1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
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.
8 *
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).
14 *
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.
18 *
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
21 * questions.
22 *
23 */
24
25 #include "asm/assembler.inline.hpp"
26 #include "code/aotCodeCache.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compilerDirectives.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "gc/shared/barrierSet.hpp"
35 #include "gc/shared/c2/barrierSetC2.hpp"
36 #include "gc/shared/gc_globals.hpp"
37 #include "memory/allocation.hpp"
38 #include "opto/ad.hpp"
39 #include "opto/block.hpp"
40 #include "opto/c2_MacroAssembler.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callnode.hpp"
43 #include "opto/cfgnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/node.hpp"
47 #include "opto/optoreg.hpp"
48 #include "opto/output.hpp"
49 #include "opto/regalloc.hpp"
50 #include "opto/type.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/macros.hpp"
53 #include "utilities/powerOfTwo.hpp"
54 #include "utilities/xmlstream.hpp"
55
56 #ifndef PRODUCT
57 #define DEBUG_ARG(x) , x
58 #else
59 #define DEBUG_ARG(x)
60 #endif
61
62 //------------------------------Scheduling----------------------------------
63 // This class contains all the information necessary to implement instruction
64 // scheduling and bundling.
65 class Scheduling {
66
67 private:
68 // Arena to use
69 Arena *_arena;
70
71 // Control-Flow Graph info
72 PhaseCFG *_cfg;
73
74 // Register Allocation info
75 PhaseRegAlloc *_regalloc;
76
77 // Number of nodes in the method
78 uint _node_bundling_limit;
79
80 // List of scheduled nodes. Generated in reverse order
81 Node_List _scheduled;
82
83 // List of nodes currently available for choosing for scheduling
84 Node_List _available;
85
86 // For each instruction beginning a bundle, the number of following
87 // nodes to be bundled with it.
88 Bundle *_node_bundling_base;
89
90 // Mapping from register to Node
91 Node_List _reg_node;
92
93 // Free list for pinch nodes.
94 Node_List _pinch_free_list;
95
96 // Number of uses of this node within the containing basic block.
97 short *_uses;
98
99 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
100 // front, branch+proj at end. Also skips Catch/CProj (same as
101 // branch-at-end), plus just-prior exception-throwing call.
102 uint _bb_start, _bb_end;
103
104 // Latency from the end of the basic block as scheduled
105 unsigned short *_current_latency;
106
107 // Remember the next node
108 Node *_next_node;
109
110 // Length of the current bundle, in instructions
111 uint _bundle_instr_count;
112
113 // Current Cycle number, for computing latencies and bundling
114 uint _bundle_cycle_number;
115
116 // Bundle information
117 Pipeline_Use_Element _bundle_use_elements[resource_count];
118 Pipeline_Use _bundle_use;
119
120 // Dump the available list
121 void dump_available() const;
122
123 public:
124 Scheduling(Arena *arena, Compile &compile);
125
126 // Step ahead "i" cycles
127 void step(uint i);
128
129 // Step ahead 1 cycle, and clear the bundle state (for example,
130 // at a branch target)
131 void step_and_clear();
132
133 Bundle* node_bundling(const Node *n) {
134 assert(valid_bundle_info(n), "oob");
135 return (&_node_bundling_base[n->_idx]);
136 }
137
138 bool valid_bundle_info(const Node *n) const {
139 return (_node_bundling_limit > n->_idx);
140 }
141
142 bool starts_bundle(const Node *n) const {
143 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
144 }
145
146 // Do the scheduling
147 void DoScheduling();
148
149 // Compute the register antidependencies within a basic block
150 void ComputeRegisterAntidependencies(Block *bb);
151 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
152 void verify_good_schedule( Block *b, const char *msg );
153 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
154 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
155
156 // Add a node to the current bundle
157 void AddNodeToBundle(Node *n, const Block *bb);
158
159 // Return an integer less than, equal to, or greater than zero
160 // if the stack offset of the first argument is respectively
161 // less than, equal to, or greater than the second.
162 int compare_two_spill_nodes(Node* first, Node* second);
163
164 // Add a node to the list of available nodes
165 void AddNodeToAvailableList(Node *n);
166
167 // Compute the local use count for the nodes in a block, and compute
168 // the list of instructions with no uses in the block as available
169 void ComputeUseCount(const Block *bb);
170
171 // Choose an instruction from the available list to add to the bundle
172 Node * ChooseNodeToBundle();
173
174 // See if this Node fits into the currently accumulating bundle
175 bool NodeFitsInBundle(Node *n);
176
177 // Decrement the use count for a node
178 void DecrementUseCounts(Node *n, const Block *bb);
179
180 // Garbage collect pinch nodes for reuse by other blocks.
181 void garbage_collect_pinch_nodes();
182 // Clean up a pinch node for reuse (helper for above).
183 void cleanup_pinch( Node *pinch );
184
185 // Information for statistics gathering
186 #ifndef PRODUCT
187 private:
188 // Gather information on size of nops relative to total
189 static uint _total_nop_size, _total_method_size;
190 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
191
192 public:
193 static void print_statistics();
194
195 static void increment_instructions_per_bundle(uint i) {
196 _total_instructions_per_bundle[i]++;
197 }
198
199 static void increment_nop_size(uint s) {
200 _total_nop_size += s;
201 }
202
203 static void increment_method_size(uint s) {
204 _total_method_size += s;
205 }
206 #endif
207
208 };
209
210 PhaseOutput::PhaseOutput()
211 : Phase(Phase::Output),
212 _code_buffer("Compile::Fill_buffer"),
213 _first_block_size(0),
214 _handler_table(),
215 _inc_table(),
216 _stub_list(),
217 _oop_map_set(nullptr),
218 _scratch_buffer_blob(nullptr),
219 _scratch_locs_memory(nullptr),
220 _scratch_const_size(-1),
221 _in_scratch_emit_size(false),
222 _frame_slots(0),
223 _code_offsets(),
224 _node_bundling_limit(0),
225 _node_bundling_base(nullptr),
226 _orig_pc_slot(0),
227 _orig_pc_slot_offset_in_bytes(0),
228 _buf_sizes(),
229 _block(nullptr),
230 _index(0) {
231 C->set_output(this);
232 if (C->stub_name() == nullptr) {
233 int fixed_slots = C->fixed_slots();
234 if (C->needs_stack_repair()) {
235 fixed_slots -= 2;
236 }
237 // TODO 8284443 Only reserve extra slot if needed
238 if (InlineTypeReturnedAsFields) {
239 fixed_slots -= 2;
240 }
241 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
242 }
243 }
244
245 PhaseOutput::~PhaseOutput() {
246 C->set_output(nullptr);
247 if (_scratch_buffer_blob != nullptr) {
248 BufferBlob::free(_scratch_buffer_blob);
249 }
250 }
251
252 void PhaseOutput::perform_mach_node_analysis() {
253 // Late barrier analysis must be done after schedule and bundle
254 // Otherwise liveness based spilling will fail
255 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
256 bs->late_barrier_analysis();
257
258 pd_perform_mach_node_analysis();
259
260 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
261 }
262
263 // Convert Nodes to instruction bits and pass off to the VM
264 void PhaseOutput::Output() {
265 // RootNode goes
266 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
267
268 // The number of new nodes (mostly MachNop) is proportional to
269 // the number of java calls and inner loops which are aligned.
270 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
271 C->inner_loops()*(OptoLoopAlignment-1)),
272 "out of nodes before code generation" ) ) {
273 return;
274 }
275 // Make sure I can find the Start Node
276 Block *entry = C->cfg()->get_block(1);
277 Block *broot = C->cfg()->get_root_block();
278
279 const StartNode *start = entry->head()->as_Start();
280
281 // Replace StartNode with prolog
282 Label verified_entry;
283 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
284 entry->map_node(prolog, 0);
285 C->cfg()->map_node_to_block(prolog, entry);
286 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
287
288 // Virtual methods need an unverified entry point
289 if (C->is_osr_compilation()) {
290 if (PoisonOSREntry) {
291 // TODO: Should use a ShouldNotReachHereNode...
292 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
293 }
294 } else {
295 if (C->method()) {
296 if (C->method()->has_scalarized_args()) {
297 // Add entry point to unpack all inline type arguments
298 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
299 if (!C->method()->is_static()) {
300 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
301 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
302 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
303 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
304 }
305 } else if (!C->method()->is_static()) {
306 // Insert unvalidated entry point
307 C->cfg()->insert(broot, 0, new MachUEPNode());
308 }
309 }
310 }
311
312 // Break before main entry point
313 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
314 (OptoBreakpoint && C->is_method_compilation()) ||
315 (OptoBreakpointOSR && C->is_osr_compilation()) ||
316 (OptoBreakpointC2R && !C->method()) ) {
317 // checking for C->method() means that OptoBreakpoint does not apply to
318 // runtime stubs or frame converters
319 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
320 }
321
322 // Insert epilogs before every return
323 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
324 Block* block = C->cfg()->get_block(i);
325 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
326 Node* m = block->end();
327 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
328 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
329 block->add_inst(epilog);
330 C->cfg()->map_node_to_block(epilog, block);
331 }
332 }
333 }
334
335 // Keeper of sizing aspects
336 _buf_sizes = BufferSizingData();
337
338 // Initialize code buffer
339 estimate_buffer_size(_buf_sizes._const);
340 if (C->failing()) return;
341
342 // Pre-compute the length of blocks and replace
343 // long branches with short if machine supports it.
344 // Must be done before ScheduleAndBundle due to SPARC delay slots
345 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
346 blk_starts[0] = 0;
347 shorten_branches(blk_starts);
348
349 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
350 // Compute the offsets of the entry points required by the inline type calling convention
351 if (!C->method()->is_static()) {
352 // We have entries at the beginning of the method, implemented by the first 4 nodes.
353 // Entry (unverified) @ offset 0
354 // Verified_Inline_Entry_RO
355 // Inline_Entry (unverified)
356 // Verified_Inline_Entry
357 uint offset = 0;
358 _code_offsets.set_value(CodeOffsets::Entry, offset);
359
360 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
361 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
362
363 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
364 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
365
366 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
367 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
368 } else {
369 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
370 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
371 }
372 }
373
374 ScheduleAndBundle();
375 if (C->failing()) {
376 return;
377 }
378
379 perform_mach_node_analysis();
380
381 // Complete sizing of codebuffer
382 CodeBuffer* cb = init_buffer();
383 if (cb == nullptr || C->failing()) {
384 return;
385 }
386
387 BuildOopMaps();
388
389 if (C->failing()) {
390 return;
391 }
392
393 C2_MacroAssembler masm(cb);
394 fill_buffer(&masm, blk_starts);
395 }
396
397 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
398 // Determine if we need to generate a stack overflow check.
399 // Do it if the method is not a stub function and
400 // has java calls or has frame size > vm_page_size/8.
401 // The debug VM checks that deoptimization doesn't trigger an
402 // unexpected stack overflow (compiled method stack banging should
403 // guarantee it doesn't happen) so we always need the stack bang in
404 // a debug VM.
405 return (C->stub_function() == nullptr &&
406 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
407 DEBUG_ONLY(|| true)));
408 }
409
410 bool PhaseOutput::need_register_stack_bang() const {
411 // Determine if we need to generate a register stack overflow check.
412 // This is only used on architectures which have split register
413 // and memory stacks.
414 // Bang if the method is not a stub function and has java calls
415 return (C->stub_function() == nullptr && C->has_java_calls());
416 }
417
418
419 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
420 // of a loop. When aligning a loop we need to provide enough instructions
421 // in cpu's fetch buffer to feed decoders. The loop alignment could be
422 // avoided if we have enough instructions in fetch buffer at the head of a loop.
423 // By default, the size is set to 999999 by Block's constructor so that
424 // a loop will be aligned if the size is not reset here.
425 //
426 // Note: Mach instructions could contain several HW instructions
427 // so the size is estimated only.
428 //
429 void PhaseOutput::compute_loop_first_inst_sizes() {
430 // The next condition is used to gate the loop alignment optimization.
431 // Don't aligned a loop if there are enough instructions at the head of a loop
432 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
433 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
434 // equal to 11 bytes which is the largest address NOP instruction.
435 if (MaxLoopPad < OptoLoopAlignment - 1) {
436 uint last_block = C->cfg()->number_of_blocks() - 1;
437 for (uint i = 1; i <= last_block; i++) {
438 Block* block = C->cfg()->get_block(i);
439 // Check the first loop's block which requires an alignment.
440 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
441 uint sum_size = 0;
442 uint inst_cnt = NumberOfLoopInstrToAlign;
443 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
444
445 // Check subsequent fallthrough blocks if the loop's first
446 // block(s) does not have enough instructions.
447 Block *nb = block;
448 while(inst_cnt > 0 &&
449 i < last_block &&
450 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
451 !nb->has_successor(block)) {
452 i++;
453 nb = C->cfg()->get_block(i);
454 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
455 } // while( inst_cnt > 0 && i < last_block )
456
457 block->set_first_inst_size(sum_size);
458 } // f( b->head()->is_Loop() )
459 } // for( i <= last_block )
460 } // if( MaxLoopPad < OptoLoopAlignment-1 )
461 }
462
463 // The architecture description provides short branch variants for some long
464 // branch instructions. Replace eligible long branches with short branches.
465 void PhaseOutput::shorten_branches(uint* blk_starts) {
466
467 Compile::TracePhase tp(_t_shortenBranches);
468
469 // Compute size of each block, method size, and relocation information size
470 uint nblocks = C->cfg()->number_of_blocks();
471
472 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
473 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
474 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
475
476 // Collect worst case block paddings
477 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
478 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
479
480 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
481 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
482
483 bool has_short_branch_candidate = false;
484
485 // Initialize the sizes to 0
486 int code_size = 0; // Size in bytes of generated code
487 int stub_size = 0; // Size in bytes of all stub entries
488 // Size in bytes of all relocation entries, including those in local stubs.
489 // Start with 2-bytes of reloc info for the unvalidated entry point
490 int reloc_size = 1; // Number of relocation entries
491
492 // Make three passes. The first computes pessimistic blk_starts,
493 // relative jmp_offset and reloc_size information. The second performs
494 // short branch substitution using the pessimistic sizing. The
495 // third inserts nops where needed.
496
497 // Step one, perform a pessimistic sizing pass.
498 uint last_call_adr = max_juint;
499 uint last_avoid_back_to_back_adr = max_juint;
500 uint nop_size = (new MachNopNode())->size(C->regalloc());
501 for (uint i = 0; i < nblocks; i++) { // For all blocks
502 Block* block = C->cfg()->get_block(i);
503 _block = block;
504
505 // During short branch replacement, we store the relative (to blk_starts)
506 // offset of jump in jmp_offset, rather than the absolute offset of jump.
507 // This is so that we do not need to recompute sizes of all nodes when
508 // we compute correct blk_starts in our next sizing pass.
509 jmp_offset[i] = 0;
510 jmp_size[i] = 0;
511 jmp_nidx[i] = -1;
512 DEBUG_ONLY( jmp_target[i] = 0; )
513 DEBUG_ONLY( jmp_rule[i] = 0; )
514
515 // Sum all instruction sizes to compute block size
516 uint last_inst = block->number_of_nodes();
517 uint blk_size = 0;
518 for (uint j = 0; j < last_inst; j++) {
519 _index = j;
520 Node* nj = block->get_node(_index);
521 // Handle machine instruction nodes
522 if (nj->is_Mach()) {
523 MachNode* mach = nj->as_Mach();
524 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
525 reloc_size += mach->reloc();
526 if (mach->is_MachCall()) {
527 // add size information for trampoline stub
528 // class CallStubImpl is platform-specific and defined in the *.ad files.
529 stub_size += CallStubImpl::size_call_trampoline();
530 reloc_size += CallStubImpl::reloc_call_trampoline();
531
532 MachCallNode *mcall = mach->as_MachCall();
533 // This destination address is NOT PC-relative
534
535 if (mcall->entry_point() != nullptr) {
536 mcall->method_set((intptr_t)mcall->entry_point());
537 }
538
539 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
540 stub_size += CompiledDirectCall::to_interp_stub_size();
541 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
542 }
543 } else if (mach->is_MachSafePoint()) {
544 // If call/safepoint are adjacent, account for possible
545 // nop to disambiguate the two safepoints.
546 // ScheduleAndBundle() can rearrange nodes in a block,
547 // check for all offsets inside this block.
548 if (last_call_adr >= blk_starts[i]) {
549 blk_size += nop_size;
550 }
551 }
552 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
553 // Nop is inserted between "avoid back to back" instructions.
554 // ScheduleAndBundle() can rearrange nodes in a block,
555 // check for all offsets inside this block.
556 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
557 blk_size += nop_size;
558 }
559 }
560 if (mach->may_be_short_branch()) {
561 if (!nj->is_MachBranch()) {
562 #ifndef PRODUCT
563 nj->dump(3);
564 #endif
565 Unimplemented();
566 }
567 assert(jmp_nidx[i] == -1, "block should have only one branch");
568 jmp_offset[i] = blk_size;
569 jmp_size[i] = nj->size(C->regalloc());
570 jmp_nidx[i] = j;
571 has_short_branch_candidate = true;
572 }
573 }
574 blk_size += nj->size(C->regalloc());
575 // Remember end of call offset
576 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
577 last_call_adr = blk_starts[i]+blk_size;
578 }
579 // Remember end of avoid_back_to_back offset
580 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
581 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
582 }
583 }
584
585 // When the next block starts a loop, we may insert pad NOP
586 // instructions. Since we cannot know our future alignment,
587 // assume the worst.
588 if (i < nblocks - 1) {
589 Block* nb = C->cfg()->get_block(i + 1);
590 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
591 if (max_loop_pad > 0) {
592 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
593 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
594 // If either is the last instruction in this block, bump by
595 // max_loop_pad in lock-step with blk_size, so sizing
596 // calculations in subsequent blocks still can conservatively
597 // detect that it may the last instruction in this block.
598 if (last_call_adr == blk_starts[i]+blk_size) {
599 last_call_adr += max_loop_pad;
600 }
601 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
602 last_avoid_back_to_back_adr += max_loop_pad;
603 }
604 blk_size += max_loop_pad;
605 block_worst_case_pad[i + 1] = max_loop_pad;
606 }
607 }
608
609 // Save block size; update total method size
610 blk_starts[i+1] = blk_starts[i]+blk_size;
611 }
612
613 // Step two, replace eligible long jumps.
614 bool progress = true;
615 uint last_may_be_short_branch_adr = max_juint;
616 while (has_short_branch_candidate && progress) {
617 progress = false;
618 has_short_branch_candidate = false;
619 int adjust_block_start = 0;
620 for (uint i = 0; i < nblocks; i++) {
621 Block* block = C->cfg()->get_block(i);
622 int idx = jmp_nidx[i];
623 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
624 if (mach != nullptr && mach->may_be_short_branch()) {
625 #ifdef ASSERT
626 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
627 int j;
628 // Find the branch; ignore trailing NOPs.
629 for (j = block->number_of_nodes()-1; j>=0; j--) {
630 Node* n = block->get_node(j);
631 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
632 break;
633 }
634 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
635 #endif
636 int br_size = jmp_size[i];
637 int br_offs = blk_starts[i] + jmp_offset[i];
638
639 // This requires the TRUE branch target be in succs[0]
640 uint bnum = block->non_connector_successor(0)->_pre_order;
641 int offset = blk_starts[bnum] - br_offs;
642 if (bnum > i) { // adjust following block's offset
643 offset -= adjust_block_start;
644 }
645
646 // This block can be a loop header, account for the padding
647 // in the previous block.
648 int block_padding = block_worst_case_pad[i];
649 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
650 // In the following code a nop could be inserted before
651 // the branch which will increase the backward distance.
652 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
653 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
654
655 if (needs_padding && offset <= 0)
656 offset -= nop_size;
657
658 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
659 // We've got a winner. Replace this branch.
660 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
661
662 // Update the jmp_size.
663 int new_size = replacement->size(C->regalloc());
664 int diff = br_size - new_size;
665 assert(diff >= (int)nop_size, "short_branch size should be smaller");
666 // Conservatively take into account padding between
667 // avoid_back_to_back branches. Previous branch could be
668 // converted into avoid_back_to_back branch during next
669 // rounds.
670 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
671 jmp_offset[i] += nop_size;
672 diff -= nop_size;
673 }
674 adjust_block_start += diff;
675 block->map_node(replacement, idx);
676 mach->subsume_by(replacement, C);
677 mach = replacement;
678 progress = true;
679
680 jmp_size[i] = new_size;
681 DEBUG_ONLY( jmp_target[i] = bnum; );
682 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
683 } else {
684 // The jump distance is not short, try again during next iteration.
685 has_short_branch_candidate = true;
686 }
687 } // (mach->may_be_short_branch())
688 if (mach != nullptr && (mach->may_be_short_branch() ||
689 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
690 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
691 }
692 blk_starts[i+1] -= adjust_block_start;
693 }
694 }
695
696 #ifdef ASSERT
697 for (uint i = 0; i < nblocks; i++) { // For all blocks
698 if (jmp_target[i] != 0) {
699 int br_size = jmp_size[i];
700 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
701 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
702 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]);
703 }
704 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
705 }
706 }
707 #endif
708
709 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
710 // after ScheduleAndBundle().
711
712 // ------------------
713 // Compute size for code buffer
714 code_size = blk_starts[nblocks];
715
716 // Relocation records
717 reloc_size += 1; // Relo entry for exception handler
718
719 // Adjust reloc_size to number of record of relocation info
720 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
721 // a relocation index.
722 // The CodeBuffer will expand the locs array if this estimate is too low.
723 reloc_size *= 10 / sizeof(relocInfo);
724
725 _buf_sizes._reloc = reloc_size;
726 _buf_sizes._code = code_size;
727 _buf_sizes._stub = stub_size;
728 }
729
730 //------------------------------FillLocArray-----------------------------------
731 // Create a bit of debug info and append it to the array. The mapping is from
732 // Java local or expression stack to constant, register or stack-slot. For
733 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
734 // entry has been taken care of and caller should skip it).
735 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
736 // This should never have accepted Bad before
737 assert(OptoReg::is_valid(regnum), "location must be valid");
738 return (OptoReg::is_reg(regnum))
739 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
740 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
741 }
742
743
744 ObjectValue*
745 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
746 for (int i = 0; i < objs->length(); i++) {
747 assert(objs->at(i)->is_object(), "corrupt object cache");
748 ObjectValue* sv = objs->at(i)->as_ObjectValue();
749 if (sv->id() == id) {
750 return sv;
751 }
752 }
753 // Otherwise..
754 return nullptr;
755 }
756
757 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
758 ObjectValue* sv ) {
759 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
760 objs->append(sv);
761 }
762
763
764 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
765 GrowableArray<ScopeValue*> *array,
766 GrowableArray<ScopeValue*> *objs ) {
767 assert( local, "use _top instead of null" );
768 if (array->length() != idx) {
769 assert(array->length() == idx + 1, "Unexpected array count");
770 // Old functionality:
771 // return
772 // New functionality:
773 // Assert if the local is not top. In product mode let the new node
774 // override the old entry.
775 assert(local == C->top(), "LocArray collision");
776 if (local == C->top()) {
777 return;
778 }
779 array->pop();
780 }
781 const Type *t = local->bottom_type();
782
783 // Is it a safepoint scalar object node?
784 if (local->is_SafePointScalarObject()) {
785 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
786
787 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
788 if (sv == nullptr) {
789 ciKlass* cik = t->is_oopptr()->exact_klass();
790 assert(cik->is_instance_klass() ||
791 cik->is_array_klass(), "Not supported allocation.");
792 uint first_ind = spobj->first_index(sfpt->jvms());
793 // Nullable, scalarized inline types have a null_marker input
794 // that needs to be checked before using the field values.
795 ScopeValue* properties = nullptr;
796 if (cik->is_inlinetype()) {
797 Node* null_marker_node = sfpt->in(first_ind++);
798 assert(null_marker_node != nullptr, "null_marker node not found");
799 if (!null_marker_node->is_top()) {
800 const TypeInt* null_marker_type = null_marker_node->bottom_type()->is_int();
801 if (null_marker_node->is_Con()) {
802 properties = new ConstantIntValue(null_marker_type->get_con());
803 } else {
804 OptoReg::Name null_marker_reg = C->regalloc()->get_reg_first(null_marker_node);
805 properties = new_loc_value(C->regalloc(), null_marker_reg, Location::normal);
806 }
807 }
808 }
809 if (cik->is_array_klass() && !cik->is_type_array_klass()) {
810 jint props = ArrayKlass::ArrayProperties::DEFAULT;
811 if (cik->as_array_klass()->element_klass()->is_inlinetype()) {
812 if (cik->as_array_klass()->is_elem_null_free()) {
813 props |= ArrayKlass::ArrayProperties::NULL_RESTRICTED;
814 }
815 if (!cik->as_array_klass()->is_elem_atomic()) {
816 props |= ArrayKlass::ArrayProperties::NON_ATOMIC;
817 }
818 }
819 properties = new ConstantIntValue(props);
820 }
821 sv = new ObjectValue(spobj->_idx,
822 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, properties);
823 set_sv_for_object_node(objs, sv);
824
825 for (uint i = 0; i < spobj->n_fields(); i++) {
826 Node* fld_node = sfpt->in(first_ind+i);
827 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
828 }
829 }
830 array->append(sv);
831 return;
832 } else if (local->is_SafePointScalarMerge()) {
833 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
834 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
835
836 if (mv == nullptr) {
837 GrowableArray<ScopeValue*> deps;
838
839 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
840 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
841 assert(deps.length() == 1, "missing value");
842
843 int selector_idx = smerge->selector_idx(sfpt->jvms());
844 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
845 assert(deps.length() == 2, "missing value");
846
847 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
848 set_sv_for_object_node(objs, mv);
849
850 for (uint i = 1; i < smerge->req(); i++) {
851 Node* obj_node = smerge->in(i);
852 int idx = mv->possible_objects()->length();
853 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
854
855 // By default ObjectValues that are in 'possible_objects' are not root objects.
856 // They will be marked as root later if they are directly referenced in a JVMS.
857 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
858 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
859 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
860 }
861 }
862 array->append(mv);
863 return;
864 }
865
866 // Grab the register number for the local
867 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
868 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
869 // Record the double as two float registers.
870 // The register mask for such a value always specifies two adjacent
871 // float registers, with the lower register number even.
872 // Normally, the allocation of high and low words to these registers
873 // is irrelevant, because nearly all operations on register pairs
874 // (e.g., StoreD) treat them as a single unit.
875 // Here, we assume in addition that the words in these two registers
876 // stored "naturally" (by operations like StoreD and double stores
877 // within the interpreter) such that the lower-numbered register
878 // is written to the lower memory address. This may seem like
879 // a machine dependency, but it is not--it is a requirement on
880 // the author of the <arch>.ad file to ensure that, for every
881 // even/odd double-register pair to which a double may be allocated,
882 // the word in the even single-register is stored to the first
883 // memory word. (Note that register numbers are completely
884 // arbitrary, and are not tied to any machine-level encodings.)
885 #ifdef _LP64
886 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
887 array->append(new ConstantIntValue((jint)0));
888 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
889 } else if ( t->base() == Type::Long ) {
890 array->append(new ConstantIntValue((jint)0));
891 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
892 } else if ( t->base() == Type::RawPtr ) {
893 // jsr/ret return address which must be restored into the full
894 // width 64-bit stack slot.
895 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
896 }
897 #else //_LP64
898 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
899 // Repack the double/long as two jints.
900 // The convention the interpreter uses is that the second local
901 // holds the first raw word of the native double representation.
902 // This is actually reasonable, since locals and stack arrays
903 // grow downwards in all implementations.
904 // (If, on some machine, the interpreter's Java locals or stack
905 // were to grow upwards, the embedded doubles would be word-swapped.)
906 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
907 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
908 }
909 #endif //_LP64
910 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
911 OptoReg::is_reg(regnum) ) {
912 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
913 ? Location::float_in_dbl : Location::normal ));
914 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
915 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
916 ? Location::int_in_long : Location::normal ));
917 } else if( t->base() == Type::NarrowOop ) {
918 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
919 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
920 t->base() == Type::VectorD || t->base() == Type::VectorX ||
921 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
922 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
923 } else if (C->regalloc()->is_oop(local)) {
924 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
925 t->base() == Type::AryPtr,
926 "Unexpected type: %s", t->msg());
927 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
928 } else {
929 assert(t->base() == Type::Int || t->base() == Type::Half ||
930 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
931 "Unexpected type: %s", t->msg());
932 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
933 }
934 return;
935 }
936
937 // No register. It must be constant data.
938 switch (t->base()) {
939 case Type::Half: // Second half of a double
940 ShouldNotReachHere(); // Caller should skip 2nd halves
941 break;
942 case Type::AnyPtr:
943 array->append(new ConstantOopWriteValue(nullptr));
944 break;
945 case Type::AryPtr:
946 case Type::InstPtr: // fall through
947 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
948 break;
949 case Type::NarrowOop:
950 if (t == TypeNarrowOop::NULL_PTR) {
951 array->append(new ConstantOopWriteValue(nullptr));
952 } else {
953 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
954 }
955 break;
956 case Type::Int:
957 array->append(new ConstantIntValue(t->is_int()->get_con()));
958 break;
959 case Type::RawPtr:
960 // A return address (T_ADDRESS).
961 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
962 #ifdef _LP64
963 // Must be restored to the full-width 64-bit stack slot.
964 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
965 #else
966 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
967 #endif
968 break;
969 case Type::FloatCon: {
970 float f = t->is_float_constant()->getf();
971 array->append(new ConstantIntValue(jint_cast(f)));
972 break;
973 }
974 case Type::DoubleCon: {
975 jdouble d = t->is_double_constant()->getd();
976 #ifdef _LP64
977 array->append(new ConstantIntValue((jint)0));
978 array->append(new ConstantDoubleValue(d));
979 #else
980 // Repack the double as two jints.
981 // The convention the interpreter uses is that the second local
982 // holds the first raw word of the native double representation.
983 // This is actually reasonable, since locals and stack arrays
984 // grow downwards in all implementations.
985 // (If, on some machine, the interpreter's Java locals or stack
986 // were to grow upwards, the embedded doubles would be word-swapped.)
987 jlong_accessor acc;
988 acc.long_value = jlong_cast(d);
989 array->append(new ConstantIntValue(acc.words[1]));
990 array->append(new ConstantIntValue(acc.words[0]));
991 #endif
992 break;
993 }
994 case Type::Long: {
995 jlong d = t->is_long()->get_con();
996 #ifdef _LP64
997 array->append(new ConstantIntValue((jint)0));
998 array->append(new ConstantLongValue(d));
999 #else
1000 // Repack the long as two jints.
1001 // The convention the interpreter uses is that the second local
1002 // holds the first raw word of the native double representation.
1003 // This is actually reasonable, since locals and stack arrays
1004 // grow downwards in all implementations.
1005 // (If, on some machine, the interpreter's Java locals or stack
1006 // were to grow upwards, the embedded doubles would be word-swapped.)
1007 jlong_accessor acc;
1008 acc.long_value = d;
1009 array->append(new ConstantIntValue(acc.words[1]));
1010 array->append(new ConstantIntValue(acc.words[0]));
1011 #endif
1012 break;
1013 }
1014 case Type::Top: // Add an illegal value here
1015 array->append(new LocationValue(Location()));
1016 break;
1017 default:
1018 ShouldNotReachHere();
1019 break;
1020 }
1021 }
1022
1023 // Determine if this node starts a bundle
1024 bool PhaseOutput::starts_bundle(const Node *n) const {
1025 return (_node_bundling_limit > n->_idx &&
1026 _node_bundling_base[n->_idx].starts_bundle());
1027 }
1028
1029 // Determine if there is a monitor that has 'ov' as its owner.
1030 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
1031 for (int k = 0; k < monarray->length(); k++) {
1032 MonitorValue* mv = monarray->at(k);
1033 if (mv->owner() == ov) {
1034 return true;
1035 }
1036 }
1037
1038 return false;
1039 }
1040
1041 // Determine if there is a scalar replaced object description represented by 'ov'.
1042 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
1043 GrowableArray<ScopeValue*>* objs,
1044 ObjectValue* ov) const {
1045 for (int i = 0; i < jvms->scl_size(); i++) {
1046 Node* n = sfn->scalarized_obj(jvms, i);
1047 // Other kinds of nodes that we may encounter here, for instance constants
1048 // representing values of fields of objects scalarized, aren't relevant for
1049 // us, since they don't map to ObjectValue.
1050 if (!n->is_SafePointScalarObject()) {
1051 continue;
1052 }
1053
1054 ObjectValue* other = sv_for_node_id(objs, n->_idx);
1055 if (ov == other) {
1056 return true;
1057 }
1058 }
1059 return false;
1060 }
1061
1062 //--------------------------Process_OopMap_Node--------------------------------
1063 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1064 // Handle special safepoint nodes for synchronization
1065 MachSafePointNode *sfn = mach->as_MachSafePoint();
1066 MachCallNode *mcall;
1067
1068 int safepoint_pc_offset = current_offset;
1069 bool return_oop = false;
1070 bool return_scalarized = false;
1071 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1072 bool arg_escape = false;
1073
1074 // Add the safepoint in the DebugInfoRecorder
1075 if( !mach->is_MachCall() ) {
1076 mcall = nullptr;
1077 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1078 } else {
1079 mcall = mach->as_MachCall();
1080
1081 if (mcall->is_MachCallJava()) {
1082 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1083 }
1084
1085 // Check if a call returns an object.
1086 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1087 return_oop = true;
1088 }
1089 if (mcall->returns_scalarized()) {
1090 return_scalarized = true;
1091 }
1092 safepoint_pc_offset += mcall->ret_addr_offset();
1093 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1094 }
1095
1096 // Loop over the JVMState list to add scope information
1097 // Do not skip safepoints with a null method, they need monitor info
1098 JVMState* youngest_jvms = sfn->jvms();
1099 int max_depth = youngest_jvms->depth();
1100
1101 // Allocate the object pool for scalar-replaced objects -- the map from
1102 // small-integer keys (which can be recorded in the local and ostack
1103 // arrays) to descriptions of the object state.
1104 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1105
1106 // Visit scopes from oldest to youngest.
1107 for (int depth = 1; depth <= max_depth; depth++) {
1108 JVMState* jvms = youngest_jvms->of_depth(depth);
1109 int idx;
1110 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1111 // Safepoints that do not have method() set only provide oop-map and monitor info
1112 // to support GC; these do not support deoptimization.
1113 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1114 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1115 int num_mon = jvms->nof_monitors();
1116 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1117 "JVMS local count must match that of the method");
1118
1119 // Add Local and Expression Stack Information
1120
1121 // Insert locals into the locarray
1122 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1123 for( idx = 0; idx < num_locs; idx++ ) {
1124 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1125 }
1126
1127 // Insert expression stack entries into the exparray
1128 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1129 for( idx = 0; idx < num_exps; idx++ ) {
1130 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1131 }
1132
1133 // Add in mappings of the monitors
1134 assert( !method ||
1135 !method->is_synchronized() ||
1136 method->is_native() ||
1137 num_mon > 0,
1138 "monitors must always exist for synchronized methods");
1139
1140 // Build the growable array of ScopeValues for exp stack
1141 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1142
1143 // Loop over monitors and insert into array
1144 for (idx = 0; idx < num_mon; idx++) {
1145 // Grab the node that defines this monitor
1146 Node* box_node = sfn->monitor_box(jvms, idx);
1147 Node* obj_node = sfn->monitor_obj(jvms, idx);
1148
1149 // Create ScopeValue for object
1150 ScopeValue *scval = nullptr;
1151
1152 if (obj_node->is_SafePointScalarObject()) {
1153 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1154 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1155 if (scval == nullptr) {
1156 const Type *t = spobj->bottom_type();
1157 ciKlass* cik = t->is_oopptr()->exact_klass();
1158 assert(cik->is_instance_klass() ||
1159 cik->is_array_klass(), "Not supported allocation.");
1160 assert(!cik->is_inlinetype(), "Synchronization on value object?");
1161 ScopeValue* properties = nullptr;
1162 if (cik->is_array_klass() && !cik->is_type_array_klass()) {
1163 jint props = ArrayKlass::ArrayProperties::DEFAULT;
1164 if (cik->as_array_klass()->element_klass()->is_inlinetype()) {
1165 if (cik->as_array_klass()->is_elem_null_free()) {
1166 props |= ArrayKlass::ArrayProperties::NULL_RESTRICTED;
1167 }
1168 if (!cik->as_array_klass()->is_elem_atomic()) {
1169 props |= ArrayKlass::ArrayProperties::NON_ATOMIC;
1170 }
1171 }
1172 properties = new ConstantIntValue(props);
1173 }
1174 ObjectValue* sv = new ObjectValue(spobj->_idx,
1175 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, properties);
1176 PhaseOutput::set_sv_for_object_node(objs, sv);
1177
1178 uint first_ind = spobj->first_index(youngest_jvms);
1179 for (uint i = 0; i < spobj->n_fields(); i++) {
1180 Node* fld_node = sfn->in(first_ind+i);
1181 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1182 }
1183 scval = sv;
1184 }
1185 } else if (obj_node->is_SafePointScalarMerge()) {
1186 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1187 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1188
1189 if (mv == nullptr) {
1190 GrowableArray<ScopeValue*> deps;
1191
1192 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1193 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1194 assert(deps.length() == 1, "missing value");
1195
1196 int selector_idx = smerge->selector_idx(youngest_jvms);
1197 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1198 assert(deps.length() == 2, "missing value");
1199
1200 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1201 set_sv_for_object_node(objs, mv);
1202
1203 for (uint i = 1; i < smerge->req(); i++) {
1204 Node* obj_node = smerge->in(i);
1205 int idx = mv->possible_objects()->length();
1206 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1207
1208 // By default ObjectValues that are in 'possible_objects' are not root objects.
1209 // They will be marked as root later if they are directly referenced in a JVMS.
1210 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1211 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1212 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1213 }
1214 }
1215 scval = mv;
1216 } else if (!obj_node->is_Con()) {
1217 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1218 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1219 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1220 } else {
1221 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1222 }
1223 } else {
1224 const TypePtr *tp = obj_node->get_ptr_type();
1225 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1226 }
1227
1228 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1229 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1230 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1231 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1232 }
1233
1234 // Mark ObjectValue nodes as root nodes if they are directly
1235 // referenced in the JVMS.
1236 for (int i = 0; i < objs->length(); i++) {
1237 ScopeValue* sv = objs->at(i);
1238 if (sv->is_object_merge()) {
1239 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1240
1241 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1242 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1243 if (ov->is_root()) {
1244 // Already flagged as 'root' by something else. We shouldn't change it
1245 // to non-root in a younger JVMS because it may need to be alive in
1246 // a younger JVMS.
1247 } else {
1248 bool is_root = locarray->contains(ov) ||
1249 exparray->contains(ov) ||
1250 contains_as_owner(monarray, ov) ||
1251 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1252 ov->set_root(is_root);
1253 }
1254 }
1255 }
1256 }
1257
1258 // We dump the object pool first, since deoptimization reads it in first.
1259 C->debug_info()->dump_object_pool(objs);
1260
1261 // Build first class objects to pass to scope
1262 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1263 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1264 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1265
1266 // Make method available for all Safepoints
1267 ciMethod* scope_method = method ? method : C->method();
1268 // Describe the scope here
1269 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1270 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1271 // Now we can describe the scope.
1272 methodHandle null_mh;
1273 bool rethrow_exception = false;
1274 C->debug_info()->describe_scope(
1275 safepoint_pc_offset,
1276 null_mh,
1277 scope_method,
1278 jvms->bci(),
1279 jvms->should_reexecute(),
1280 rethrow_exception,
1281 return_oop,
1282 return_scalarized,
1283 has_ea_local_in_scope,
1284 arg_escape,
1285 locvals,
1286 expvals,
1287 monvals
1288 );
1289 } // End jvms loop
1290
1291 // Mark the end of the scope set.
1292 C->debug_info()->end_safepoint(safepoint_pc_offset);
1293 }
1294
1295
1296
1297 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1298 class NonSafepointEmitter {
1299 Compile* C;
1300 JVMState* _pending_jvms;
1301 int _pending_offset;
1302
1303 void emit_non_safepoint();
1304
1305 public:
1306 NonSafepointEmitter(Compile* compile) {
1307 this->C = compile;
1308 _pending_jvms = nullptr;
1309 _pending_offset = 0;
1310 }
1311
1312 void observe_instruction(Node* n, int pc_offset) {
1313 if (!C->debug_info()->recording_non_safepoints()) return;
1314
1315 Node_Notes* nn = C->node_notes_at(n->_idx);
1316 if (nn == nullptr || nn->jvms() == nullptr) return;
1317 if (_pending_jvms != nullptr &&
1318 _pending_jvms->same_calls_as(nn->jvms())) {
1319 // Repeated JVMS? Stretch it up here.
1320 _pending_offset = pc_offset;
1321 } else {
1322 if (_pending_jvms != nullptr &&
1323 _pending_offset < pc_offset) {
1324 emit_non_safepoint();
1325 }
1326 _pending_jvms = nullptr;
1327 if (pc_offset > C->debug_info()->last_pc_offset()) {
1328 // This is the only way _pending_jvms can become non-null:
1329 _pending_jvms = nn->jvms();
1330 _pending_offset = pc_offset;
1331 }
1332 }
1333 }
1334
1335 // Stay out of the way of real safepoints:
1336 void observe_safepoint(JVMState* jvms, int pc_offset) {
1337 if (_pending_jvms != nullptr &&
1338 !_pending_jvms->same_calls_as(jvms) &&
1339 _pending_offset < pc_offset) {
1340 emit_non_safepoint();
1341 }
1342 _pending_jvms = nullptr;
1343 }
1344
1345 void flush_at_end() {
1346 if (_pending_jvms != nullptr) {
1347 emit_non_safepoint();
1348 }
1349 _pending_jvms = nullptr;
1350 }
1351 };
1352
1353 void NonSafepointEmitter::emit_non_safepoint() {
1354 JVMState* youngest_jvms = _pending_jvms;
1355 int pc_offset = _pending_offset;
1356
1357 // Clear it now:
1358 _pending_jvms = nullptr;
1359
1360 DebugInformationRecorder* debug_info = C->debug_info();
1361 assert(debug_info->recording_non_safepoints(), "sanity");
1362
1363 debug_info->add_non_safepoint(pc_offset);
1364 int max_depth = youngest_jvms->depth();
1365
1366 // Visit scopes from oldest to youngest.
1367 for (int depth = 1; depth <= max_depth; depth++) {
1368 JVMState* jvms = youngest_jvms->of_depth(depth);
1369 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1370 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1371 methodHandle null_mh;
1372 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1373 }
1374
1375 // Mark the end of the scope set.
1376 debug_info->end_non_safepoint(pc_offset);
1377 }
1378
1379 //------------------------------init_buffer------------------------------------
1380 void PhaseOutput::estimate_buffer_size(int& const_req) {
1381
1382 // Set the initially allocated size
1383 const_req = initial_const_capacity;
1384
1385 // The extra spacing after the code is necessary on some platforms.
1386 // Sometimes we need to patch in a jump after the last instruction,
1387 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1388
1389 // Compute the byte offset where we can store the deopt pc.
1390 if (C->fixed_slots() != 0) {
1391 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1392 }
1393
1394 // Compute prolog code size
1395 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1396 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1397
1398 if (C->has_mach_constant_base_node()) {
1399 uint add_size = 0;
1400 // Fill the constant table.
1401 // Note: This must happen before shorten_branches.
1402 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1403 Block* b = C->cfg()->get_block(i);
1404
1405 for (uint j = 0; j < b->number_of_nodes(); j++) {
1406 Node* n = b->get_node(j);
1407
1408 // If the node is a MachConstantNode evaluate the constant
1409 // value section.
1410 if (n->is_MachConstant()) {
1411 MachConstantNode* machcon = n->as_MachConstant();
1412 machcon->eval_constant(C);
1413 } else if (n->is_Mach()) {
1414 // On Power there are more nodes that issue constants.
1415 add_size += (n->as_Mach()->ins_num_consts() * 8);
1416 }
1417 }
1418 }
1419
1420 // Calculate the offsets of the constants and the size of the
1421 // constant table (including the padding to the next section).
1422 constant_table().calculate_offsets_and_size();
1423 const_req = constant_table().alignment() + constant_table().size() + add_size;
1424 }
1425
1426 // Initialize the space for the BufferBlob used to find and verify
1427 // instruction size in MachNode::emit_size()
1428 init_scratch_buffer_blob(const_req);
1429 }
1430
1431 CodeBuffer* PhaseOutput::init_buffer() {
1432 int stub_req = _buf_sizes._stub;
1433 int code_req = _buf_sizes._code;
1434 int const_req = _buf_sizes._const;
1435
1436 int pad_req = NativeCall::byte_size();
1437
1438 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1439 stub_req += bs->estimate_stub_size();
1440
1441 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1442 // class HandlerImpl is platform-specific and defined in the *.ad files.
1443 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1444 stub_req += MAX_stubs_size; // ensure per-stub margin
1445 code_req += MAX_inst_size; // ensure per-instruction margin
1446
1447 if (StressCodeBuffers)
1448 code_req = const_req = stub_req = deopt_handler_req = 0x10; // force expansion
1449
1450 int total_req =
1451 const_req +
1452 code_req +
1453 pad_req +
1454 stub_req +
1455 deopt_handler_req; // deopt handler
1456
1457 CodeBuffer* cb = code_buffer();
1458 cb->set_const_section_alignment(constant_table().alignment());
1459 cb->initialize(total_req, _buf_sizes._reloc);
1460
1461 // Have we run out of code space?
1462 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1463 C->record_failure("CodeCache is full");
1464 return nullptr;
1465 }
1466 // Configure the code buffer.
1467 cb->initialize_consts_size(const_req);
1468 cb->initialize_stubs_size(stub_req);
1469 cb->initialize_oop_recorder(C->env()->oop_recorder());
1470
1471 return cb;
1472 }
1473
1474 //------------------------------fill_buffer------------------------------------
1475 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1476 // blk_starts[] contains offsets calculated during short branches processing,
1477 // offsets should not be increased during following steps.
1478
1479 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1480 // of a loop. It is used to determine the padding for loop alignment.
1481 Compile::TracePhase tp(_t_fillBuffer);
1482
1483 compute_loop_first_inst_sizes();
1484
1485 // Create oopmap set.
1486 _oop_map_set = new OopMapSet();
1487
1488 // !!!!! This preserves old handling of oopmaps for now
1489 C->debug_info()->set_oopmaps(_oop_map_set);
1490
1491 uint nblocks = C->cfg()->number_of_blocks();
1492 // Count and start of implicit null check instructions
1493 uint inct_cnt = 0;
1494 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1495
1496 // Count and start of calls
1497 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1498
1499 uint return_offset = 0;
1500 int nop_size = (new MachNopNode())->size(C->regalloc());
1501
1502 int previous_offset = 0;
1503 int current_offset = 0;
1504 int last_call_offset = -1;
1505 int last_avoid_back_to_back_offset = -1;
1506 #ifdef ASSERT
1507 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1508 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1509 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1510 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1511 #endif
1512
1513 // Create an array of unused labels, one for each basic block, if printing is enabled
1514 #if defined(SUPPORT_OPTO_ASSEMBLY)
1515 int* node_offsets = nullptr;
1516 uint node_offset_limit = C->unique();
1517
1518 if (C->print_assembly()) {
1519 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1520 }
1521 if (node_offsets != nullptr) {
1522 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1523 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1524 }
1525 #endif
1526
1527 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1528
1529 // Emit the constant table.
1530 if (C->has_mach_constant_base_node()) {
1531 if (!constant_table().emit(masm)) {
1532 C->record_failure("consts section overflow");
1533 return;
1534 }
1535 }
1536
1537 // Create an array of labels, one for each basic block
1538 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1539 for (uint i = 0; i <= nblocks; i++) {
1540 blk_labels[i].init();
1541 }
1542
1543 // Now fill in the code buffer
1544 for (uint i = 0; i < nblocks; i++) {
1545 Block* block = C->cfg()->get_block(i);
1546 _block = block;
1547 Node* head = block->head();
1548
1549 // If this block needs to start aligned (i.e, can be reached other
1550 // than by falling-thru from the previous block), then force the
1551 // start of a new bundle.
1552 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1553 masm->code()->flush_bundle(true);
1554 }
1555
1556 #ifdef ASSERT
1557 if (!block->is_connector()) {
1558 stringStream st;
1559 block->dump_head(C->cfg(), &st);
1560 masm->block_comment(st.freeze());
1561 }
1562 jmp_target[i] = 0;
1563 jmp_offset[i] = 0;
1564 jmp_size[i] = 0;
1565 jmp_rule[i] = 0;
1566 #endif
1567 int blk_offset = current_offset;
1568
1569 // Define the label at the beginning of the basic block
1570 masm->bind(blk_labels[block->_pre_order]);
1571
1572 uint last_inst = block->number_of_nodes();
1573
1574 // Emit block normally, except for last instruction.
1575 // Emit means "dump code bits into code buffer".
1576 for (uint j = 0; j<last_inst; j++) {
1577 _index = j;
1578
1579 // Get the node
1580 Node* n = block->get_node(j);
1581
1582 // If this starts a new instruction group, then flush the current one
1583 // (but allow split bundles)
1584 if (Pipeline::requires_bundling() && starts_bundle(n))
1585 masm->code()->flush_bundle(false);
1586
1587 // Special handling for SafePoint/Call Nodes
1588 bool is_mcall = false;
1589 if (n->is_Mach()) {
1590 MachNode *mach = n->as_Mach();
1591 is_mcall = n->is_MachCall();
1592 bool is_sfn = n->is_MachSafePoint();
1593
1594 // If this requires all previous instructions be flushed, then do so
1595 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1596 masm->code()->flush_bundle(true);
1597 current_offset = masm->offset();
1598 }
1599
1600 // align the instruction if necessary
1601 int padding = mach->compute_padding(current_offset);
1602 // Make sure safepoint node for polling is distinct from a call's
1603 // return by adding a nop if needed.
1604 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1605 padding = nop_size;
1606 }
1607 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1608 current_offset == last_avoid_back_to_back_offset) {
1609 // Avoid back to back some instructions.
1610 padding = nop_size;
1611 }
1612
1613 if (padding > 0) {
1614 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1615 int nops_cnt = padding / nop_size;
1616 MachNode *nop = new MachNopNode(nops_cnt);
1617 block->insert_node(nop, j++);
1618 last_inst++;
1619 C->cfg()->map_node_to_block(nop, block);
1620 // Ensure enough space.
1621 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1622 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1623 C->record_failure("CodeCache is full");
1624 return;
1625 }
1626 nop->emit(masm, C->regalloc());
1627 masm->code()->flush_bundle(true);
1628 current_offset = masm->offset();
1629 }
1630
1631 bool observe_safepoint = is_sfn;
1632 // Remember the start of the last call in a basic block
1633 if (is_mcall) {
1634 MachCallNode *mcall = mach->as_MachCall();
1635
1636 if (mcall->entry_point() != nullptr) {
1637 // This destination address is NOT PC-relative
1638 mcall->method_set((intptr_t)mcall->entry_point());
1639 }
1640
1641 // Save the return address
1642 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1643
1644 observe_safepoint = mcall->guaranteed_safepoint();
1645 }
1646
1647 // sfn will be valid whenever mcall is valid now because of inheritance
1648 if (observe_safepoint) {
1649 // Handle special safepoint nodes for synchronization
1650 if (!is_mcall) {
1651 MachSafePointNode *sfn = mach->as_MachSafePoint();
1652 // !!!!! Stubs only need an oopmap right now, so bail out
1653 if (sfn->jvms()->method() == nullptr) {
1654 // Write the oopmap directly to the code blob??!!
1655 continue;
1656 }
1657 } // End synchronization
1658
1659 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1660 current_offset);
1661 Process_OopMap_Node(mach, current_offset);
1662 } // End if safepoint
1663
1664 // If this is a null check, then add the start of the previous instruction to the list
1665 else if( mach->is_MachNullCheck() ) {
1666 inct_starts[inct_cnt++] = previous_offset;
1667 }
1668
1669 // If this is a branch, then fill in the label with the target BB's label
1670 else if (mach->is_MachBranch()) {
1671 // This requires the TRUE branch target be in succs[0]
1672 uint block_num = block->non_connector_successor(0)->_pre_order;
1673
1674 // Try to replace long branch,
1675 // it is mostly for back branches since forward branch's
1676 // distance is not updated yet.
1677 if (mach->may_be_short_branch()) {
1678 int br_size = n->size(C->regalloc());
1679 int offset = blk_starts[block_num] - current_offset;
1680 if (block_num >= i) {
1681 // Current and following block's offset are not
1682 // finalized yet, adjust distance by the difference
1683 // between calculated and final offsets of current block.
1684 offset -= (blk_starts[i] - blk_offset);
1685 }
1686 // In the following code a nop could be inserted before
1687 // the branch which will increase the backward distance.
1688 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1689 if (needs_padding && offset <= 0)
1690 offset -= nop_size;
1691
1692 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1693 // We've got a winner. Replace this branch.
1694 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1695
1696 // Update the jmp_size.
1697 int new_size = replacement->size(C->regalloc());
1698 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1699 // Insert padding between avoid_back_to_back branches.
1700 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1701 MachNode *nop = new MachNopNode();
1702 block->insert_node(nop, j++);
1703 C->cfg()->map_node_to_block(nop, block);
1704 last_inst++;
1705 nop->emit(masm, C->regalloc());
1706 masm->code()->flush_bundle(true);
1707 current_offset = masm->offset();
1708 }
1709 #ifdef ASSERT
1710 jmp_target[i] = block_num;
1711 jmp_offset[i] = current_offset - blk_offset;
1712 jmp_size[i] = new_size;
1713 jmp_rule[i] = mach->rule();
1714 #endif
1715 block->map_node(replacement, j);
1716 mach->subsume_by(replacement, C);
1717 n = replacement;
1718 mach = replacement;
1719 }
1720 }
1721 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1722 } else if (mach->ideal_Opcode() == Op_Jump) {
1723 for (uint h = 0; h < block->_num_succs; h++) {
1724 Block* succs_block = block->_succs[h];
1725 for (uint j = 1; j < succs_block->num_preds(); j++) {
1726 Node* jpn = succs_block->pred(j);
1727 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1728 uint block_num = succs_block->non_connector()->_pre_order;
1729 Label *blkLabel = &blk_labels[block_num];
1730 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1731 }
1732 }
1733 }
1734 } else if (!n->is_Proj()) {
1735 // Remember the beginning of the previous instruction, in case
1736 // it's followed by a flag-kill and a null-check. Happens on
1737 // Intel all the time, with add-to-memory kind of opcodes.
1738 previous_offset = current_offset;
1739 }
1740
1741 // Not an else-if!
1742 // If this is a trap based cmp then add its offset to the list.
1743 if (mach->is_TrapBasedCheckNode()) {
1744 inct_starts[inct_cnt++] = current_offset;
1745 }
1746 }
1747
1748 // Verify that there is sufficient space remaining
1749 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1750 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1751 C->record_failure("CodeCache is full");
1752 return;
1753 }
1754
1755 // Save the offset for the listing
1756 #if defined(SUPPORT_OPTO_ASSEMBLY)
1757 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1758 node_offsets[n->_idx] = masm->offset();
1759 }
1760 #endif
1761 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1762
1763 // "Normal" instruction case
1764 DEBUG_ONLY(uint instr_offset = masm->offset());
1765 n->emit(masm, C->regalloc());
1766 current_offset = masm->offset();
1767
1768 // Above we only verified that there is enough space in the instruction section.
1769 // However, the instruction may emit stubs that cause code buffer expansion.
1770 // Bail out here if expansion failed due to a lack of code cache space.
1771 if (C->failing()) {
1772 return;
1773 }
1774
1775 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1776 "ret_addr_offset() not within emitted code");
1777 #ifdef ASSERT
1778 uint n_size = n->size(C->regalloc());
1779 if (n_size < (current_offset-instr_offset)) {
1780 MachNode* mach = n->as_Mach();
1781 n->dump();
1782 mach->dump_format(C->regalloc(), tty);
1783 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1784 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1785 tty->print_cr(" ------------------- ");
1786 BufferBlob* blob = this->scratch_buffer_blob();
1787 address blob_begin = blob->content_begin();
1788 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1789 assert(false, "wrong size of mach node");
1790 }
1791 #endif
1792 non_safepoints.observe_instruction(n, current_offset);
1793
1794 // mcall is last "call" that can be a safepoint
1795 // record it so we can see if a poll will directly follow it
1796 // in which case we'll need a pad to make the PcDesc sites unique
1797 // see 5010568. This can be slightly inaccurate but conservative
1798 // in the case that return address is not actually at current_offset.
1799 // This is a small price to pay.
1800
1801 if (is_mcall) {
1802 last_call_offset = current_offset;
1803 }
1804
1805 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1806 // Avoid back to back some instructions.
1807 last_avoid_back_to_back_offset = current_offset;
1808 }
1809
1810 } // End for all instructions in block
1811
1812 // If the next block is the top of a loop, pad this block out to align
1813 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1814 if (i < nblocks-1) {
1815 Block *nb = C->cfg()->get_block(i + 1);
1816 int padding = nb->alignment_padding(current_offset);
1817 if( padding > 0 ) {
1818 MachNode *nop = new MachNopNode(padding / nop_size);
1819 block->insert_node(nop, block->number_of_nodes());
1820 C->cfg()->map_node_to_block(nop, block);
1821 nop->emit(masm, C->regalloc());
1822 current_offset = masm->offset();
1823 }
1824 }
1825 // Verify that the distance for generated before forward
1826 // short branches is still valid.
1827 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1828
1829 // Save new block start offset
1830 blk_starts[i] = blk_offset;
1831 } // End of for all blocks
1832 blk_starts[nblocks] = current_offset;
1833
1834 non_safepoints.flush_at_end();
1835
1836 // Offset too large?
1837 if (C->failing()) return;
1838
1839 // Define a pseudo-label at the end of the code
1840 masm->bind( blk_labels[nblocks] );
1841
1842 // Compute the size of the first block
1843 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1844
1845 #ifdef ASSERT
1846 for (uint i = 0; i < nblocks; i++) { // For all blocks
1847 if (jmp_target[i] != 0) {
1848 int br_size = jmp_size[i];
1849 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1850 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1851 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]);
1852 assert(false, "Displacement too large for short jmp");
1853 }
1854 }
1855 }
1856 #endif
1857
1858 if (!masm->code()->finalize_stubs()) {
1859 C->record_failure("CodeCache is full");
1860 return;
1861 }
1862
1863 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1864 bs->emit_stubs(*masm->code());
1865 if (C->failing()) return;
1866
1867 // Fill in stubs.
1868 assert(masm->inst_mark() == nullptr, "should be.");
1869 _stub_list.emit(*masm);
1870 if (C->failing()) return;
1871
1872 #ifndef PRODUCT
1873 // Information on the size of the method, without the extraneous code
1874 Scheduling::increment_method_size(masm->offset());
1875 #endif
1876
1877 // ------------------
1878 // Fill in exception table entries.
1879 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1880
1881 // Only java methods have exception handlers and deopt handlers
1882 // class HandlerImpl is platform-specific and defined in the *.ad files.
1883 if (C->method()) {
1884 if (C->failing()) {
1885 return; // CodeBuffer::expand failed
1886 }
1887 // Emit the deopt handler code.
1888 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1889 }
1890
1891 // One last check for failed CodeBuffer::expand:
1892 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1893 C->record_failure("CodeCache is full");
1894 return;
1895 }
1896
1897 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1898 if (C->print_assembly()) {
1899 tty->cr();
1900 tty->print_cr("============================= C2-compiled nmethod ==============================");
1901 }
1902 #endif
1903
1904 #if defined(SUPPORT_OPTO_ASSEMBLY)
1905 // Dump the assembly code, including basic-block numbers
1906 if (C->print_assembly()) {
1907 ttyLocker ttyl; // keep the following output all in one block
1908 if (!VMThread::should_terminate()) { // test this under the tty lock
1909 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1910 // We call them first and write to a stringStream, then we retake the lock to
1911 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1912 ResourceMark rm;
1913 stringStream method_metadata_str;
1914 if (C->method() != nullptr) {
1915 C->method()->print_metadata(&method_metadata_str);
1916 }
1917 stringStream dump_asm_str;
1918 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1919
1920 NoSafepointVerifier nsv;
1921 ttyLocker ttyl2;
1922 // This output goes directly to the tty, not the compiler log.
1923 // To enable tools to match it up with the compilation activity,
1924 // be sure to tag this tty output with the compile ID.
1925 if (xtty != nullptr) {
1926 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1927 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1928 }
1929 if (C->method() != nullptr) {
1930 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1931 tty->print_raw(method_metadata_str.freeze());
1932 } else if (C->stub_name() != nullptr) {
1933 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1934 }
1935 tty->cr();
1936 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1937 tty->print_raw(dump_asm_str.freeze());
1938 tty->print_cr("--------------------------------------------------------------------------------");
1939 if (xtty != nullptr) {
1940 xtty->tail("opto_assembly");
1941 }
1942 }
1943 }
1944 #endif
1945 }
1946
1947 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1948 _inc_table.set_size(cnt);
1949
1950 uint inct_cnt = 0;
1951 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1952 Block* block = C->cfg()->get_block(i);
1953 Node *n = nullptr;
1954 int j;
1955
1956 // Find the branch; ignore trailing NOPs.
1957 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1958 n = block->get_node(j);
1959 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1960 break;
1961 }
1962 }
1963
1964 // If we didn't find anything, continue
1965 if (j < 0) {
1966 continue;
1967 }
1968
1969 // Compute ExceptionHandlerTable subtable entry and add it
1970 // (skip empty blocks)
1971 if (n->is_Catch()) {
1972
1973 // Get the offset of the return from the call
1974 uint call_return = call_returns[block->_pre_order];
1975 #ifdef ASSERT
1976 assert( call_return > 0, "no call seen for this basic block" );
1977 while (block->get_node(--j)->is_MachProj()) ;
1978 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1979 #endif
1980 // last instruction is a CatchNode, find it's CatchProjNodes
1981 int nof_succs = block->_num_succs;
1982 // allocate space
1983 GrowableArray<intptr_t> handler_bcis(nof_succs);
1984 GrowableArray<intptr_t> handler_pcos(nof_succs);
1985 // iterate through all successors
1986 for (int j = 0; j < nof_succs; j++) {
1987 Block* s = block->_succs[j];
1988 bool found_p = false;
1989 for (uint k = 1; k < s->num_preds(); k++) {
1990 Node* pk = s->pred(k);
1991 if (pk->is_CatchProj() && pk->in(0) == n) {
1992 const CatchProjNode* p = pk->as_CatchProj();
1993 found_p = true;
1994 // add the corresponding handler bci & pco information
1995 if (p->_con != CatchProjNode::fall_through_index) {
1996 // p leads to an exception handler (and is not fall through)
1997 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1998 // no duplicates, please
1999 if (!handler_bcis.contains(p->handler_bci())) {
2000 uint block_num = s->non_connector()->_pre_order;
2001 handler_bcis.append(p->handler_bci());
2002 handler_pcos.append(blk_labels[block_num].loc_pos());
2003 }
2004 }
2005 }
2006 }
2007 assert(found_p, "no matching predecessor found");
2008 // Note: Due to empty block removal, one block may have
2009 // several CatchProj inputs, from the same Catch.
2010 }
2011
2012 // Set the offset of the return from the call
2013 assert(handler_bcis.find(-1) != -1, "must have default handler");
2014 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2015 continue;
2016 }
2017
2018 // Handle implicit null exception table updates
2019 if (n->is_MachNullCheck()) {
2020 MachNode* access = n->in(1)->as_Mach();
2021 assert(access->barrier_data() == 0 ||
2022 access->is_late_expanded_null_check_candidate(),
2023 "Implicit null checks on memory accesses with barriers are only supported on nodes explicitly marked as null-check candidates");
2024 uint block_num = block->non_connector_successor(0)->_pre_order;
2025 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2026 continue;
2027 }
2028 // Handle implicit exception table updates: trap instructions.
2029 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2030 uint block_num = block->non_connector_successor(0)->_pre_order;
2031 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2032 continue;
2033 }
2034 } // End of for all blocks fill in exception table entries
2035 }
2036
2037 // Static Variables
2038 #ifndef PRODUCT
2039 uint Scheduling::_total_nop_size = 0;
2040 uint Scheduling::_total_method_size = 0;
2041 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2042 #endif
2043
2044 // Initializer for class Scheduling
2045
2046 Scheduling::Scheduling(Arena *arena, Compile &compile)
2047 : _arena(arena),
2048 _cfg(compile.cfg()),
2049 _regalloc(compile.regalloc()),
2050 _scheduled(arena),
2051 _available(arena),
2052 _reg_node(arena),
2053 _pinch_free_list(arena),
2054 _next_node(nullptr),
2055 _bundle_instr_count(0),
2056 _bundle_cycle_number(0),
2057 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2058 {
2059 // Save the count
2060 _node_bundling_limit = compile.unique();
2061 uint node_max = _regalloc->node_regs_max_index();
2062
2063 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2064
2065 // This one is persistent within the Compile class
2066 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2067
2068 // Allocate space for fixed-size arrays
2069 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2070 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2071
2072 // Clear the arrays
2073 for (uint i = 0; i < node_max; i++) {
2074 ::new (&_node_bundling_base[i]) Bundle();
2075 }
2076 memset(_uses, 0, node_max * sizeof(short));
2077 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2078
2079 // Clear the bundling information
2080 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2081
2082 // Get the last node
2083 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2084
2085 _next_node = block->get_node(block->number_of_nodes() - 1);
2086 }
2087
2088 // Step ahead "i" cycles
2089 void Scheduling::step(uint i) {
2090
2091 Bundle *bundle = node_bundling(_next_node);
2092 bundle->set_starts_bundle();
2093
2094 // Update the bundle record, but leave the flags information alone
2095 if (_bundle_instr_count > 0) {
2096 bundle->set_instr_count(_bundle_instr_count);
2097 bundle->set_resources_used(_bundle_use.resourcesUsed());
2098 }
2099
2100 // Update the state information
2101 _bundle_instr_count = 0;
2102 _bundle_cycle_number += i;
2103 _bundle_use.step(i);
2104 }
2105
2106 void Scheduling::step_and_clear() {
2107 Bundle *bundle = node_bundling(_next_node);
2108 bundle->set_starts_bundle();
2109
2110 // Update the bundle record
2111 if (_bundle_instr_count > 0) {
2112 bundle->set_instr_count(_bundle_instr_count);
2113 bundle->set_resources_used(_bundle_use.resourcesUsed());
2114
2115 _bundle_cycle_number += 1;
2116 }
2117
2118 // Clear the bundling information
2119 _bundle_instr_count = 0;
2120 _bundle_use.reset();
2121
2122 memcpy(_bundle_use_elements,
2123 Pipeline_Use::elaborated_elements,
2124 sizeof(Pipeline_Use::elaborated_elements));
2125 }
2126
2127 // Perform instruction scheduling and bundling over the sequence of
2128 // instructions in backwards order.
2129 void PhaseOutput::ScheduleAndBundle() {
2130
2131 // Don't optimize this if it isn't a method
2132 if (!C->method())
2133 return;
2134
2135 // Don't optimize this if scheduling is disabled
2136 if (!C->do_scheduling())
2137 return;
2138
2139 // Scheduling code works only with pairs (8 bytes) maximum.
2140 // And when the scalable vector register is used, we may spill/unspill
2141 // the whole reg regardless of the max vector size.
2142 if (C->max_vector_size() > 8 ||
2143 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2144 return;
2145 }
2146
2147 Compile::TracePhase tp(_t_instrSched);
2148
2149 // Create a data structure for all the scheduling information
2150 Scheduling scheduling(Thread::current()->resource_area(), *C);
2151
2152 // Walk backwards over each basic block, computing the needed alignment
2153 // Walk over all the basic blocks
2154 scheduling.DoScheduling();
2155
2156 #ifndef PRODUCT
2157 if (C->trace_opto_output()) {
2158 // Buffer and print all at once
2159 ResourceMark rm;
2160 stringStream ss;
2161 ss.print("\n---- After ScheduleAndBundle ----\n");
2162 print_scheduling(&ss);
2163 tty->print("%s", ss.as_string());
2164 }
2165 #endif
2166 }
2167
2168 #ifndef PRODUCT
2169 // Separated out so that it can be called directly from debugger
2170 void PhaseOutput::print_scheduling() {
2171 print_scheduling(tty);
2172 }
2173
2174 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2175 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2176 output_stream->print("\nBB#%03d:\n", i);
2177 Block* block = C->cfg()->get_block(i);
2178 for (uint j = 0; j < block->number_of_nodes(); j++) {
2179 Node* n = block->get_node(j);
2180 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2181 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2182 n->dump("\n", false, output_stream);
2183 }
2184 }
2185 }
2186 #endif
2187
2188 // See if this node fits into the present instruction bundle
2189 bool Scheduling::NodeFitsInBundle(Node *n) {
2190 uint n_idx = n->_idx;
2191
2192 // If the node cannot be scheduled this cycle, skip it
2193 if (_current_latency[n_idx] > _bundle_cycle_number) {
2194 #ifndef PRODUCT
2195 if (_cfg->C->trace_opto_output())
2196 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2197 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2198 #endif
2199 return (false);
2200 }
2201
2202 const Pipeline *node_pipeline = n->pipeline();
2203
2204 uint instruction_count = node_pipeline->instructionCount();
2205 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2206 instruction_count = 0;
2207
2208 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2209 #ifndef PRODUCT
2210 if (_cfg->C->trace_opto_output())
2211 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2212 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2213 #endif
2214 return (false);
2215 }
2216
2217 // Don't allow non-machine nodes to be handled this way
2218 if (!n->is_Mach() && instruction_count == 0)
2219 return (false);
2220
2221 // See if there is any overlap
2222 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2223
2224 if (delay > 0) {
2225 #ifndef PRODUCT
2226 if (_cfg->C->trace_opto_output())
2227 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2228 #endif
2229 return false;
2230 }
2231
2232 #ifndef PRODUCT
2233 if (_cfg->C->trace_opto_output())
2234 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2235 #endif
2236
2237 return true;
2238 }
2239
2240 Node * Scheduling::ChooseNodeToBundle() {
2241 uint siz = _available.size();
2242
2243 if (siz == 0) {
2244
2245 #ifndef PRODUCT
2246 if (_cfg->C->trace_opto_output())
2247 tty->print("# ChooseNodeToBundle: null\n");
2248 #endif
2249 return (nullptr);
2250 }
2251
2252 // Fast path, if only 1 instruction in the bundle
2253 if (siz == 1) {
2254 #ifndef PRODUCT
2255 if (_cfg->C->trace_opto_output()) {
2256 tty->print("# ChooseNodeToBundle (only 1): ");
2257 _available[0]->dump();
2258 }
2259 #endif
2260 return (_available[0]);
2261 }
2262
2263 // Don't bother, if the bundle is already full
2264 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2265 for ( uint i = 0; i < siz; i++ ) {
2266 Node *n = _available[i];
2267
2268 // Skip projections, we'll handle them another way
2269 if (n->is_Proj())
2270 continue;
2271
2272 // This presupposed that instructions are inserted into the
2273 // available list in a legality order; i.e. instructions that
2274 // must be inserted first are at the head of the list
2275 if (NodeFitsInBundle(n)) {
2276 #ifndef PRODUCT
2277 if (_cfg->C->trace_opto_output()) {
2278 tty->print("# ChooseNodeToBundle: ");
2279 n->dump();
2280 }
2281 #endif
2282 return (n);
2283 }
2284 }
2285 }
2286
2287 // Nothing fits in this bundle, choose the highest priority
2288 #ifndef PRODUCT
2289 if (_cfg->C->trace_opto_output()) {
2290 tty->print("# ChooseNodeToBundle: ");
2291 _available[0]->dump();
2292 }
2293 #endif
2294
2295 return _available[0];
2296 }
2297
2298 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2299 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2300
2301 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2302 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2303 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2304 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2305
2306 // Comparison between stack -> reg and stack -> reg
2307 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2308 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2309 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2310 }
2311
2312 // Comparison between reg -> stack and reg -> stack
2313 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2314 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2315 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2316 }
2317
2318 return 0; // Not comparable
2319 }
2320
2321 void Scheduling::AddNodeToAvailableList(Node *n) {
2322 assert( !n->is_Proj(), "projections never directly made available" );
2323 #ifndef PRODUCT
2324 if (_cfg->C->trace_opto_output()) {
2325 tty->print("# AddNodeToAvailableList: ");
2326 n->dump();
2327 }
2328 #endif
2329
2330 int latency = _current_latency[n->_idx];
2331
2332 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2333 // for stack spilling or unspilling have the same latency, we sort
2334 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2335 // have more opportunities to do ld/st merging
2336 uint i;
2337 for (i = 0; i < _available.size(); i++) {
2338 if (_current_latency[_available[i]->_idx] > latency) {
2339 break;
2340 } else if (_current_latency[_available[i]->_idx] == latency &&
2341 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2342 compare_two_spill_nodes(n, _available[i]) > 0) {
2343 break;
2344 }
2345 }
2346
2347 // Special Check for compares following branches
2348 if( n->is_Mach() && _scheduled.size() > 0 ) {
2349 int op = n->as_Mach()->ideal_Opcode();
2350 Node *last = _scheduled[0];
2351 if( last->is_MachIf() && last->in(1) == n &&
2352 ( op == Op_CmpI ||
2353 op == Op_CmpU ||
2354 op == Op_CmpUL ||
2355 op == Op_CmpP ||
2356 op == Op_CmpF ||
2357 op == Op_CmpD ||
2358 op == Op_CmpL ) ) {
2359
2360 // Recalculate position, moving to front of same latency
2361 for ( i=0 ; i < _available.size(); i++ )
2362 if (_current_latency[_available[i]->_idx] >= latency)
2363 break;
2364 }
2365 }
2366
2367 // Insert the node in the available list
2368 _available.insert(i, n);
2369
2370 #ifndef PRODUCT
2371 if (_cfg->C->trace_opto_output())
2372 dump_available();
2373 #endif
2374 }
2375
2376 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2377 for ( uint i=0; i < n->len(); i++ ) {
2378 Node *def = n->in(i);
2379 if (!def) continue;
2380 if( def->is_Proj() ) // If this is a machine projection, then
2381 def = def->in(0); // propagate usage thru to the base instruction
2382
2383 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2384 continue;
2385 }
2386
2387 // Compute the latency
2388 uint l = _bundle_cycle_number + n->latency(i);
2389 if (_current_latency[def->_idx] < l)
2390 _current_latency[def->_idx] = l;
2391
2392 // If this does not have uses then schedule it
2393 if ((--_uses[def->_idx]) == 0)
2394 AddNodeToAvailableList(def);
2395 }
2396 }
2397
2398 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2399 #ifndef PRODUCT
2400 if (_cfg->C->trace_opto_output()) {
2401 tty->print("# AddNodeToBundle: ");
2402 n->dump();
2403 }
2404 #endif
2405
2406 // Remove this from the available list
2407 uint i;
2408 for (i = 0; i < _available.size(); i++)
2409 if (_available[i] == n)
2410 break;
2411 assert(i < _available.size(), "entry in _available list not found");
2412 _available.remove(i);
2413
2414 // See if this fits in the current bundle
2415 const Pipeline *node_pipeline = n->pipeline();
2416 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2417
2418
2419 // Get the number of instructions
2420 uint instruction_count = node_pipeline->instructionCount();
2421 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2422 instruction_count = 0;
2423
2424 // Compute the latency information
2425 uint delay = 0;
2426
2427 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2428 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2429 if (relative_latency < 0)
2430 relative_latency = 0;
2431
2432 delay = _bundle_use.full_latency(relative_latency, node_usage);
2433
2434 // Does not fit in this bundle, start a new one
2435 if (delay > 0) {
2436 step(delay);
2437
2438 #ifndef PRODUCT
2439 if (_cfg->C->trace_opto_output())
2440 tty->print("# *** STEP(%d) ***\n", delay);
2441 #endif
2442 }
2443 }
2444
2445 if (delay == 0) {
2446 if (node_pipeline->hasMultipleBundles()) {
2447 #ifndef PRODUCT
2448 if (_cfg->C->trace_opto_output())
2449 tty->print("# *** STEP(multiple instructions) ***\n");
2450 #endif
2451 step(1);
2452 }
2453
2454 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2455 #ifndef PRODUCT
2456 if (_cfg->C->trace_opto_output())
2457 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2458 instruction_count + _bundle_instr_count,
2459 Pipeline::_max_instrs_per_cycle);
2460 #endif
2461 step(1);
2462 }
2463 }
2464
2465 // Set the node's latency
2466 _current_latency[n->_idx] = _bundle_cycle_number;
2467
2468 // Now merge the functional unit information
2469 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2470 _bundle_use.add_usage(node_usage);
2471
2472 // Increment the number of instructions in this bundle
2473 _bundle_instr_count += instruction_count;
2474
2475 // Remember this node for later
2476 if (n->is_Mach())
2477 _next_node = n;
2478
2479 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2480 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2481 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2482 // into the block. All other scheduled nodes get put in the schedule here.
2483 int op = n->Opcode();
2484 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2485 (op != Op_Node && // Not an unused antidepedence node and
2486 // not an unallocated boxlock
2487 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2488
2489 // Push any trailing projections
2490 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2491 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2492 Node *foi = n->fast_out(i);
2493 if( foi->is_Proj() )
2494 _scheduled.push(foi);
2495 }
2496 }
2497
2498 // Put the instruction in the schedule list
2499 _scheduled.push(n);
2500 }
2501
2502 #ifndef PRODUCT
2503 if (_cfg->C->trace_opto_output())
2504 dump_available();
2505 #endif
2506
2507 // Walk all the definitions, decrementing use counts, and
2508 // if a definition has a 0 use count, place it in the available list.
2509 DecrementUseCounts(n,bb);
2510 }
2511
2512 // This method sets the use count within a basic block. We will ignore all
2513 // uses outside the current basic block. As we are doing a backwards walk,
2514 // any node we reach that has a use count of 0 may be scheduled. This also
2515 // avoids the problem of cyclic references from phi nodes, as long as phi
2516 // nodes are at the front of the basic block. This method also initializes
2517 // the available list to the set of instructions that have no uses within this
2518 // basic block.
2519 void Scheduling::ComputeUseCount(const Block *bb) {
2520 #ifndef PRODUCT
2521 if (_cfg->C->trace_opto_output())
2522 tty->print("# -> ComputeUseCount\n");
2523 #endif
2524
2525 // Clear the list of available and scheduled instructions, just in case
2526 _available.clear();
2527 _scheduled.clear();
2528
2529 #ifdef ASSERT
2530 for( uint i=0; i < bb->number_of_nodes(); i++ )
2531 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2532 #endif
2533
2534 // Force the _uses count to never go to zero for unscheduable pieces
2535 // of the block
2536 for( uint k = 0; k < _bb_start; k++ )
2537 _uses[bb->get_node(k)->_idx] = 1;
2538 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2539 _uses[bb->get_node(l)->_idx] = 1;
2540
2541 // Iterate backwards over the instructions in the block. Don't count the
2542 // branch projections at end or the block header instructions.
2543 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2544 Node *n = bb->get_node(j);
2545 if( n->is_Proj() ) continue; // Projections handled another way
2546
2547 // Account for all uses
2548 for ( uint k = 0; k < n->len(); k++ ) {
2549 Node *inp = n->in(k);
2550 if (!inp) continue;
2551 assert(inp != n, "no cycles allowed" );
2552 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2553 if (inp->is_Proj()) { // Skip through Proj's
2554 inp = inp->in(0);
2555 }
2556 ++_uses[inp->_idx]; // Count 1 block-local use
2557 }
2558 }
2559
2560 // If this instruction has a 0 use count, then it is available
2561 if (!_uses[n->_idx]) {
2562 _current_latency[n->_idx] = _bundle_cycle_number;
2563 AddNodeToAvailableList(n);
2564 }
2565
2566 #ifndef PRODUCT
2567 if (_cfg->C->trace_opto_output()) {
2568 tty->print("# uses: %3d: ", _uses[n->_idx]);
2569 n->dump();
2570 }
2571 #endif
2572 }
2573
2574 #ifndef PRODUCT
2575 if (_cfg->C->trace_opto_output())
2576 tty->print("# <- ComputeUseCount\n");
2577 #endif
2578 }
2579
2580 // This routine performs scheduling on each basic block in reverse order,
2581 // using instruction latencies and taking into account function unit
2582 // availability.
2583 void Scheduling::DoScheduling() {
2584 #ifndef PRODUCT
2585 if (_cfg->C->trace_opto_output())
2586 tty->print("# -> DoScheduling\n");
2587 #endif
2588
2589 Block *succ_bb = nullptr;
2590 Block *bb;
2591 Compile* C = Compile::current();
2592
2593 // Walk over all the basic blocks in reverse order
2594 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2595 bb = _cfg->get_block(i);
2596
2597 #ifndef PRODUCT
2598 if (_cfg->C->trace_opto_output()) {
2599 tty->print("# Schedule BB#%03d (initial)\n", i);
2600 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2601 bb->get_node(j)->dump();
2602 }
2603 }
2604 #endif
2605
2606 // On the head node, skip processing
2607 if (bb == _cfg->get_root_block()) {
2608 continue;
2609 }
2610
2611 // Skip empty, connector blocks
2612 if (bb->is_connector())
2613 continue;
2614
2615 // If the following block is not the sole successor of
2616 // this one, then reset the pipeline information
2617 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2618 #ifndef PRODUCT
2619 if (_cfg->C->trace_opto_output()) {
2620 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2621 _next_node->_idx, _bundle_instr_count);
2622 }
2623 #endif
2624 step_and_clear();
2625 }
2626
2627 // Leave untouched the starting instruction, any Phis, a CreateEx node
2628 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2629 _bb_end = bb->number_of_nodes()-1;
2630 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2631 Node *n = bb->get_node(_bb_start);
2632 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2633 // Also, MachIdealNodes do not get scheduled
2634 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2635 MachNode *mach = n->as_Mach();
2636 int iop = mach->ideal_Opcode();
2637 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2638 if( iop == Op_Con ) continue; // Do not schedule Top
2639 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2640 mach->pipeline() == MachNode::pipeline_class() &&
2641 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2642 continue;
2643 break; // Funny loop structure to be sure...
2644 }
2645 // Compute last "interesting" instruction in block - last instruction we
2646 // might schedule. _bb_end points just after last schedulable inst.
2647 Node *last = bb->get_node(_bb_end);
2648 // Ignore trailing NOPs.
2649 while (_bb_end > 0 && last->is_Mach() &&
2650 last->as_Mach()->ideal_Opcode() == Op_Con) {
2651 last = bb->get_node(--_bb_end);
2652 }
2653 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2654 if( last->is_Catch() ||
2655 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2656 // There might be a prior call. Skip it.
2657 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2658 } else if( last->is_MachNullCheck() ) {
2659 // Backup so the last null-checked memory instruction is
2660 // outside the schedulable range. Skip over the nullcheck,
2661 // projection, and the memory nodes.
2662 Node *mem = last->in(1);
2663 do {
2664 _bb_end--;
2665 } while (mem != bb->get_node(_bb_end));
2666 } else {
2667 // Set _bb_end to point after last schedulable inst.
2668 _bb_end++;
2669 }
2670
2671 assert( _bb_start <= _bb_end, "inverted block ends" );
2672
2673 // Compute the register antidependencies for the basic block
2674 ComputeRegisterAntidependencies(bb);
2675 if (C->failing()) return; // too many D-U pinch points
2676
2677 // Compute the usage within the block, and set the list of all nodes
2678 // in the block that have no uses within the block.
2679 ComputeUseCount(bb);
2680
2681 // Schedule the remaining instructions in the block
2682 while ( _available.size() > 0 ) {
2683 Node *n = ChooseNodeToBundle();
2684 guarantee(n != nullptr, "no nodes available");
2685 AddNodeToBundle(n,bb);
2686 }
2687
2688 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2689 #ifdef ASSERT
2690 for( uint l = _bb_start; l < _bb_end; l++ ) {
2691 Node *n = bb->get_node(l);
2692 uint m;
2693 for( m = 0; m < _bb_end-_bb_start; m++ )
2694 if( _scheduled[m] == n )
2695 break;
2696 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2697 }
2698 #endif
2699
2700 // Now copy the instructions (in reverse order) back to the block
2701 for ( uint k = _bb_start; k < _bb_end; k++ )
2702 bb->map_node(_scheduled[_bb_end-k-1], k);
2703
2704 #ifndef PRODUCT
2705 if (_cfg->C->trace_opto_output()) {
2706 tty->print("# Schedule BB#%03d (final)\n", i);
2707 uint current = 0;
2708 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2709 Node *n = bb->get_node(j);
2710 if( valid_bundle_info(n) ) {
2711 Bundle *bundle = node_bundling(n);
2712 if (bundle->instr_count() > 0) {
2713 tty->print("*** Bundle: ");
2714 bundle->dump();
2715 }
2716 n->dump();
2717 }
2718 }
2719 }
2720 #endif
2721 #ifdef ASSERT
2722 verify_good_schedule(bb,"after block local scheduling");
2723 #endif
2724 }
2725
2726 #ifndef PRODUCT
2727 if (_cfg->C->trace_opto_output())
2728 tty->print("# <- DoScheduling\n");
2729 #endif
2730
2731 // Record final node-bundling array location
2732 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2733
2734 } // end DoScheduling
2735
2736 // Verify that no live-range used in the block is killed in the block by a
2737 // wrong DEF. This doesn't verify live-ranges that span blocks.
2738
2739 // Check for edge existence. Used to avoid adding redundant precedence edges.
2740 static bool edge_from_to( Node *from, Node *to ) {
2741 for( uint i=0; i<from->len(); i++ )
2742 if( from->in(i) == to )
2743 return true;
2744 return false;
2745 }
2746
2747 #ifdef ASSERT
2748 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2749 // Check for bad kills
2750 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2751 Node *prior_use = _reg_node[def];
2752 if( prior_use && !edge_from_to(prior_use,n) ) {
2753 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2754 n->dump();
2755 tty->print_cr("...");
2756 prior_use->dump();
2757 assert(edge_from_to(prior_use,n), "%s", msg);
2758 }
2759 _reg_node.map(def,nullptr); // Kill live USEs
2760 }
2761 }
2762
2763 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2764
2765 // Zap to something reasonable for the verify code
2766 _reg_node.clear();
2767
2768 // Walk over the block backwards. Check to make sure each DEF doesn't
2769 // kill a live value (other than the one it's supposed to). Add each
2770 // USE to the live set.
2771 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2772 Node *n = b->get_node(i);
2773 int n_op = n->Opcode();
2774 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2775 // Fat-proj kills a slew of registers
2776 RegMaskIterator rmi(n->out_RegMask());
2777 while (rmi.has_next()) {
2778 OptoReg::Name kill = rmi.next();
2779 verify_do_def(n, kill, msg);
2780 }
2781 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2782 // Get DEF'd registers the normal way
2783 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2784 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2785 }
2786
2787 // Now make all USEs live
2788 for( uint i=1; i<n->req(); i++ ) {
2789 Node *def = n->in(i);
2790 assert(def != nullptr, "input edge required");
2791 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2792 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2793 if( OptoReg::is_valid(reg_lo) ) {
2794 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2795 _reg_node.map(reg_lo,n);
2796 }
2797 if( OptoReg::is_valid(reg_hi) ) {
2798 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2799 _reg_node.map(reg_hi,n);
2800 }
2801 }
2802
2803 }
2804
2805 // Zap to something reasonable for the Antidependence code
2806 _reg_node.clear();
2807 }
2808 #endif
2809
2810 // Conditionally add precedence edges. Avoid putting edges on Projs.
2811 static void add_prec_edge_from_to( Node *from, Node *to ) {
2812 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2813 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2814 from = from->in(0);
2815 }
2816 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2817 !edge_from_to( from, to ) ) // Avoid duplicate edge
2818 from->add_prec(to);
2819 }
2820
2821 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2822 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2823 return;
2824
2825 if (OptoReg::is_reg(def_reg)) {
2826 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2827 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2828 // This is one of the high slots of a vector register.
2829 // ScheduleAndBundle already checked there are no live wide
2830 // vectors in this method so it can be safely ignored.
2831 return;
2832 }
2833 }
2834
2835 Node *pinch = _reg_node[def_reg]; // Get pinch point
2836 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2837 is_def ) { // Check for a true def (not a kill)
2838 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2839 return;
2840 }
2841
2842 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2843 DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); )
2844
2845 // After some number of kills there _may_ be a later def
2846 Node *later_def = nullptr;
2847
2848 Compile* C = Compile::current();
2849
2850 // Finding a kill requires a real pinch-point.
2851 // Check for not already having a pinch-point.
2852 // Pinch points are Op_Node's.
2853 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2854 later_def = pinch; // Must be def/kill as optimistic pinch-point
2855 if ( _pinch_free_list.size() > 0) {
2856 pinch = _pinch_free_list.pop();
2857 } else {
2858 pinch = new Node(1); // Pinch point to-be
2859 }
2860 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2861 DEBUG_ONLY( pinch->dump(); );
2862 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2863 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2864 return;
2865 }
2866 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2867 _reg_node.map(def_reg,pinch); // Record pinch-point
2868 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2869 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2870 pinch->init_req(0, C->top()); // set not null for the next call
2871 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2872 later_def = nullptr; // and no later def
2873 }
2874 pinch->set_req(0,later_def); // Hook later def so we can find it
2875 } else { // Else have valid pinch point
2876 if( pinch->in(0) ) // If there is a later-def
2877 later_def = pinch->in(0); // Get it
2878 }
2879
2880 // Add output-dependence edge from later def to kill
2881 if( later_def ) // If there is some original def
2882 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2883
2884 // See if current kill is also a use, and so is forced to be the pinch-point.
2885 if( pinch->Opcode() == Op_Node ) {
2886 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2887 for( uint i=1; i<uses->req(); i++ ) {
2888 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2889 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2890 // Yes, found a use/kill pinch-point
2891 pinch->set_req(0,nullptr); //
2892 pinch->replace_by(kill); // Move anti-dep edges up
2893 pinch = kill;
2894 _reg_node.map(def_reg,pinch);
2895 return;
2896 }
2897 }
2898 }
2899
2900 // Add edge from kill to pinch-point
2901 add_prec_edge_from_to(kill,pinch);
2902 }
2903
2904 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2905 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2906 return;
2907 Node *pinch = _reg_node[use_reg]; // Get pinch point
2908 // Check for no later def_reg/kill in block
2909 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
2910 // Use has to be block-local as well
2911 _cfg->get_block_for_node(use) == b) {
2912 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2913 pinch->req() == 1 ) { // pinch not yet in block?
2914 pinch->del_req(0); // yank pointer to later-def, also set flag
2915 // Insert the pinch-point in the block just after the last use
2916 b->insert_node(pinch, b->find_node(use) + 1);
2917 _bb_end++; // Increase size scheduled region in block
2918 }
2919
2920 add_prec_edge_from_to(pinch,use);
2921 }
2922 }
2923
2924 // We insert antidependences between the reads and following write of
2925 // allocated registers to prevent illegal code motion. Hopefully, the
2926 // number of added references should be fairly small, especially as we
2927 // are only adding references within the current basic block.
2928 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2929
2930 #ifdef ASSERT
2931 verify_good_schedule(b,"before block local scheduling");
2932 #endif
2933
2934 // A valid schedule, for each register independently, is an endless cycle
2935 // of: a def, then some uses (connected to the def by true dependencies),
2936 // then some kills (defs with no uses), finally the cycle repeats with a new
2937 // def. The uses are allowed to float relative to each other, as are the
2938 // kills. No use is allowed to slide past a kill (or def). This requires
2939 // antidependencies between all uses of a single def and all kills that
2940 // follow, up to the next def. More edges are redundant, because later defs
2941 // & kills are already serialized with true or antidependencies. To keep
2942 // the edge count down, we add a 'pinch point' node if there's more than
2943 // one use or more than one kill/def.
2944
2945 // We add dependencies in one bottom-up pass.
2946
2947 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2948
2949 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2950 // register. If not, we record the DEF/KILL in _reg_node, the
2951 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2952 // "pinch point", a new Node that's in the graph but not in the block.
2953 // We put edges from the prior and current DEF/KILLs to the pinch point.
2954 // We put the pinch point in _reg_node. If there's already a pinch point
2955 // we merely add an edge from the current DEF/KILL to the pinch point.
2956
2957 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2958 // put an edge from the pinch point to the USE.
2959
2960 // To be expedient, the _reg_node array is pre-allocated for the whole
2961 // compilation. _reg_node is lazily initialized; it either contains a null,
2962 // or a valid def/kill/pinch-point, or a leftover node from some prior
2963 // block. Leftover node from some prior block is treated like a null (no
2964 // prior def, so no anti-dependence needed). Valid def is distinguished by
2965 // it being in the current block.
2966 bool fat_proj_seen = false;
2967 uint last_safept = _bb_end-1;
2968 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
2969 Node* last_safept_node = end_node;
2970 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2971 Node *n = b->get_node(i);
2972 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2973 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2974 // Fat-proj kills a slew of registers
2975 // This can add edges to 'n' and obscure whether or not it was a def,
2976 // hence the is_def flag.
2977 fat_proj_seen = true;
2978 RegMaskIterator rmi(n->out_RegMask());
2979 while (rmi.has_next()) {
2980 OptoReg::Name kill = rmi.next();
2981 anti_do_def(b, n, kill, is_def);
2982 }
2983 } else {
2984 // Get DEF'd registers the normal way
2985 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2986 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2987 }
2988
2989 // Kill projections on a branch should appear to occur on the
2990 // branch, not afterwards, so grab the masks from the projections
2991 // and process them.
2992 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
2993 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2994 Node* use = n->fast_out(i);
2995 if (use->is_Proj()) {
2996 RegMaskIterator rmi(use->out_RegMask());
2997 while (rmi.has_next()) {
2998 OptoReg::Name kill = rmi.next();
2999 anti_do_def(b, n, kill, false);
3000 }
3001 }
3002 }
3003 }
3004
3005 // Check each register used by this instruction for a following DEF/KILL
3006 // that must occur afterward and requires an anti-dependence edge.
3007 for( uint j=0; j<n->req(); j++ ) {
3008 Node *def = n->in(j);
3009 if( def ) {
3010 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3011 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3012 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3013 }
3014 }
3015 // Do not allow defs of new derived values to float above GC
3016 // points unless the base is definitely available at the GC point.
3017
3018 Node *m = b->get_node(i);
3019
3020 // Add precedence edge from following safepoint to use of derived pointer
3021 if( last_safept_node != end_node &&
3022 m != last_safept_node) {
3023 for (uint k = 1; k < m->req(); k++) {
3024 const Type *t = m->in(k)->bottom_type();
3025 if( t->isa_oop_ptr() &&
3026 t->is_ptr()->offset() != 0 ) {
3027 last_safept_node->add_prec( m );
3028 break;
3029 }
3030 }
3031
3032 // Do not allow a CheckCastPP node whose input is a raw pointer to
3033 // float past a safepoint. This can occur when a buffered inline
3034 // type is allocated in a loop and the CheckCastPP from that
3035 // allocation is reused outside the loop. If the use inside the
3036 // loop is scalarized the CheckCastPP will no longer be connected
3037 // to the loop safepoint. See JDK-8264340.
3038 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3039 Node *def = m->in(1);
3040 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3041 last_safept_node->add_prec(m);
3042 }
3043 }
3044 }
3045
3046 if( n->jvms() ) { // Precedence edge from derived to safept
3047 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3048 if( b->get_node(last_safept) != last_safept_node ) {
3049 last_safept = b->find_node(last_safept_node);
3050 }
3051 for( uint j=last_safept; j > i; j-- ) {
3052 Node *mach = b->get_node(j);
3053 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3054 mach->add_prec( n );
3055 }
3056 last_safept = i;
3057 last_safept_node = m;
3058 }
3059 }
3060
3061 if (fat_proj_seen) {
3062 // Garbage collect pinch nodes that were not consumed.
3063 // They are usually created by a fat kill MachProj for a call.
3064 garbage_collect_pinch_nodes();
3065 }
3066 }
3067
3068 // Garbage collect pinch nodes for reuse by other blocks.
3069 //
3070 // The block scheduler's insertion of anti-dependence
3071 // edges creates many pinch nodes when the block contains
3072 // 2 or more Calls. A pinch node is used to prevent a
3073 // combinatorial explosion of edges. If a set of kills for a
3074 // register is anti-dependent on a set of uses (or defs), rather
3075 // than adding an edge in the graph between each pair of kill
3076 // and use (or def), a pinch is inserted between them:
3077 //
3078 // use1 use2 use3
3079 // \ | /
3080 // \ | /
3081 // pinch
3082 // / | \
3083 // / | \
3084 // kill1 kill2 kill3
3085 //
3086 // One pinch node is created per register killed when
3087 // the second call is encountered during a backwards pass
3088 // over the block. Most of these pinch nodes are never
3089 // wired into the graph because the register is never
3090 // used or def'ed in the block.
3091 //
3092 void Scheduling::garbage_collect_pinch_nodes() {
3093 #ifndef PRODUCT
3094 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3095 #endif
3096 int trace_cnt = 0;
3097 for (uint k = 0; k < _reg_node.max(); k++) {
3098 Node* pinch = _reg_node[k];
3099 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3100 // no predecence input edges
3101 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3102 cleanup_pinch(pinch);
3103 _pinch_free_list.push(pinch);
3104 _reg_node.map(k, nullptr);
3105 #ifndef PRODUCT
3106 if (_cfg->C->trace_opto_output()) {
3107 trace_cnt++;
3108 if (trace_cnt > 40) {
3109 tty->print("\n");
3110 trace_cnt = 0;
3111 }
3112 tty->print(" %d", pinch->_idx);
3113 }
3114 #endif
3115 }
3116 }
3117 #ifndef PRODUCT
3118 if (_cfg->C->trace_opto_output()) tty->print("\n");
3119 #endif
3120 }
3121
3122 // Clean up a pinch node for reuse.
3123 void Scheduling::cleanup_pinch( Node *pinch ) {
3124 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3125
3126 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3127 Node* use = pinch->last_out(i);
3128 uint uses_found = 0;
3129 for (uint j = use->req(); j < use->len(); j++) {
3130 if (use->in(j) == pinch) {
3131 use->rm_prec(j);
3132 uses_found++;
3133 }
3134 }
3135 assert(uses_found > 0, "must be a precedence edge");
3136 i -= uses_found; // we deleted 1 or more copies of this edge
3137 }
3138 // May have a later_def entry
3139 pinch->set_req(0, nullptr);
3140 }
3141
3142 #ifndef PRODUCT
3143
3144 void Scheduling::dump_available() const {
3145 tty->print("#Availist ");
3146 for (uint i = 0; i < _available.size(); i++)
3147 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3148 tty->cr();
3149 }
3150
3151 // Print Scheduling Statistics
3152 void Scheduling::print_statistics() {
3153 // Print the size added by nops for bundling
3154 tty->print("Nops added %d bytes to total of %d bytes",
3155 _total_nop_size, _total_method_size);
3156 if (_total_method_size > 0)
3157 tty->print(", for %.2f%%",
3158 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3159 tty->print("\n");
3160
3161 uint total_instructions = 0, total_bundles = 0;
3162
3163 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3164 uint bundle_count = _total_instructions_per_bundle[i];
3165 total_instructions += bundle_count * i;
3166 total_bundles += bundle_count;
3167 }
3168
3169 if (total_bundles > 0)
3170 tty->print("Average ILP (excluding nops) is %.2f\n",
3171 ((double)total_instructions) / ((double)total_bundles));
3172 }
3173 #endif
3174
3175 //-----------------------init_scratch_buffer_blob------------------------------
3176 // Construct a temporary BufferBlob and cache it for this compile.
3177 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3178 // If there is already a scratch buffer blob allocated and the
3179 // constant section is big enough, use it. Otherwise free the
3180 // current and allocate a new one.
3181 BufferBlob* blob = scratch_buffer_blob();
3182 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3183 // Use the current blob.
3184 } else {
3185 if (blob != nullptr) {
3186 BufferBlob::free(blob);
3187 }
3188
3189 ResourceMark rm;
3190 _scratch_const_size = const_size;
3191 int size = C2Compiler::initial_code_buffer_size(const_size);
3192 if (C->has_scalarized_args()) {
3193 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3194 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3195 ciMethod* method = C->method();
3196 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3197 int arg_num = 0;
3198 if (!method->is_static()) {
3199 if (method->is_scalarized_arg(arg_num)) {
3200 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3201 }
3202 arg_num++;
3203 }
3204 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3205 if (method->is_scalarized_arg(arg_num)) {
3206 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3207 }
3208 arg_num++;
3209 }
3210 }
3211 blob = BufferBlob::create("Compile::scratch_buffer", size);
3212 // Record the buffer blob for next time.
3213 set_scratch_buffer_blob(blob);
3214 // Have we run out of code space?
3215 if (scratch_buffer_blob() == nullptr) {
3216 // Let CompilerBroker disable further compilations.
3217 C->record_failure("Not enough space for scratch buffer in CodeCache");
3218 return;
3219 }
3220 }
3221
3222 // Initialize the relocation buffers
3223 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3224 set_scratch_locs_memory(locs_buf);
3225 }
3226
3227
3228 //-----------------------scratch_emit_size-------------------------------------
3229 // Helper function that computes size by emitting code
3230 uint PhaseOutput::scratch_emit_size(const Node* n) {
3231 // Start scratch_emit_size section.
3232 set_in_scratch_emit_size(true);
3233
3234 // Emit into a trash buffer and count bytes emitted.
3235 // This is a pretty expensive way to compute a size,
3236 // but it works well enough if seldom used.
3237 // All common fixed-size instructions are given a size
3238 // method by the AD file.
3239 // Note that the scratch buffer blob and locs memory are
3240 // allocated at the beginning of the compile task, and
3241 // may be shared by several calls to scratch_emit_size.
3242 // The allocation of the scratch buffer blob is particularly
3243 // expensive, since it has to grab the code cache lock.
3244 BufferBlob* blob = this->scratch_buffer_blob();
3245 assert(blob != nullptr, "Initialize BufferBlob at start");
3246 assert(blob->size() > MAX_inst_size, "sanity");
3247 relocInfo* locs_buf = scratch_locs_memory();
3248 address blob_begin = blob->content_begin();
3249 address blob_end = (address)locs_buf;
3250 assert(blob->contains(blob_end), "sanity");
3251 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3252 buf.initialize_consts_size(_scratch_const_size);
3253 buf.initialize_stubs_size(MAX_stubs_size);
3254 assert(locs_buf != nullptr, "sanity");
3255 int lsize = MAX_locs_size / 3;
3256 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3257 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3258 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3259 // Mark as scratch buffer.
3260 buf.consts()->set_scratch_emit();
3261 buf.insts()->set_scratch_emit();
3262 buf.stubs()->set_scratch_emit();
3263
3264 // Do the emission.
3265
3266 Label fakeL; // Fake label for branch instructions.
3267 Label* saveL = nullptr;
3268 uint save_bnum = 0;
3269 bool is_branch = n->is_MachBranch();
3270 C2_MacroAssembler masm(&buf);
3271 masm.bind(fakeL);
3272 if (is_branch) {
3273 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3274 n->as_MachBranch()->label_set(&fakeL, 0);
3275 }
3276 n->emit(&masm, C->regalloc());
3277
3278 // Emitting into the scratch buffer should not fail
3279 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3280
3281 // Restore label.
3282 if (is_branch) {
3283 n->as_MachBranch()->label_set(saveL, save_bnum);
3284 }
3285
3286 // End scratch_emit_size section.
3287 set_in_scratch_emit_size(false);
3288
3289 return buf.insts_size();
3290 }
3291
3292 void PhaseOutput::install() {
3293 if (!C->should_install_code()) {
3294 return;
3295 } else if (C->stub_function() != nullptr) {
3296 install_stub(C->stub_name());
3297 } else {
3298 install_code(C->method(),
3299 C->entry_bci(),
3300 CompileBroker::compiler2(),
3301 C->has_unsafe_access(),
3302 SharedRuntime::is_wide_vector(C->max_vector_size()));
3303 }
3304 }
3305
3306 void PhaseOutput::install_code(ciMethod* target,
3307 int entry_bci,
3308 AbstractCompiler* compiler,
3309 bool has_unsafe_access,
3310 bool has_wide_vectors) {
3311 // Check if we want to skip execution of all compiled code.
3312 {
3313 #ifndef PRODUCT
3314 if (OptoNoExecute) {
3315 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3316 return;
3317 }
3318 #endif
3319 Compile::TracePhase tp(_t_registerMethod);
3320
3321 if (C->is_osr_compilation()) {
3322 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3323 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3324 } else {
3325 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3326 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3327 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3328 }
3329 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3330 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3331 }
3332 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3333 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3334 }
3335 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3336 }
3337
3338 C->env()->register_method(target,
3339 entry_bci,
3340 &_code_offsets,
3341 _orig_pc_slot_offset_in_bytes,
3342 code_buffer(),
3343 frame_size_in_words(),
3344 _oop_map_set,
3345 &_handler_table,
3346 inc_table(),
3347 compiler,
3348 has_unsafe_access,
3349 SharedRuntime::is_wide_vector(C->max_vector_size()),
3350 C->has_monitors(),
3351 C->has_scoped_access(),
3352 0);
3353
3354 if (C->log() != nullptr) { // Print code cache state into compiler log
3355 C->log()->code_cache_state();
3356 }
3357 }
3358 }
3359 void PhaseOutput::install_stub(const char* stub_name) {
3360 // Entry point will be accessed using stub_entry_point();
3361 if (code_buffer() == nullptr) {
3362 Matcher::soft_match_failure();
3363 } else {
3364 if (PrintAssembly && (WizardMode || Verbose))
3365 tty->print_cr("### Stub::%s", stub_name);
3366
3367 if (!C->failing()) {
3368 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3369
3370 // Make the NMethod
3371 // For now we mark the frame as never safe for profile stackwalking
3372 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3373 code_buffer(),
3374 CodeOffsets::frame_never_safe,
3375 // _code_offsets.value(CodeOffsets::Frame_Complete),
3376 frame_size_in_words(),
3377 oop_map_set(),
3378 false,
3379 false);
3380
3381 if (rs == nullptr) {
3382 C->record_failure("CodeCache is full");
3383 } else {
3384 assert(rs->is_runtime_stub(), "sanity check");
3385 C->set_stub_entry_point(rs->entry_point());
3386 BlobId blob_id = StubInfo::blob(C->stub_id());
3387 AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, blob_id);
3388 }
3389 }
3390 }
3391 }
3392
3393 // Support for bundling info
3394 Bundle* PhaseOutput::node_bundling(const Node *n) {
3395 assert(valid_bundle_info(n), "oob");
3396 return &_node_bundling_base[n->_idx];
3397 }
3398
3399 bool PhaseOutput::valid_bundle_info(const Node *n) {
3400 return (_node_bundling_limit > n->_idx);
3401 }
3402
3403 //------------------------------frame_size_in_words-----------------------------
3404 // frame_slots in units of words
3405 int PhaseOutput::frame_size_in_words() const {
3406 // shift is 0 in LP32 and 1 in LP64
3407 const int shift = (LogBytesPerWord - LogBytesPerInt);
3408 int words = _frame_slots >> shift;
3409 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3410 return words;
3411 }
3412
3413 // To bang the stack of this compiled method we use the stack size
3414 // that the interpreter would need in case of a deoptimization. This
3415 // removes the need to bang the stack in the deoptimization blob which
3416 // in turn simplifies stack overflow handling.
3417 int PhaseOutput::bang_size_in_bytes() const {
3418 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3419 }
3420
3421 //------------------------------dump_asm---------------------------------------
3422 // Dump formatted assembly
3423 #if defined(SUPPORT_OPTO_ASSEMBLY)
3424 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3425
3426 int pc_digits = 3; // #chars required for pc
3427 int sb_chars = 3; // #chars for "start bundle" indicator
3428 int tab_size = 8;
3429 if (pcs != nullptr) {
3430 int max_pc = 0;
3431 for (uint i = 0; i < pc_limit; i++) {
3432 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3433 }
3434 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3435 }
3436 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3437
3438 bool cut_short = false;
3439 st->print_cr("#");
3440 st->print("# "); C->tf()->dump_on(st); st->cr();
3441 st->print_cr("#");
3442
3443 // For all blocks
3444 int pc = 0x0; // Program counter
3445 char starts_bundle = ' ';
3446 C->regalloc()->dump_frame();
3447
3448 Node *n = nullptr;
3449 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3450 if (VMThread::should_terminate()) {
3451 cut_short = true;
3452 break;
3453 }
3454 Block* block = C->cfg()->get_block(i);
3455 if (block->is_connector() && !Verbose) {
3456 continue;
3457 }
3458 n = block->head();
3459 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3460 pc = pcs[n->_idx];
3461 st->print("%*.*x", pc_digits, pc_digits, pc);
3462 }
3463 st->fill_to(prefix_len);
3464 block->dump_head(C->cfg(), st);
3465 if (block->is_connector()) {
3466 st->fill_to(prefix_len);
3467 st->print_cr("# Empty connector block");
3468 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3469 st->fill_to(prefix_len);
3470 st->print_cr("# Block is sole successor of call");
3471 }
3472
3473 // For all instructions
3474 for (uint j = 0; j < block->number_of_nodes(); j++) {
3475 if (VMThread::should_terminate()) {
3476 cut_short = true;
3477 break;
3478 }
3479 n = block->get_node(j);
3480 if (valid_bundle_info(n)) {
3481 Bundle* bundle = node_bundling(n);
3482 if (bundle->starts_bundle()) {
3483 starts_bundle = '+';
3484 }
3485 }
3486
3487 if (WizardMode) {
3488 n->dump();
3489 }
3490
3491 if( !n->is_Region() && // Dont print in the Assembly
3492 !n->is_Phi() && // a few noisely useless nodes
3493 !n->is_Proj() &&
3494 !n->is_MachTemp() &&
3495 !n->is_SafePointScalarObject() &&
3496 !n->is_Catch() && // Would be nice to print exception table targets
3497 !n->is_MergeMem() && // Not very interesting
3498 !n->is_top() && // Debug info table constants
3499 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3500 ) {
3501 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3502 pc = pcs[n->_idx];
3503 st->print("%*.*x", pc_digits, pc_digits, pc);
3504 } else {
3505 st->fill_to(pc_digits);
3506 }
3507 st->print(" %c ", starts_bundle);
3508 starts_bundle = ' ';
3509 st->fill_to(prefix_len);
3510 n->format(C->regalloc(), st);
3511 st->cr();
3512 }
3513
3514 // Dump the exception table as well
3515 if( n->is_Catch() && (Verbose || WizardMode) ) {
3516 // Print the exception table for this offset
3517 _handler_table.print_subtable_for(pc);
3518 }
3519 st->bol(); // Make sure we start on a new line
3520 }
3521 st->cr(); // one empty line between blocks
3522 } // End of per-block dump
3523
3524 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3525 }
3526 #endif
3527
3528 #ifndef PRODUCT
3529 void PhaseOutput::print_statistics() {
3530 Scheduling::print_statistics();
3531 }
3532 #endif