_count(0),
2059
_bundle_cycle_number(0),
2060
_bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2063
, _unconditional_delays(0)
2066
// Create a MachNopNode
2067
_nop = new MachNopNode();
2069
// Now that the nops are in the array, save the count
2070
// (but allow entries for the nops)
2071
_node_bundling_limit = compile.unique();
2072
uint node_max = _regalloc->node_regs_max_index();
2074
compile.output()->set_node_bundling_limit(_node_bundling_limit);
2076
// This one is persistent within the Compile class
2077
_node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2079
// Allocate space for fixed-size arrays
2080
_uses = NEW_ARENA_ARRAY(arena, short, node_max);
2081
_current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2084
for (uint i = 0; i < node_max; i++) {
2085
::new (&_node_bundling_base[i]) Bundle();
2087
memset(_uses, 0, node_max * sizeof(short));
2088
memset(_current_latency, 0, node_max * sizeof(unsigned short));
2090
// Clear the bundling information
2091
memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2093
// Get the last node
2094
Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2096
_next_node = block->get_node(block->number_of_nodes() - 1);
2100
// Scheduling destructor
2101
Scheduling::~Scheduling() {
2102
_total_branches += _branches;
2103
_total_unconditional_delays += _unconditional_delays;
2107
// Step ahead "i" cycles
2108
void Scheduling::step(uint i) {
2110
Bundle *bundle = node_bundling(_next_node);
2111
bundle->set_starts_bundle();
2113
// Update the bundle record, but leave the flags information alone
2114
if (_bundle_instr_count > 0) {
2115
bundle->set_instr_count(_bundle_instr_count);
2116
bundle->set_resources_used(_bundle_use.resourcesUsed());
2119
// Update the state information
2120
_bundle_instr_count = 0;
2121
_bundle_cycle_number += i;
2122
_bundle_use.step(i);
2125
void Scheduling::step_and_clear() {
2126
Bundle *bundle = node_bundling(_next_node);
2127
bundle->set_starts_bundle();
2129
// Update the bundle record
2130
if (_bundle_instr_count > 0) {
2131
bundle->set_instr_count(_bundle_instr_count);
2132
bundle->set_resources_used(_bundle_use.resourcesUsed());
2134
_bundle_cycle_number += 1;
2137
// Clear the bundling information
2138
_bundle_instr_count = 0;
2139
_bundle_use.reset();
2141
memcpy(_bundle_use_elements,
2142
Pipeline_Use::elaborated_elements,
2143
sizeof(Pipeline_Use::elaborated_elements));
2146
// Perform instruction scheduling and bundling over the sequence of
2147
// instructions in backwards order.
2148
void PhaseOutput::ScheduleAndBundle() {
2150
// Don't optimize this if it isn't a method
2154
// Don't optimize this if scheduling is disabled
2155
if (!C->do_scheduling())
2158
// Scheduling code works only with pairs (8 bytes) maximum.
2159
// And when the scalable vector register is used, we may spill/unspill
2160
// the whole reg regardless of the max vector size.
2161
if (C->max_vector_size() > 8 ||
2162
(C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2166
Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2168
// Create a data structure for all the scheduling information
2169
Scheduling scheduling(Thread::current()->resource_area(), *C);
2171
// Walk backwards over each basic block, computing the needed alignment
2172
// Walk over all the basic blocks
2173
scheduling.DoScheduling();
2176
if (C->trace_opto_output()) {
2177
// Buffer and print all at once
2180
ss.print("\n---- After ScheduleAndBundle ----\n");
2181
print_scheduling(&ss);
2182
tty->print("%s", ss.as_string());
2188
// Separated out so that it can be called directly from debugger
2189
void PhaseOutput::print_scheduling() {
2190
print_scheduling(tty);
2193
void PhaseOutput::print_scheduling(outputStream* output_stream) {
2194
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2195
output_stream->print("\nBB#%03d:\n", i);
2196
Block* block = C->cfg()->get_block(i);
2197
for (uint j = 0; j < block->number_of_nodes(); j++) {
2198
Node* n = block->get_node(j);
2199
OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2200
output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2201
n->dump("\n", false, output_stream);
2207
// See if this node fits into the present instruction bundle
2208
bool Scheduling::NodeFitsInBundle(Node *n) {
2209
uint n_idx = n->_idx;
2211
// If this is the unconditional delay instruction, then it fits
2212
if (n == _unconditional_delay_slot) {
2214
if (_cfg->C->trace_opto_output())
2215
tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2220
// If the node cannot be scheduled this cycle, skip it
2221
if (_current_latency[n_idx] > _bundle_cycle_number) {
2223
if (_cfg->C->trace_opto_output())
2224
tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2225
n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2230
const Pipeline *node_pipeline = n->pipeline();
2232
uint instruction_count = node_pipeline->instructionCount();
2233
if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2234
instruction_count = 0;
2235
else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2236
instruction_count++;
2238
if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2240
if (_cfg->C->trace_opto_output())
2241
tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2242
n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2247
// Don't allow non-machine nodes to be handled this way
2248
if (!n->is_Mach() && instruction_count == 0)
2251
// See if there is any overlap
2252
uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2256
if (_cfg->C->trace_opto_output())
2257
tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2263
if (_cfg->C->trace_opto_output())
2264
tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2270
Node * Scheduling::ChooseNodeToBundle() {
2271
uint siz = _available.size();
2276
if (_cfg->C->trace_opto_output())
2277
tty->print("# ChooseNodeToBundle: null\n");
2282
// Fast path, if only 1 instruction in the bundle
2285
if (_cfg->C->trace_opto_output()) {
2286
tty->print("# ChooseNodeToBundle (only 1): ");
2287
_available[0]->dump();
2290
return (_available[0]);
2293
// Don't bother, if the bundle is already full
2294
if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2295
for ( uint i = 0; i < siz; i++ ) {
2296
Node *n = _available[i];
2298
// Skip projections, we'll handle them another way
2302
// This presupposed that instructions are inserted into the
2303
// available list in a legality order; i.e. instructions that
2304
// must be inserted first are at the head of the list
2305
if (NodeFitsInBundle(n)) {
2307
if (_cfg->C->trace_opto_output()) {
2308
tty->print("# ChooseNodeToBundle: ");
2317
// Nothing fits in this bundle, choose the highest priority
2319
if (_cfg->C->trace_opto_output()) {
2320
tty->print("# ChooseNodeToBundle: ");
2321
_available[0]->dump();
2325
return _available[0];
2328
int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2329
assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2331
OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2332
OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2333
OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2334
OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2336
// Comparison between stack -> reg and stack -> reg
2337
if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2338
OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2339
return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2342
// Comparison between reg -> stack and reg -> stack
2343
if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2344
OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2345
return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2348
return 0; // Not comparable
2351
void Scheduling::AddNodeToAvailableList(Node *n) {
2352
assert( !n->is_Proj(), "projections never directly made available" );
2354
if (_cfg->C->trace_opto_output()) {
2355
tty->print("# AddNodeToAvailableList: ");
2360
int latency = _current_latency[n->_idx];
2362
// Insert in latency order (insertion sort). If two MachSpillCopyNodes
2363
// for stack spilling or unspilling have the same latency, we sort
2364
// them in the order of stack offset. Some ports (e.g. aarch64) may also
2365
// have more opportunities to do ld/st merging
2367
for (i = 0; i < _available.size(); i++) {
2368
if (_current_latency[_available[i]->_idx] > latency) {
2370
} else if (_current_latency[_available[i]->_idx] == latency &&
2371
n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2372
compare_two_spill_nodes(n, _available[i]) > 0) {
2377
// Special Check for compares following branches
2378
if( n->is_Mach() && _scheduled.size() > 0 ) {
2379
int op = n->as_Mach()->ideal_Opcode();
2380
Node *last = _scheduled[0];
2381
if( last->is_MachIf() && last->in(1) == n &&
2390
// Recalculate position, moving to front of same latency
2391
for ( i=0 ; i < _available.size(); i++ )
2392
if (_current_latency[_available[i]->_idx] >= latency)
2397
// Insert the node in the available list
2398
_available.insert(i, n);
2401
if (_cfg->C->trace_opto_output())
2406
void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2407
for ( uint i=0; i < n->len(); i++ ) {
2408
Node *def = n->in(i);
2410
if( def->is_Proj() ) // If this is a machine projection, then
2411
def = def->in(0); // propagate usage thru to the base instruction
2413
if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2417
// Compute the latency
2418
uint l = _bundle_cycle_number + n->latency(i);
2419
if (_current_latency[def->_idx] < l)
2420
_current_latency[def->_idx] = l;
2422
// If this does not have uses then schedule it
2423
if ((--_uses[def->_idx]) == 0)
2424
AddNodeToAvailableList(def);
2428
void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2430
if (_cfg->C->trace_opto_output()) {
2431
tty->print("# AddNodeToBundle: ");
2436
// Remove this from the available list
2438
for (i = 0; i < _available.size(); i++)
2439
if (_available[i] == n)
2441
assert(i < _available.size(), "entry in _available list not found");
2442
_available.remove(i);
2444
// See if this fits in the current bundle
2445
const Pipeline *node_pipeline = n->pipeline();
2446
const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2448
// Check for instructions to be placed in the delay slot. We
2449
// do this before we actually schedule the current instruction,
2450
// because the delay slot follows the current instruction.
2451
if (Pipeline::_branch_has_delay_slot &&
2452
node_pipeline->hasBranchDelay() &&
2453
!_unconditional_delay_slot) {
2455
uint siz = _available.size();
2457
// Conditional branches can support an instruction that
2458
// is unconditionally executed and not dependent by the
2459
// branch, OR a conditionally executed instruction if
2460
// the branch is taken. In practice, this means that
2461
// the first instruction at the branch target is
2462
// copied to the delay slot, and the branch goes to
2463
// the instruction after that at the branch target
2464
if ( n->is_MachBranch() ) {
2466
assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2467
assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2473
// At least 1 instruction is on the available list
2474
// that is not dependent on the branch
2475
for (uint i = 0; i < siz; i++) {
2476
Node *d = _available[i];
2477
const Pipeline *avail_pipeline = d->pipeline();
2479
// Don't allow safepoints in the branch shadow, that will
2480
// cause a number of difficulties
2481
if ( avail_pipeline->instructionCount() == 1 &&
2482
!avail_pipeline->hasMultipleBundles() &&
2483
!avail_pipeline->hasBranchDelay() &&
2484
Pipeline::instr_has_unit_size() &&
2485
d->size(_regalloc) == Pipeline::instr_unit_size() &&
2486
NodeFitsInBundle(d) &&
2487
!node_bundling(d)->used_in_delay()) {
2489
if (d->is_Mach() && !d->is_MachSafePoint()) {
2490
// A node that fits in the delay slot was found, so we need to
2491
// set the appropriate bits in the bundle pipeline information so
2492
// that it correctly indicates resource usage. Later, when we
2493
// attempt to add this instruction to the bundle, we will skip
2494
// setting the resource usage.
2495
_unconditional_delay_slot = d;
2496
node_bundling(n)->set_use_unconditional_delay();
2497
node_bundling(d)->set_used_in_unconditional_delay();
2498
_bundle_use.add_usage(avail_pipeline->resourceUse());
2499
_current_latency[d->_idx] = _bundle_cycle_number;
2501
++_bundle_instr_count;
2503
_unconditional_delays++;
2511
// No delay slot, add a nop to the usage
2512
if (!_unconditional_delay_slot) {
2513
// See if adding an instruction in the delay slot will overflow
2515
if (!NodeFitsInBundle(_nop)) {
2517
if (_cfg->C->trace_opto_output())
2518
tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2523
_bundle_use.add_usage(_nop->pipeline()->resourceUse());
2525
++_bundle_instr_count;
2528
// See if the instruction in the delay slot requires a
2529
// step of the bundles
2530
if (!NodeFitsInBundle(n)) {
2532
if (_cfg->C->trace_opto_output())
2533
tty->print("# *** STEP(branch won't fit) ***\n");
2535
// Update the state information
2536
_bundle_instr_count = 0;
2537
_bundle_cycle_number += 1;
2538
_bundle_use.step(1);
2542
// Get the number of instructions
2543
uint instruction_count = node_pipeline->instructionCount();
2544
if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2545
instruction_count = 0;
2547
// Compute the latency information
2550
if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2551
int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2552
if (relative_latency < 0)
2553
relative_latency = 0;
2555
delay = _bundle_use.full_latency(relative_latency, node_usage);
2557
// Does not fit in this bundle, start a new one
2562
if (_cfg->C->trace_opto_output())
2563
tty->print("# *** STEP(%d) ***\n", delay);
2568
// If this was placed in the delay slot, ignore it
2569
if (n != _unconditional_delay_slot) {
2572
if (node_pipeline->hasMultipleBundles()) {
2574
if (_cfg->C->trace_opto_output())
2575
tty->print("# *** STEP(multiple instructions) ***\n");
2580
else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2582
if (_cfg->C->trace_opto_output())
2583
tty->print("# *** STEP(%d >= %d instructions) ***\n",
2584
instruction_count + _bundle_instr_count,
2585
Pipeline::_max_instrs_per_cycle);
2591
if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2592
_bundle_instr_count++;
2594
// Set the node's latency
2595
_current_latency[n->_idx] = _bundle_cycle_number;
2597
// Now merge the functional unit information
2598
if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2599
_bundle_use.add_usage(node_usage);
2601
// Increment the number of instructions in this bundle
2602
_bundle_instr_count += instruction_count;
2604
// Remember this node for later
2609
// It's possible to have a BoxLock in the graph and in the _bbs mapping but
2610
// not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2611
// 'Schedule' them (basically ignore in the schedule) but do not insert them
2612
// into the block. All other scheduled nodes get put in the schedule here.
2613
int op = n->Opcode();
2614
if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2615
(op != Op_Node && // Not an unused antidepedence node and
2616
// not an unallocated boxlock
2617
(OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2619
// Push any trailing projections
2620
if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2621
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2622
Node *foi = n->fast_out(i);
2623
if( foi->is_Proj() )
2624
_scheduled.push(foi);
2628
// Put the instruction in the schedule list
2633
if (_cfg->C->trace_opto_output())
2637
// Walk all the definitions, decrementing use counts, and
2638
// if a definition has a 0 use count, place it in the available list.
2639
DecrementUseCounts(n,bb);
2642
// This method sets the use count within a basic block. We will ignore all
2643
// uses outside the current basic block. As we are doing a backwards walk,
2644
// any node we reach that has a use count of 0 may be scheduled. This also
2645
// avoids the problem of cyclic references from phi nodes, as long as phi
2646
// nodes are at the front of the basic block. This method also initializes
2647
// the available list to the set of instructions that have no uses within this
2649
void Scheduling::ComputeUseCount(const Block *bb) {
2651
if (_cfg->C->trace_opto_output())
2652
tty->print("# -> ComputeUseCount\n");
2655
// Clear the list of available and scheduled instructions, just in case
2659
// No delay slot specified
2660
_unconditional_delay_slot = nullptr;
2663
for( uint i=0; i < bb->number_of_nodes(); i++ )
2664
assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2667
// Force the _uses count to never go to zero for unscheduable pieces
2669
for( uint k = 0; k < _bb_start; k++ )
2670
_uses[bb->get_node(k)->_idx] = 1;
2671
for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2672
_uses[bb->get_node(l)->_idx] = 1;
2674
// Iterate backwards over the instructions in the block. Don't count the
2675
// branch projections at end or the block header instructions.
2676
for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2677
Node *n = bb->get_node(j);
2678
if( n->is_Proj() ) continue; // Projections handled another way
2680
// Account for all uses
2681
for ( uint k = 0; k < n->len(); k++ ) {
2682
Node *inp = n->in(k);
2684
assert(inp != n, "no cycles allowed" );
2685
if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2686
if (inp->is_Proj()) { // Skip through Proj's
2689
++_uses[inp->_idx]; // Count 1 block-local use
2693
// If this instruction has a 0 use count, then it is available
2694
if (!_uses[n->_idx]) {
2695
_current_latency[n->_idx] = _bundle_cycle_number;
2696
AddNodeToAvailableList(n);
2700
if (_cfg->C->trace_opto_output()) {
2701
tty->print("# uses: %3d: ", _uses[n->_idx]);
2708
if (_cfg->C->trace_opto_output())
2709
tty->print("# <- ComputeUseCount\n");
2713
// This routine performs scheduling on each basic block in reverse order,
2714
// using instruction latencies and taking into account function unit
2716
void Scheduling::DoScheduling() {
2718
if (_cfg->C->trace_opto_output())
2719
tty->print("# -> DoScheduling\n");
2722
Block *succ_bb = nullptr;
2724
Compile* C = Compile::current();
2726
// Walk over all the basic blocks in reverse order
2727
for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2728
bb = _cfg->get_block(i);
2731
if (_cfg->C->trace_opto_output()) {
2732
tty->print("# Schedule BB#%03d (initial)\n", i);
2733
for (uint j = 0; j < bb->number_of_nodes(); j++) {
2734
bb->get_node(j)->dump();
2739
// On the head node, skip processing
2740
if (bb == _cfg->get_root_block()) {
2744
// Skip empty, connector blocks
2745
if (bb->is_connector())
2748
// If the following block is not the sole successor of
2749
// this one, then reset the pipeline information
2750
if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2752
if (_cfg->C->trace_opto_output()) {
2753
tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2754
_next_node->_idx, _bundle_instr_count);
2760
// Leave untouched the starting instruction, any Phis, a CreateEx node
2761
// or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2762
_bb_end = bb->number_of_nodes()-1;
2763
for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2764
Node *n = bb->get_node(_bb_start);
2765
// Things not matched, like Phinodes and ProjNodes don't get scheduled.
2766
// Also, MachIdealNodes do not get scheduled
2767
if( !n->is_Mach() ) continue; // Skip non-machine nodes
2768
MachNode *mach = n->as_Mach();
2769
int iop = mach->ideal_Opcode();
2770
if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2771
if( iop == Op_Con ) continue; // Do not schedule Top
2772
if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2773
mach->pipeline() == MachNode::pipeline_class() &&
2774
!n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2776
break; // Funny loop structure to be sure...
2778
// Compute last "interesting" instruction in block - last instruction we
2779
// might schedule. _bb_end points just after last schedulable inst. We
2780
// normally schedule conditional branches (despite them being forced last
2781
// in the block), because they have delay slots we can fill. Calls all
2782
// have their delay slots filled in the template expansions, so we don't
2783
// bother scheduling them.
2784
Node *last = bb->get_node(_bb_end);
2785
// Ignore trailing NOPs.
2786
while (_bb_end > 0 && last->is_Mach() &&
2787
last->as_Mach()->ideal_Opcode() == Op_Con) {
2788
last = bb->get_node(--_bb_end);
2790
assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2791
if( last->is_Catch() ||
2792
(last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2793
// There might be a prior call. Skip it.
2794
while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2795
} else if( last->is_MachNullCheck() ) {
2796
// Backup so the last null-checked memory instruction is
2797
// outside the schedulable range. Skip over the nullcheck,
2798
// projection, and the memory nodes.
2799
Node *mem = last->in(1);
2802
} while (mem != bb->get_node(_bb_end));
2804
// Set _bb_end to point after last schedulable inst.
2808
assert( _bb_start <= _bb_end, "inverted block ends" );
2810
// Compute the register antidependencies for the basic block
2811
ComputeRegisterAntidependencies(bb);
2812
if (C->failing()) return; // too many D-U pinch points
2814
// Compute the usage within the block, and set the list of all nodes
2815
// in the block that have no uses within the block.
2816
ComputeUseCount(bb);
2818
// Schedule the remaining instructions in the block
2819
while ( _available.size() > 0 ) {
2820
Node *n = ChooseNodeToBundle();
2821
guarantee(n != nullptr, "no nodes available");
2822
AddNodeToBundle(n,bb);
2825
assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2827
for( uint l = _bb_start; l < _bb_end; l++ ) {
2828
Node *n = bb->get_node(l);
2830
for( m = 0; m < _bb_end-_bb_start; m++ )
2831
if( _scheduled[m] == n )
2833
assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2837
// Now copy the instructions (in reverse order) back to the block
2838
for ( uint k = _bb_start; k < _bb_end; k++ )
2839
bb->map_node(_scheduled[_bb_end-k-1], k);
2842
if (_cfg->C->trace_opto_output()) {
2843
tty->print("# Schedule BB#%03d (final)\n", i);
2845
for (uint j = 0; j < bb->number_of_nodes(); j++) {
2846
Node *n = bb->get_node(j);
2847
if( valid_bundle_info(n) ) {
2848
Bundle *bundle = node_bundling(n);
2849
if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2850
tty->print("*** Bundle: ");
2859
verify_good_schedule(bb,"after block local scheduling");
2864
if (_cfg->C->trace_opto_output())
2865
tty->print("# <- DoScheduling\n");
2868
// Record final node-bundling array location
2869
_regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2871
} // end DoScheduling
2873
// Verify that no live-range used in the block is killed in the block by a
2874
// wrong DEF. This doesn't verify live-ranges that span blocks.
2876
// Check for edge existence. Used to avoid adding redundant precedence edges.
2877
static bool edge_from_to( Node *from, Node *to ) {
2878
for( uint i=0; i<from->len(); i++ )
2879
if( from->in(i) == to )
2885
void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2886
// Check for bad kills
2887
if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2888
Node *prior_use = _reg_node[def];
2889
if( prior_use && !edge_from_to(prior_use,n) ) {
2890
tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2892
tty->print_cr("...");
2894
assert(edge_from_to(prior_use,n), "%s", msg);
2896
_reg_node.map(def,nullptr); // Kill live USEs
2900
void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2902
// Zap to something reasonable for the verify code
2905
// Walk over the block backwards. Check to make sure each DEF doesn't
2906
// kill a live value (other than the one it's supposed to). Add each
2907
// USE to the live set.
2908
for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2909
Node *n = b->get_node(i);
2910
int n_op = n->Opcode();
2911
if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2912
// Fat-proj kills a slew of registers
2913
RegMaskIterator rmi(n->out_RegMask());
2914
while (rmi.has_next()) {
2915
OptoReg::Name kill = rmi.next();
2916
verify_do_def(n, kill, msg);
2918
} else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2919
// Get DEF'd registers the normal way
2920
verify_do_def( n, _regalloc->get_reg_first(n), msg );
2921
verify_do_def( n, _regalloc->get_reg_second(n), msg );
2924
// Now make all USEs live
2925
for( uint i=1; i<n->req(); i++ ) {
2926
Node *def = n->in(i);
2927
assert(def != nullptr, "input edge required");
2928
OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2929
OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2930
if( OptoReg::is_valid(reg_lo) ) {
2931
assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2932
_reg_node.map(reg_lo,n);
2934
if( OptoReg::is_valid(reg_hi) ) {
2935
assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2936
_reg_node.map(reg_hi,n);
2942
// Zap to something reasonable for the Antidependence code
2947
// Conditionally add precedence edges. Avoid putting edges on Projs.
2948
static void add_prec_edge_from_to( Node *from, Node *to ) {
2949
if( from->is_Proj() ) { // Put precedence edge on Proj's input
2950
assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2953
if( from != to && // No cycles (for things like LD L0,[L0+4] )
2954
!edge_from_to( from, to ) ) // Avoid duplicate edge
2958
void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2959
if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2962
if (OptoReg::is_reg(def_reg)) {
2963
VMReg vmreg = OptoReg::as_VMReg(def_reg);
2964
if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2965
// This is one of the high slots of a vector register.
2966
// ScheduleAndBundle already checked there are no live wide
2967
// vectors in this method so it can be safely ignored.
2972
Node *pinch = _reg_node[def_reg]; // Get pinch point
2973
if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2974
is_def ) { // Check for a true def (not a kill)
2975
_reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2979
Node *kill = def; // Rename 'def' to more descriptive 'kill'
2980
debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2982
// After some number of kills there _may_ be a later def
2983
Node *later_def = nullptr;
2985
Compile* C = Compile::current();
2987
// Finding a kill requires a real pinch-point.
2988
// Check for not already having a pinch-point.
2989
// Pinch points are Op_Node's.
2990
if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2991
later_def = pinch; // Must be def/kill as optimistic pinch-point
2992
if ( _pinch_free_list.size() > 0) {
2993
pinch = _pinch_free_list.pop();
2995
pinch = new Node(1); // Pinch point to-be
2997
if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2998
DEBUG_ONLY( pinch->dump(); );
2999
assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
3000
_cfg->C->record_method_not_compilable("too many D-U pinch points");
3003
_cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3004
_reg_node.map(def_reg,pinch); // Record pinch-point
3005
//regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3006
if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3007
pinch->init_req(0, C->top()); // set not null for the next call
3008
add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3009
later_def = nullptr; // and no later def
3011
pinch->set_req(0,later_def); // Hook later def so we can find it
3012
} else { // Else have valid pinch point
3013
if( pinch->in(0) ) // If there is a later-def
3014
later_def = pinch->in(0); // Get it
3017
// Add output-dependence edge from later def to kill
3018
if( later_def ) // If there is some original def
3019
add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3021
// See if current kill is also a use, and so is forced to be the pinch-point.
3022
if( pinch->Opcode() == Op_Node ) {
3023
Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3024
for( uint i=1; i<uses->req(); i++ ) {
3025
if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3026
_regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3027
// Yes, found a use/kill pinch-point
3028
pinch->set_req(0,nullptr); //
3029
pinch->replace_by(kill); // Move anti-dep edges up
3031
_reg_node.map(def_reg,pinch);
3037
// Add edge from kill to pinch-point
3038
add_prec_edge_from_to(kill,pinch);
3041
void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3042
if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3044
Node *pinch = _reg_node[use_reg]; // Get pinch point
3045
// Check for no later def_reg/kill in block
3046
if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3047
// Use has to be block-local as well
3048
_cfg->get_block_for_node(use) == b) {
3049
if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3050
pinch->req() == 1 ) { // pinch not yet in block?
3051
pinch->del_req(0); // yank pointer to later-def, also set flag
3052
// Insert the pinch-point in the block just after the last use
3053
b->insert_node(pinch, b->find_node(use) + 1);
3054
_bb_end++; // Increase size scheduled region in block
3057
add_prec_edge_from_to(pinch,use);
3061
// We insert antidependences between the reads and following write of
3062
// allocated registers to prevent illegal code motion. Hopefully, the
3063
// number of added references should be fairly small, especially as we
3064
// are only adding references within the current basic block.
3065
void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3068
verify_good_schedule(b,"before block local scheduling");
3071
// A valid schedule, for each register independently, is an endless cycle
3072
// of: a def, then some uses (connected to the def by true dependencies),
3073
// then some kills (defs with no uses), finally the cycle repeats with a new
3074
// def. The uses are allowed to float relative to each other, as are the
3075
// kills. No use is allowed to slide past a kill (or def). This requires
3076
// antidependencies between all uses of a single def and all kills that
3077
// follow, up to the next def. More edges are redundant, because later defs
3078
// & kills are already serialized with true or antidependencies. To keep
3079
// the edge count down, we add a 'pinch point' node if there's more than
3080
// one use or more than one kill/def.
3082
// We add dependencies in one bottom-up pass.
3084
// For each instruction we handle it's DEFs/KILLs, then it's USEs.
3086
// For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3087
// register. If not, we record the DEF/KILL in _reg_node, the
3088
// register-to-def mapping. If there is a prior DEF/KILL, we insert a
3089
// "pinch point", a new Node that's in the graph but not in the block.
3090
// We put edges from the prior and current DEF/KILLs to the pinch point.
3091
// We put the pinch point in _reg_node. If there's already a pinch point
3092
// we merely add an edge from the current DEF/KILL to the pinch point.
3094
// After doing the DEF/KILLs, we handle USEs. For each used register, we
3095
// put an edge from the pinch point to the USE.
3097
// To be expedient, the _reg_node array is pre-allocated for the whole
3098
// compilation. _reg_node is lazily initialized; it either contains a null,
3099
// or a valid def/kill/pinch-point, or a leftover node from some prior
3100
// block. Leftover node from some prior block is treated like a null (no
3101
// prior def, so no anti-dependence needed). Valid def is distinguished by
3102
// it being in the current block.
3103
bool fat_proj_seen = false;
3104
uint last_safept = _bb_end-1;
3105
Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3106
Node* last_safept_node = end_node;
3107
for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3108
Node *n = b->get_node(i);
3109
int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3110
if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3111
// Fat-proj kills a slew of registers
3112
// This can add edges to 'n' and obscure whether or not it was a def,
3113
// hence the is_def flag.
3114
fat_proj_seen = true;
3115
RegMaskIterator rmi(n->out_RegMask());
3116
while (rmi.has_next()) {
3117
OptoReg::Name kill = rmi.next();
3118
anti_do_def(b, n, kill, is_def);
3121
// Get DEF'd registers the normal way
3122
anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3123
anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3126
// Kill projections on a branch should appear to occur on the
3127
// branch, not afterwards, so grab the masks from the projections
3128
// and process them.
3129
if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3130
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3131
Node* use = n->fast_out(i);
3132
if (use->is_Proj()) {
3133
RegMaskIterator rmi(use->out_RegMask());
3134
while (rmi.has_next()) {
3135
OptoReg::Name kill = rmi.next();
3136
anti_do_def(b, n, kill, false);
3142
// Check each register used by this instruction for a following DEF/KILL
3143
// that must occur afterward and requires an anti-dependence edge.
3144
for( uint j=0; j<n->req(); j++ ) {
3145
Node *def = n->in(j);
3147
assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3148
anti_do_use( b, n, _regalloc->get_reg_first(def) );
3149
anti_do_use( b, n, _regalloc->get_reg_second(def) );
3152
// Do not allow defs of new derived values to float above GC
3153
// points unless the base is definitely available at the GC point.
3155
Node *m = b->get_node(i);
3157
// Add precedence edge from following safepoint to use of derived pointer
3158
if( last_safept_node != end_node &&
3159
m != last_safept_node) {
3160
for (uint k = 1; k < m->req(); k++) {
3161
const Type *t = m->in(k)->bottom_type();
3162
if( t->isa_oop_ptr() &&
3163
t->is_ptr()->offset() != 0 ) {
3164
last_safept_node->add_prec( m );
3170
if( n->jvms() ) { // Precedence edge from derived to safept
3171
// Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3172
if( b->get_node(last_safept) != last_safept_node ) {
3173
last_safept = b->find_node(last_safept_node);
3175
for( uint j=last_safept; j > i; j-- ) {
3176
Node *mach = b->get_node(j);
3177
if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3178
mach->add_prec( n );
3181
last_safept_node = m;
3185
if (fat_proj_seen) {
3186
// Garbage collect pinch nodes that were not consumed.
3187
// They are usually created by a fat kill MachProj for a call.
3188
garbage_collect_pinch_nodes();
3192
// Garbage collect pinch nodes for reuse by other blocks.
3194
// The block scheduler's insertion of anti-dependence
3195
// edges creates many pinch nodes when the block contains
3196
// 2 or more Calls. A pinch node is used to prevent a
3197
// combinatorial explosion of edges. If a set of kills for a
3198
// register is anti-dependent on a set of uses (or defs), rather
3199
// than adding an edge in the graph between each pair of kill
3200
// and use (or def), a pinch is inserted between them:
3210
// One pinch node is created per register killed when
3211
// the second call is encountered during a backwards pass
3212
// over the block. Most of these pinch nodes are never
3213
// wired into the graph because the register is never
3214
// used or def'ed in the block.
3216
void Scheduling::garbage_collect_pinch_nodes() {
3218
if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3221
for (uint k = 0; k < _reg_node.max(); k++) {
3222
Node* pinch = _reg_node[k];
3223
if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3224
// no predecence input edges
3225
(pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3226
cleanup_pinch(pinch);
3227
_pinch_free_list.push(pinch);
3228
_reg_node.map(k, nullptr);
3230
if (_cfg->C->trace_opto_output()) {
3232
if (trace_cnt > 40) {
3236
tty->print(" %d", pinch->_idx);
3242
if (_cfg->C->trace_opto_output()) tty->print("\n");
3246
// Clean up a pinch node for reuse.
3247
void Scheduling::cleanup_pinch( Node *pinch ) {
3248
assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3250
for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3251
Node* use = pinch->last_out(i);
3252
uint uses_found = 0;
3253
for (uint j = use->req(); j < use->len(); j++) {
3254
if (use->in(j) == pinch) {
3259
assert(uses_found > 0, "must be a precedence edge");
3260
i -= uses_found; // we deleted 1 or more copies of this edge
3262
// May have a later_def entry
3263
pinch->set_req(0, nullptr);
3268
void Scheduling::dump_available() const {
3269
tty->print("#Availist ");
3270
for (uint i = 0; i < _available.size(); i++)
3271
tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3275
// Print Scheduling Statistics
3276
void Scheduling::print_statistics() {
3277
// Print the size added by nops for bundling
3278
tty->print("Nops added %d bytes to total of %d bytes",
3279
_total_nop_size, _total_method_size);
3280
if (_total_method_size > 0)
3281
tty->print(", for %.2f%%",
3282
((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3285
// Print the number of branch shadows filled
3286
if (Pipeline::_branch_has_delay_slot) {
3287
tty->print("Of %d branches, %d had unconditional delay slots filled",
3288
_total_branches, _total_unconditional_delays);
3289
if (_total_branches > 0)
3290
tty->print(", for %.2f%%",
3291
((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3295
uint total_instructions = 0, total_bundles = 0;
3297
for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3298
uint bundle_count = _total_instructions_per_bundle[i];
3299
total_instructions += bundle_count * i;
3300
total_bundles += bundle_count;
3303
if (total_bundles > 0)
3304
tty->print("Average ILP (excluding nops) is %.2f\n",
3305
((double)total_instructions) / ((double)total_bundles));
3309
//-----------------------init_scratch_buffer_blob------------------------------
3310
// Construct a temporary BufferBlob and cache it for this compile.
3311
void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3312
// If there is already a scratch buffer blob allocated and the
3313
// constant section is big enough, use it. Otherwise free the
3314
// current and allocate a new one.
3315
BufferBlob* blob = scratch_buffer_blob();
3316
if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3317
// Use the current blob.
3319
if (blob != nullptr) {
3320
BufferBlob::free(blob);
3324
_scratch_const_size = const_size;
3325
int size = C2Compiler::initial_code_buffer_size(const_size);
3326
blob = BufferBlob::create("Compile::scratch_buffer", size);
3327
// Record the buffer blob for next time.
3328
set_scratch_buffer_blob(blob);
3329
// Have we run out of code space?
3330
if (scratch_buffer_blob() == nullptr) {
3331
// Let CompilerBroker disable further compilations.
3332
C->record_failure("Not enough space for scratch buffer in CodeCache");
3337
// Initialize the relocation buffers
3338
relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3339
set_scratch_locs_memory(locs_buf);
3343
//-----------------------scratch_emit_size-------------------------------------
3344
// Helper function that computes size by emitting code
3345
uint PhaseOutput::scratch_emit_size(const Node* n) {
3346
// Start scratch_emit_size section.
3347
set_in_scratch_emit_size(true);
3349
// Emit into a trash buffer and count bytes emitted.
3350
// This is a pretty expensive way to compute a size,
3351
// but it works well enough if seldom used.
3352
// All common fixed-size instructions are given a size
3353
// method by the AD file.
3354
// Note that the scratch buffer blob and locs memory are
3355
// allocated at the beginning of the compile task, and
3356
// may be shared by several calls to scratch_emit_size.
3357
// The allocation of the scratch buffer blob is particularly
3358
// expensive, since it has to grab the code cache lock.
3359
BufferBlob* blob = this->scratch_buffer_blob();
3360
assert(blob != nullptr, "Initialize BufferBlob at start");
3361
assert(blob->size() > MAX_inst_size, "sanity");
3362
relocInfo* locs_buf = scratch_locs_memory();
3363
address blob_begin = blob->content_begin();
3364
address blob_end = (address)locs_buf;
3365
assert(blob->contains(blob_end), "sanity");
3366
CodeBuffer buf(blob_begin, blob_end - blob_begin);
3367
buf.initialize_consts_size(_scratch_const_size);
3368
buf.initialize_stubs_size(MAX_stubs_size);
3369
assert(locs_buf != nullptr, "sanity");
3370
int lsize = MAX_locs_size / 3;
3371
buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3372
buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3373
buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3374
// Mark as scratch buffer.
3375
buf.consts()->set_scratch_emit();
3376
buf.insts()->set_scratch_emit();
3377
buf.stubs()->set_scratch_emit();
3381
Label fakeL; // Fake label for branch instructions.
3382
Label* saveL = nullptr;
3384
bool is_branch = n->is_MachBranch();
3385
C2_MacroAssembler masm(&buf);
3388
n->as_MachBranch()->save_label(&saveL, &save_bnum);
3389
n->as_MachBranch()->label_set(&fakeL, 0);
3391
n->emit(&masm, C->regalloc());
3393
// Emitting into the scratch buffer should not fail
3394
assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3396
if (is_branch) // Restore label.
3397
n->as_MachBranch()->label_set(saveL, save_bnum);
3399
// End scratch_emit_size section.
3400
set_in_scratch_emit_size(false);
3402
return buf.insts_size();
3405
void PhaseOutput::install() {
3406
if (!C->should_install_code()) {
3408
} else if (C->stub_function() != nullptr) {
3409
install_stub(C->stub_name());
3411
install_code(C->method(),
3413
CompileBroker::compiler2(),
3414
C->has_unsafe_access(),
3415
SharedRuntime::is_wide_vector(C->max_vector_size()));
3419
void PhaseOutput::install_code(ciMethod* target,
3421
AbstractCompiler* compiler,
3422
bool has_unsafe_access,
3423
bool has_wide_vectors) {
3424
// Check if we want to skip execution of all compiled code.
3427
if (OptoNoExecute) {
3428
C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3432
Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3434
if (C->is_osr_compilation()) {
3435
_code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3436
_code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3438
if (!target->is_static()) {
3439
// The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3440
// before the inline cache check, so we don't have to execute any nop instructions when dispatching
3441
// through the UEP, yet we can ensure that the VEP is aligned appropriately.
3442
_code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3444
_code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3445
_code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3448
C->env()->register_method(target,
3451
_orig_pc_slot_offset_in_bytes,
3453
frame_size_in_words(),
3459
SharedRuntime::is_wide_vector(C->max_vector_size()),
3461
C->has_scoped_access(),
3464
if (C->log() != nullptr) { // Print code cache state into compiler log
3465
C->log()->code_cache_state();
3469
void PhaseOutput::install_stub(const char* stub_name) {
3470
// Entry point will be accessed using stub_entry_point();
3471
if (code_buffer() == nullptr) {
3472
Matcher::soft_match_failure();
3474
if (PrintAssembly && (WizardMode || Verbose))
3475
tty->print_cr("### Stub::%s", stub_name);
3477
if (!C->failing()) {
3478
assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3481
// For now we mark the frame as never safe for profile stackwalking
3482
RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3484
CodeOffsets::frame_never_safe,
3485
// _code_offsets.value(CodeOffsets::Frame_Complete),
3486
frame_size_in_words(),
3489
assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3491
C->set_stub_entry_point(rs->entry_point());
3496
// Support for bundling info
3497
Bundle* PhaseOutput::node_bundling(const Node *n) {
3498
assert(valid_bundle_info(n), "oob");
3499
return &_node_bundling_base[n->_idx];
3502
bool PhaseOutput::valid_bundle_info(const Node *n) {
3503
return (_node_bundling_limit > n->_idx);
3506
//------------------------------frame_size_in_words-----------------------------
3507
// frame_slots in units of words
3508
int PhaseOutput::frame_size_in_words() const {
3509
// shift is 0 in LP32 and 1 in LP64
3510
const int shift = (LogBytesPerWord - LogBytesPerInt);
3511
int words = _frame_slots >> shift;
3512
assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3516
// To bang the stack of this compiled method we use the stack size
3517
// that the interpreter would need in case of a deoptimization. This
3518
// removes the need to bang the stack in the deoptimization blob which
3519
// in turn simplifies stack overflow handling.
3520
int PhaseOutput::bang_size_in_bytes() const {
3521
return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3524
//------------------------------dump_asm---------------------------------------
3525
// Dump formatted assembly
3526
#if defined(SUPPORT_OPTO_ASSEMBLY)
3527
void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3529
int pc_digits = 3; // #chars required for pc
3530
int sb_chars = 3; // #chars for "start bundle" indicator
3532
if (pcs != nullptr) {
3534
for (uint i = 0; i < pc_limit; i++) {
3535
max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3537
pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3539
int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3541
bool cut_short = false;
3543
st->print("# "); C->tf()->dump_on(st); st->cr();
3547
int pc = 0x0; // Program counter
3548
char starts_bundle = ' ';
3549
C->regalloc()->dump_frame();
3552
for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3553
if (VMThread::should_terminate()) {
3557
Block* block = C->cfg()->get_block(i);
3558
if (block->is_connector() && !Verbose) {
3562
if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3564
st->print("%*.*x", pc_digits, pc_digits, pc);
3566
st->fill_to(prefix_len);
3567
block->dump_head(C->cfg(), st);
3568
if (block->is_connector()) {
3569
st->fill_to(prefix_len);
3570
st->print_cr("# Empty connector block");
3571
} else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3572
st->fill_to(prefix_len);
3573
st->print_cr("# Block is sole successor of call");
3576
// For all instructions
3577
Node *delay = nullptr;
3578
for (uint j = 0; j < block->number_of_nodes(); j++) {
3579
if (VMThread::should_terminate()) {
3583
n = block->get_node(j);
3584
if (valid_bundle_info(n)) {
3585
Bundle* bundle = node_bundling(n);
3586
if (bundle->used_in_unconditional_delay()) {
3590
if (bundle->starts_bundle()) {
3591
starts_bundle = '+';
3599
if( !n->is_Region() && // Dont print in the Assembly
3600
!n->is_Phi() && // a few noisely useless nodes
3602
!n->is_MachTemp() &&
3603
!n->is_SafePointScalarObject() &&
3604
!n->is_Catch() && // Would be nice to print exception table targets
3605
!n->is_MergeMem() && // Not very interesting
3606
!n->is_top() && // Debug info table constants
3607
!(n->is_Con() && !n->is_Mach())// Debug info table constants
3609
if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3611
st->print("%*.*x", pc_digits, pc_digits, pc);
3613
st->fill_to(pc_digits);
3615
st->print(" %c ", starts_bundle);
3616
starts_bundle = ' ';
3617
st->fill_to(prefix_len);
3618
n->format(C->regalloc(), st);
3622
// If we have an instruction with a delay slot, and have seen a delay,
3623
// then back up and print it
3624
if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3625
// Coverity finding - Explicit null dereferenced.
3626
guarantee(delay != nullptr, "no unconditional delay instruction");
3627
if (WizardMode) delay->dump();
3629
if (node_bundling(delay)->starts_bundle())
3630
starts_bundle = '+';
3631
if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3633
st->print("%*.*x", pc_digits, pc_digits, pc);
3635
st->fill_to(pc_digits);
3637
st->print(" %c ", starts_bundle);
3638
starts_bundle = ' ';
3639
st->fill_to(prefix_len);
3640
delay->format(C->regalloc(), st);
3645
// Dump the exception table as well
3646
if( n->is_Catch() && (Verbose || WizardMode) ) {
3647
// Print the exception table for this offset
3648
_handler_table.print_subtable_for(pc);
3650
st->bol(); // Make sure we start on a new line
3652
st->cr(); // one empty line between blocks
3653
assert(cut_short || delay == nullptr, "no unconditional delay branch");
3654
} // End of per-block dump
3656
if (cut_short) st->print_cr("*** disassembly is cut short ***");
3661
void PhaseOutput::print_statistics() {
3662
Scheduling::print_statistics();