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