1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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 "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/debugInfo.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/compilerDirectives.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "gc/shared/barrierSet.hpp"
36 #include "gc/shared/c2/barrierSetC2.hpp"
37 #include "memory/allocation.inline.hpp"
38 #include "memory/allocation.hpp"
39 #include "opto/ad.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/c2_MacroAssembler.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/cfgnode.hpp"
45 #include "opto/locknode.hpp"
46 #include "opto/machnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/optoreg.hpp"
49 #include "opto/output.hpp"
50 #include "opto/regalloc.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/type.hpp"
54 #include "runtime/handles.inline.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #include "utilities/xmlstream.hpp"
59
60 #ifndef PRODUCT
61 #define DEBUG_ARG(x) , x
62 #else
63 #define DEBUG_ARG(x)
64 #endif
65
66 //------------------------------Scheduling----------------------------------
67 // This class contains all the information necessary to implement instruction
68 // scheduling and bundling.
69 class Scheduling {
70
71 private:
72 // Arena to use
73 Arena *_arena;
74
75 // Control-Flow Graph info
76 PhaseCFG *_cfg;
77
78 // Register Allocation info
79 PhaseRegAlloc *_regalloc;
80
81 // Number of nodes in the method
82 uint _node_bundling_limit;
83
84 // List of scheduled nodes. Generated in reverse order
85 Node_List _scheduled;
86
87 // List of nodes currently available for choosing for scheduling
88 Node_List _available;
89
90 // For each instruction beginning a bundle, the number of following
91 // nodes to be bundled with it.
92 Bundle *_node_bundling_base;
93
94 // Mapping from register to Node
95 Node_List _reg_node;
96
97 // Free list for pinch nodes.
98 Node_List _pinch_free_list;
99
100 // Number of uses of this node within the containing basic block.
101 short *_uses;
102
103 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
104 // front, branch+proj at end. Also skips Catch/CProj (same as
105 // branch-at-end), plus just-prior exception-throwing call.
106 uint _bb_start, _bb_end;
107
108 // Latency from the end of the basic block as scheduled
109 unsigned short *_current_latency;
110
111 // Remember the next node
112 Node *_next_node;
113
114 // Use this for an unconditional branch delay slot
115 Node *_unconditional_delay_slot;
116
117 // Pointer to a Nop
118 MachNopNode *_nop;
119
120 // Length of the current bundle, in instructions
121 uint _bundle_instr_count;
122
123 // Current Cycle number, for computing latencies and bundling
124 uint _bundle_cycle_number;
125
126 // Bundle information
127 Pipeline_Use_Element _bundle_use_elements[resource_count];
128 Pipeline_Use _bundle_use;
129
130 // Dump the available list
131 void dump_available() const;
132
133 public:
134 Scheduling(Arena *arena, Compile &compile);
135
136 // Destructor
137 NOT_PRODUCT( ~Scheduling(); )
138
139 // Step ahead "i" cycles
140 void step(uint i);
141
142 // Step ahead 1 cycle, and clear the bundle state (for example,
143 // at a branch target)
144 void step_and_clear();
145
146 Bundle* node_bundling(const Node *n) {
147 assert(valid_bundle_info(n), "oob");
148 return (&_node_bundling_base[n->_idx]);
149 }
150
151 bool valid_bundle_info(const Node *n) const {
152 return (_node_bundling_limit > n->_idx);
153 }
154
155 bool starts_bundle(const Node *n) const {
156 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
157 }
158
159 // Do the scheduling
160 void DoScheduling();
161
162 // Compute the register antidependencies within a basic block
163 void ComputeRegisterAntidependencies(Block *bb);
164 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
165 void verify_good_schedule( Block *b, const char *msg );
166 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
167 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
168
169 // Add a node to the current bundle
170 void AddNodeToBundle(Node *n, const Block *bb);
171
172 // Return an integer less than, equal to, or greater than zero
173 // if the stack offset of the first argument is respectively
174 // less than, equal to, or greater than the second.
175 int compare_two_spill_nodes(Node* first, Node* second);
176
177 // Add a node to the list of available nodes
178 void AddNodeToAvailableList(Node *n);
179
180 // Compute the local use count for the nodes in a block, and compute
181 // the list of instructions with no uses in the block as available
182 void ComputeUseCount(const Block *bb);
183
184 // Choose an instruction from the available list to add to the bundle
185 Node * ChooseNodeToBundle();
186
187 // See if this Node fits into the currently accumulating bundle
188 bool NodeFitsInBundle(Node *n);
189
190 // Decrement the use count for a node
191 void DecrementUseCounts(Node *n, const Block *bb);
192
193 // Garbage collect pinch nodes for reuse by other blocks.
194 void garbage_collect_pinch_nodes();
195 // Clean up a pinch node for reuse (helper for above).
196 void cleanup_pinch( Node *pinch );
197
198 // Information for statistics gathering
199 #ifndef PRODUCT
200 private:
201 // Gather information on size of nops relative to total
202 uint _branches, _unconditional_delays;
203
204 static uint _total_nop_size, _total_method_size;
205 static uint _total_branches, _total_unconditional_delays;
206 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
207
208 public:
209 static void print_statistics();
210
211 static void increment_instructions_per_bundle(uint i) {
212 _total_instructions_per_bundle[i]++;
213 }
214
215 static void increment_nop_size(uint s) {
216 _total_nop_size += s;
217 }
218
219 static void increment_method_size(uint s) {
220 _total_method_size += s;
221 }
222 #endif
223
224 };
225
226 PhaseOutput::PhaseOutput()
227 : Phase(Phase::Output),
228 _code_buffer("Compile::Fill_buffer"),
229 _first_block_size(0),
230 _handler_table(),
231 _inc_table(),
232 _stub_list(),
233 _oop_map_set(nullptr),
234 _scratch_buffer_blob(nullptr),
235 _scratch_locs_memory(nullptr),
236 _scratch_const_size(-1),
237 _in_scratch_emit_size(false),
238 _frame_slots(0),
239 _code_offsets(),
240 _node_bundling_limit(0),
241 _node_bundling_base(nullptr),
242 _orig_pc_slot(0),
243 _orig_pc_slot_offset_in_bytes(0),
244 _buf_sizes(),
245 _block(nullptr),
246 _index(0) {
247 C->set_output(this);
248 if (C->stub_name() == nullptr) {
249 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
250 }
251 }
252
253 PhaseOutput::~PhaseOutput() {
254 C->set_output(nullptr);
255 if (_scratch_buffer_blob != nullptr) {
256 BufferBlob::free(_scratch_buffer_blob);
257 }
258 }
259
260 void PhaseOutput::perform_mach_node_analysis() {
261 // Late barrier analysis must be done after schedule and bundle
262 // Otherwise liveness based spilling will fail
263 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
264 bs->late_barrier_analysis();
265
266 pd_perform_mach_node_analysis();
267
268 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
269 }
270
271 // Convert Nodes to instruction bits and pass off to the VM
272 void PhaseOutput::Output() {
273 // RootNode goes
274 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
275
276 // The number of new nodes (mostly MachNop) is proportional to
277 // the number of java calls and inner loops which are aligned.
278 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
279 C->inner_loops()*(OptoLoopAlignment-1)),
280 "out of nodes before code generation" ) ) {
281 return;
282 }
283 // Make sure I can find the Start Node
284 Block *entry = C->cfg()->get_block(1);
285 Block *broot = C->cfg()->get_root_block();
286
287 const StartNode *start = entry->head()->as_Start();
288
289 // Replace StartNode with prolog
290 MachPrologNode *prolog = new MachPrologNode();
291 entry->map_node(prolog, 0);
292 C->cfg()->map_node_to_block(prolog, entry);
293 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
294
295 // Virtual methods need an unverified entry point
296
297 if( C->is_osr_compilation() ) {
298 if( PoisonOSREntry ) {
299 // TODO: Should use a ShouldNotReachHereNode...
300 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
301 }
302 } else {
303 if( C->method() && !C->method()->flags().is_static() ) {
304 // Insert unvalidated entry point
305 C->cfg()->insert( broot, 0, new MachUEPNode() );
306 }
307
308 }
309
310 // Break before main entry point
311 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
312 (OptoBreakpoint && C->is_method_compilation()) ||
313 (OptoBreakpointOSR && C->is_osr_compilation()) ||
314 (OptoBreakpointC2R && !C->method()) ) {
315 // checking for C->method() means that OptoBreakpoint does not apply to
316 // runtime stubs or frame converters
317 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
318 }
319
320 // Insert epilogs before every return
321 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
322 Block* block = C->cfg()->get_block(i);
323 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
324 Node* m = block->end();
325 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
326 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
327 block->add_inst(epilog);
328 C->cfg()->map_node_to_block(epilog, block);
329 }
330 }
331 }
332
333 // Keeper of sizing aspects
334 _buf_sizes = BufferSizingData();
335
336 // Initialize code buffer
337 estimate_buffer_size(_buf_sizes._const);
338 if (C->failing()) return;
339
340 // Pre-compute the length of blocks and replace
341 // long branches with short if machine supports it.
342 // Must be done before ScheduleAndBundle due to SPARC delay slots
343 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
344 blk_starts[0] = 0;
345 shorten_branches(blk_starts);
346
347 ScheduleAndBundle();
348 if (C->failing()) {
349 return;
350 }
351
352 perform_mach_node_analysis();
353
354 // Complete sizing of codebuffer
355 CodeBuffer* cb = init_buffer();
356 if (cb == nullptr || C->failing()) {
357 return;
358 }
359
360 BuildOopMaps();
361
362 if (C->failing()) {
363 return;
364 }
365
366 C2_MacroAssembler masm(cb);
367 fill_buffer(&masm, blk_starts);
368 }
369
370 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
371 // Determine if we need to generate a stack overflow check.
372 // Do it if the method is not a stub function and
373 // has java calls or has frame size > vm_page_size/8.
374 // The debug VM checks that deoptimization doesn't trigger an
375 // unexpected stack overflow (compiled method stack banging should
376 // guarantee it doesn't happen) so we always need the stack bang in
377 // a debug VM.
378 return (C->stub_function() == nullptr &&
379 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
380 DEBUG_ONLY(|| true)));
381 }
382
383 bool PhaseOutput::need_register_stack_bang() const {
384 // Determine if we need to generate a register stack overflow check.
385 // This is only used on architectures which have split register
386 // and memory stacks (ie. IA64).
387 // Bang if the method is not a stub function and has java calls
388 return (C->stub_function() == nullptr && C->has_java_calls());
389 }
390
391
392 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
393 // of a loop. When aligning a loop we need to provide enough instructions
394 // in cpu's fetch buffer to feed decoders. The loop alignment could be
395 // avoided if we have enough instructions in fetch buffer at the head of a loop.
396 // By default, the size is set to 999999 by Block's constructor so that
397 // a loop will be aligned if the size is not reset here.
398 //
399 // Note: Mach instructions could contain several HW instructions
400 // so the size is estimated only.
401 //
402 void PhaseOutput::compute_loop_first_inst_sizes() {
403 // The next condition is used to gate the loop alignment optimization.
404 // Don't aligned a loop if there are enough instructions at the head of a loop
405 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
406 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
407 // equal to 11 bytes which is the largest address NOP instruction.
408 if (MaxLoopPad < OptoLoopAlignment - 1) {
409 uint last_block = C->cfg()->number_of_blocks() - 1;
410 for (uint i = 1; i <= last_block; i++) {
411 Block* block = C->cfg()->get_block(i);
412 // Check the first loop's block which requires an alignment.
413 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
414 uint sum_size = 0;
415 uint inst_cnt = NumberOfLoopInstrToAlign;
416 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
417
418 // Check subsequent fallthrough blocks if the loop's first
419 // block(s) does not have enough instructions.
420 Block *nb = block;
421 while(inst_cnt > 0 &&
422 i < last_block &&
423 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
424 !nb->has_successor(block)) {
425 i++;
426 nb = C->cfg()->get_block(i);
427 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
428 } // while( inst_cnt > 0 && i < last_block )
429
430 block->set_first_inst_size(sum_size);
431 } // f( b->head()->is_Loop() )
432 } // for( i <= last_block )
433 } // if( MaxLoopPad < OptoLoopAlignment-1 )
434 }
435
436 // The architecture description provides short branch variants for some long
437 // branch instructions. Replace eligible long branches with short branches.
438 void PhaseOutput::shorten_branches(uint* blk_starts) {
439
440 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
441
442 // Compute size of each block, method size, and relocation information size
443 uint nblocks = C->cfg()->number_of_blocks();
444
445 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
446 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
447 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
448
449 // Collect worst case block paddings
450 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
451 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
452
453 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
454 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
455
456 bool has_short_branch_candidate = false;
457
458 // Initialize the sizes to 0
459 int code_size = 0; // Size in bytes of generated code
460 int stub_size = 0; // Size in bytes of all stub entries
461 // Size in bytes of all relocation entries, including those in local stubs.
462 // Start with 2-bytes of reloc info for the unvalidated entry point
463 int reloc_size = 1; // Number of relocation entries
464
465 // Make three passes. The first computes pessimistic blk_starts,
466 // relative jmp_offset and reloc_size information. The second performs
467 // short branch substitution using the pessimistic sizing. The
468 // third inserts nops where needed.
469
470 // Step one, perform a pessimistic sizing pass.
471 uint last_call_adr = max_juint;
472 uint last_avoid_back_to_back_adr = max_juint;
473 uint nop_size = (new MachNopNode())->size(C->regalloc());
474 for (uint i = 0; i < nblocks; i++) { // For all blocks
475 Block* block = C->cfg()->get_block(i);
476 _block = block;
477
478 // During short branch replacement, we store the relative (to blk_starts)
479 // offset of jump in jmp_offset, rather than the absolute offset of jump.
480 // This is so that we do not need to recompute sizes of all nodes when
481 // we compute correct blk_starts in our next sizing pass.
482 jmp_offset[i] = 0;
483 jmp_size[i] = 0;
484 jmp_nidx[i] = -1;
485 DEBUG_ONLY( jmp_target[i] = 0; )
486 DEBUG_ONLY( jmp_rule[i] = 0; )
487
488 // Sum all instruction sizes to compute block size
489 uint last_inst = block->number_of_nodes();
490 uint blk_size = 0;
491 for (uint j = 0; j < last_inst; j++) {
492 _index = j;
493 Node* nj = block->get_node(_index);
494 // Handle machine instruction nodes
495 if (nj->is_Mach()) {
496 MachNode* mach = nj->as_Mach();
497 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
498 reloc_size += mach->reloc();
499 if (mach->is_MachCall()) {
500 // add size information for trampoline stub
501 // class CallStubImpl is platform-specific and defined in the *.ad files.
502 stub_size += CallStubImpl::size_call_trampoline();
503 reloc_size += CallStubImpl::reloc_call_trampoline();
504
505 MachCallNode *mcall = mach->as_MachCall();
506 // This destination address is NOT PC-relative
507
508 mcall->method_set((intptr_t)mcall->entry_point());
509
510 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
511 stub_size += CompiledDirectCall::to_interp_stub_size();
512 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
513 }
514 } else if (mach->is_MachSafePoint()) {
515 // If call/safepoint are adjacent, account for possible
516 // nop to disambiguate the two safepoints.
517 // ScheduleAndBundle() can rearrange nodes in a block,
518 // check for all offsets inside this block.
519 if (last_call_adr >= blk_starts[i]) {
520 blk_size += nop_size;
521 }
522 }
523 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
524 // Nop is inserted between "avoid back to back" instructions.
525 // ScheduleAndBundle() can rearrange nodes in a block,
526 // check for all offsets inside this block.
527 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
528 blk_size += nop_size;
529 }
530 }
531 if (mach->may_be_short_branch()) {
532 if (!nj->is_MachBranch()) {
533 #ifndef PRODUCT
534 nj->dump(3);
535 #endif
536 Unimplemented();
537 }
538 assert(jmp_nidx[i] == -1, "block should have only one branch");
539 jmp_offset[i] = blk_size;
540 jmp_size[i] = nj->size(C->regalloc());
541 jmp_nidx[i] = j;
542 has_short_branch_candidate = true;
543 }
544 }
545 blk_size += nj->size(C->regalloc());
546 // Remember end of call offset
547 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
548 last_call_adr = blk_starts[i]+blk_size;
549 }
550 // Remember end of avoid_back_to_back offset
551 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
552 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
553 }
554 }
555
556 // When the next block starts a loop, we may insert pad NOP
557 // instructions. Since we cannot know our future alignment,
558 // assume the worst.
559 if (i < nblocks - 1) {
560 Block* nb = C->cfg()->get_block(i + 1);
561 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
562 if (max_loop_pad > 0) {
563 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
564 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
565 // If either is the last instruction in this block, bump by
566 // max_loop_pad in lock-step with blk_size, so sizing
567 // calculations in subsequent blocks still can conservatively
568 // detect that it may the last instruction in this block.
569 if (last_call_adr == blk_starts[i]+blk_size) {
570 last_call_adr += max_loop_pad;
571 }
572 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
573 last_avoid_back_to_back_adr += max_loop_pad;
574 }
575 blk_size += max_loop_pad;
576 block_worst_case_pad[i + 1] = max_loop_pad;
577 }
578 }
579
580 // Save block size; update total method size
581 blk_starts[i+1] = blk_starts[i]+blk_size;
582 }
583
584 // Step two, replace eligible long jumps.
585 bool progress = true;
586 uint last_may_be_short_branch_adr = max_juint;
587 while (has_short_branch_candidate && progress) {
588 progress = false;
589 has_short_branch_candidate = false;
590 int adjust_block_start = 0;
591 for (uint i = 0; i < nblocks; i++) {
592 Block* block = C->cfg()->get_block(i);
593 int idx = jmp_nidx[i];
594 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
595 if (mach != nullptr && mach->may_be_short_branch()) {
596 #ifdef ASSERT
597 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
598 int j;
599 // Find the branch; ignore trailing NOPs.
600 for (j = block->number_of_nodes()-1; j>=0; j--) {
601 Node* n = block->get_node(j);
602 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
603 break;
604 }
605 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
606 #endif
607 int br_size = jmp_size[i];
608 int br_offs = blk_starts[i] + jmp_offset[i];
609
610 // This requires the TRUE branch target be in succs[0]
611 uint bnum = block->non_connector_successor(0)->_pre_order;
612 int offset = blk_starts[bnum] - br_offs;
613 if (bnum > i) { // adjust following block's offset
614 offset -= adjust_block_start;
615 }
616
617 // This block can be a loop header, account for the padding
618 // in the previous block.
619 int block_padding = block_worst_case_pad[i];
620 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
621 // In the following code a nop could be inserted before
622 // the branch which will increase the backward distance.
623 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
624 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
625
626 if (needs_padding && offset <= 0)
627 offset -= nop_size;
628
629 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
630 // We've got a winner. Replace this branch.
631 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
632
633 // Update the jmp_size.
634 int new_size = replacement->size(C->regalloc());
635 int diff = br_size - new_size;
636 assert(diff >= (int)nop_size, "short_branch size should be smaller");
637 // Conservatively take into account padding between
638 // avoid_back_to_back branches. Previous branch could be
639 // converted into avoid_back_to_back branch during next
640 // rounds.
641 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
642 jmp_offset[i] += nop_size;
643 diff -= nop_size;
644 }
645 adjust_block_start += diff;
646 block->map_node(replacement, idx);
647 mach->subsume_by(replacement, C);
648 mach = replacement;
649 progress = true;
650
651 jmp_size[i] = new_size;
652 DEBUG_ONLY( jmp_target[i] = bnum; );
653 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
654 } else {
655 // The jump distance is not short, try again during next iteration.
656 has_short_branch_candidate = true;
657 }
658 } // (mach->may_be_short_branch())
659 if (mach != nullptr && (mach->may_be_short_branch() ||
660 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
661 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
662 }
663 blk_starts[i+1] -= adjust_block_start;
664 }
665 }
666
667 #ifdef ASSERT
668 for (uint i = 0; i < nblocks; i++) { // For all blocks
669 if (jmp_target[i] != 0) {
670 int br_size = jmp_size[i];
671 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
672 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
673 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
674 }
675 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
676 }
677 }
678 #endif
679
680 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
681 // after ScheduleAndBundle().
682
683 // ------------------
684 // Compute size for code buffer
685 code_size = blk_starts[nblocks];
686
687 // Relocation records
688 reloc_size += 1; // Relo entry for exception handler
689
690 // Adjust reloc_size to number of record of relocation info
691 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
692 // a relocation index.
693 // The CodeBuffer will expand the locs array if this estimate is too low.
694 reloc_size *= 10 / sizeof(relocInfo);
695
696 _buf_sizes._reloc = reloc_size;
697 _buf_sizes._code = code_size;
698 _buf_sizes._stub = stub_size;
699 }
700
701 //------------------------------FillLocArray-----------------------------------
702 // Create a bit of debug info and append it to the array. The mapping is from
703 // Java local or expression stack to constant, register or stack-slot. For
704 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
705 // entry has been taken care of and caller should skip it).
706 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
707 // This should never have accepted Bad before
708 assert(OptoReg::is_valid(regnum), "location must be valid");
709 return (OptoReg::is_reg(regnum))
710 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
711 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
712 }
713
714
715 ObjectValue*
716 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
717 for (int i = 0; i < objs->length(); i++) {
718 assert(objs->at(i)->is_object(), "corrupt object cache");
719 ObjectValue* sv = objs->at(i)->as_ObjectValue();
720 if (sv->id() == id) {
721 return sv;
722 }
723 }
724 // Otherwise..
725 return nullptr;
726 }
727
728 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
729 ObjectValue* sv ) {
730 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
731 objs->append(sv);
732 }
733
734
735 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
736 GrowableArray<ScopeValue*> *array,
737 GrowableArray<ScopeValue*> *objs ) {
738 assert( local, "use _top instead of null" );
739 if (array->length() != idx) {
740 assert(array->length() == idx + 1, "Unexpected array count");
741 // Old functionality:
742 // return
743 // New functionality:
744 // Assert if the local is not top. In product mode let the new node
745 // override the old entry.
746 assert(local == C->top(), "LocArray collision");
747 if (local == C->top()) {
748 return;
749 }
750 array->pop();
751 }
752 const Type *t = local->bottom_type();
753
754 // Is it a safepoint scalar object node?
755 if (local->is_SafePointScalarObject()) {
756 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
757
758 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
759 if (sv == nullptr) {
760 ciKlass* cik = t->is_oopptr()->exact_klass();
761 assert(cik->is_instance_klass() ||
762 cik->is_array_klass(), "Not supported allocation.");
763 sv = new ObjectValue(spobj->_idx,
764 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
765 set_sv_for_object_node(objs, sv);
766
767 uint first_ind = spobj->first_index(sfpt->jvms());
768 for (uint i = 0; i < spobj->n_fields(); i++) {
769 Node* fld_node = sfpt->in(first_ind+i);
770 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
771 }
772 }
773 array->append(sv);
774 return;
775 } else if (local->is_SafePointScalarMerge()) {
776 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
777 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
778
779 if (mv == nullptr) {
780 GrowableArray<ScopeValue*> deps;
781
782 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
783 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
784 assert(deps.length() == 1, "missing value");
785
786 int selector_idx = smerge->selector_idx(sfpt->jvms());
787 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
788 assert(deps.length() == 2, "missing value");
789
790 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
791 set_sv_for_object_node(objs, mv);
792
793 for (uint i = 1; i < smerge->req(); i++) {
794 Node* obj_node = smerge->in(i);
795 (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs);
796 }
797 }
798 array->append(mv);
799 return;
800 }
801
802 // Grab the register number for the local
803 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
804 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
805 // Record the double as two float registers.
806 // The register mask for such a value always specifies two adjacent
807 // float registers, with the lower register number even.
808 // Normally, the allocation of high and low words to these registers
809 // is irrelevant, because nearly all operations on register pairs
810 // (e.g., StoreD) treat them as a single unit.
811 // Here, we assume in addition that the words in these two registers
812 // stored "naturally" (by operations like StoreD and double stores
813 // within the interpreter) such that the lower-numbered register
814 // is written to the lower memory address. This may seem like
815 // a machine dependency, but it is not--it is a requirement on
816 // the author of the <arch>.ad file to ensure that, for every
817 // even/odd double-register pair to which a double may be allocated,
818 // the word in the even single-register is stored to the first
819 // memory word. (Note that register numbers are completely
820 // arbitrary, and are not tied to any machine-level encodings.)
821 #ifdef _LP64
822 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
823 array->append(new ConstantIntValue((jint)0));
824 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
825 } else if ( t->base() == Type::Long ) {
826 array->append(new ConstantIntValue((jint)0));
827 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
828 } else if ( t->base() == Type::RawPtr ) {
829 // jsr/ret return address which must be restored into the full
830 // width 64-bit stack slot.
831 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
832 }
833 #else //_LP64
834 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
835 // Repack the double/long as two jints.
836 // The convention the interpreter uses is that the second local
837 // holds the first raw word of the native double representation.
838 // This is actually reasonable, since locals and stack arrays
839 // grow downwards in all implementations.
840 // (If, on some machine, the interpreter's Java locals or stack
841 // were to grow upwards, the embedded doubles would be word-swapped.)
842 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
843 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
844 }
845 #endif //_LP64
846 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
847 OptoReg::is_reg(regnum) ) {
848 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
849 ? Location::float_in_dbl : Location::normal ));
850 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
851 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
852 ? Location::int_in_long : Location::normal ));
853 } else if( t->base() == Type::NarrowOop ) {
854 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
855 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
856 t->base() == Type::VectorD || t->base() == Type::VectorX ||
857 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
858 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
859 } else if (C->regalloc()->is_oop(local)) {
860 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
861 t->base() == Type::AryPtr,
862 "Unexpected type: %s", t->msg());
863 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
864 } else {
865 assert(t->base() == Type::Int || t->base() == Type::Half ||
866 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
867 "Unexpected type: %s", t->msg());
868 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
869 }
870 return;
871 }
872
873 // No register. It must be constant data.
874 switch (t->base()) {
875 case Type::Half: // Second half of a double
876 ShouldNotReachHere(); // Caller should skip 2nd halves
877 break;
878 case Type::AnyPtr:
879 array->append(new ConstantOopWriteValue(nullptr));
880 break;
881 case Type::AryPtr:
882 case Type::InstPtr: // fall through
883 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
884 break;
885 case Type::NarrowOop:
886 if (t == TypeNarrowOop::NULL_PTR) {
887 array->append(new ConstantOopWriteValue(nullptr));
888 } else {
889 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
890 }
891 break;
892 case Type::Int:
893 array->append(new ConstantIntValue(t->is_int()->get_con()));
894 break;
895 case Type::RawPtr:
896 // A return address (T_ADDRESS).
897 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
898 #ifdef _LP64
899 // Must be restored to the full-width 64-bit stack slot.
900 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
901 #else
902 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
903 #endif
904 break;
905 case Type::FloatCon: {
906 float f = t->is_float_constant()->getf();
907 array->append(new ConstantIntValue(jint_cast(f)));
908 break;
909 }
910 case Type::DoubleCon: {
911 jdouble d = t->is_double_constant()->getd();
912 #ifdef _LP64
913 array->append(new ConstantIntValue((jint)0));
914 array->append(new ConstantDoubleValue(d));
915 #else
916 // Repack the double as two jints.
917 // The convention the interpreter uses is that the second local
918 // holds the first raw word of the native double representation.
919 // This is actually reasonable, since locals and stack arrays
920 // grow downwards in all implementations.
921 // (If, on some machine, the interpreter's Java locals or stack
922 // were to grow upwards, the embedded doubles would be word-swapped.)
923 jlong_accessor acc;
924 acc.long_value = jlong_cast(d);
925 array->append(new ConstantIntValue(acc.words[1]));
926 array->append(new ConstantIntValue(acc.words[0]));
927 #endif
928 break;
929 }
930 case Type::Long: {
931 jlong d = t->is_long()->get_con();
932 #ifdef _LP64
933 array->append(new ConstantIntValue((jint)0));
934 array->append(new ConstantLongValue(d));
935 #else
936 // Repack the long as two jints.
937 // The convention the interpreter uses is that the second local
938 // holds the first raw word of the native double representation.
939 // This is actually reasonable, since locals and stack arrays
940 // grow downwards in all implementations.
941 // (If, on some machine, the interpreter's Java locals or stack
942 // were to grow upwards, the embedded doubles would be word-swapped.)
943 jlong_accessor acc;
944 acc.long_value = d;
945 array->append(new ConstantIntValue(acc.words[1]));
946 array->append(new ConstantIntValue(acc.words[0]));
947 #endif
948 break;
949 }
950 case Type::Top: // Add an illegal value here
951 array->append(new LocationValue(Location()));
952 break;
953 default:
954 ShouldNotReachHere();
955 break;
956 }
957 }
958
959 // Determine if this node starts a bundle
960 bool PhaseOutput::starts_bundle(const Node *n) const {
961 return (_node_bundling_limit > n->_idx &&
962 _node_bundling_base[n->_idx].starts_bundle());
963 }
964
965 // Determine if there is a monitor that has 'ov' as its owner.
966 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
967 for (int k = 0; k < monarray->length(); k++) {
968 MonitorValue* mv = monarray->at(k);
969 if (mv->owner() == ov) {
970 return true;
971 }
972 }
973
974 return false;
975 }
976
977 // Determine if there is a scalar replaced object description represented by 'ov'.
978 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
979 GrowableArray<ScopeValue*>* objs,
980 ObjectValue* ov) const {
981 for (int i = 0; i < jvms->scl_size(); i++) {
982 Node* n = sfn->scalarized_obj(jvms, i);
983 // Other kinds of nodes that we may encounter here, for instance constants
984 // representing values of fields of objects scalarized, aren't relevant for
985 // us, since they don't map to ObjectValue.
986 if (!n->is_SafePointScalarObject()) {
987 continue;
988 }
989
990 ObjectValue* other = sv_for_node_id(objs, n->_idx);
991 if (ov == other) {
992 return true;
993 }
994 }
995 return false;
996 }
997
998 //--------------------------Process_OopMap_Node--------------------------------
999 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1000 // Handle special safepoint nodes for synchronization
1001 MachSafePointNode *sfn = mach->as_MachSafePoint();
1002 MachCallNode *mcall;
1003
1004 int safepoint_pc_offset = current_offset;
1005 bool is_method_handle_invoke = false;
1006 bool return_oop = false;
1007 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1008 bool arg_escape = false;
1009
1010 // Add the safepoint in the DebugInfoRecorder
1011 if( !mach->is_MachCall() ) {
1012 mcall = nullptr;
1013 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1014 } else {
1015 mcall = mach->as_MachCall();
1016
1017 // Is the call a MethodHandle call?
1018 if (mcall->is_MachCallJava()) {
1019 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1020 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1021 is_method_handle_invoke = true;
1022 }
1023 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1024 }
1025
1026 // Check if a call returns an object.
1027 if (mcall->returns_pointer()) {
1028 return_oop = true;
1029 }
1030 safepoint_pc_offset += mcall->ret_addr_offset();
1031 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1032 }
1033
1034 // Loop over the JVMState list to add scope information
1035 // Do not skip safepoints with a null method, they need monitor info
1036 JVMState* youngest_jvms = sfn->jvms();
1037 int max_depth = youngest_jvms->depth();
1038
1039 // Allocate the object pool for scalar-replaced objects -- the map from
1040 // small-integer keys (which can be recorded in the local and ostack
1041 // arrays) to descriptions of the object state.
1042 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1043
1044 // Visit scopes from oldest to youngest.
1045 for (int depth = 1; depth <= max_depth; depth++) {
1046 JVMState* jvms = youngest_jvms->of_depth(depth);
1047 int idx;
1048 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1049 // Safepoints that do not have method() set only provide oop-map and monitor info
1050 // to support GC; these do not support deoptimization.
1051 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1052 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1053 int num_mon = jvms->nof_monitors();
1054 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1055 "JVMS local count must match that of the method");
1056
1057 // Add Local and Expression Stack Information
1058
1059 // Insert locals into the locarray
1060 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1061 for( idx = 0; idx < num_locs; idx++ ) {
1062 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1063 }
1064
1065 // Insert expression stack entries into the exparray
1066 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1067 for( idx = 0; idx < num_exps; idx++ ) {
1068 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1069 }
1070
1071 // Add in mappings of the monitors
1072 assert( !method ||
1073 !method->is_synchronized() ||
1074 method->is_native() ||
1075 num_mon > 0 ||
1076 !GenerateSynchronizationCode,
1077 "monitors must always exist for synchronized methods");
1078
1079 // Build the growable array of ScopeValues for exp stack
1080 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1081
1082 // Loop over monitors and insert into array
1083 for (idx = 0; idx < num_mon; idx++) {
1084 // Grab the node that defines this monitor
1085 Node* box_node = sfn->monitor_box(jvms, idx);
1086 Node* obj_node = sfn->monitor_obj(jvms, idx);
1087
1088 // Create ScopeValue for object
1089 ScopeValue *scval = nullptr;
1090
1091 if (obj_node->is_SafePointScalarObject()) {
1092 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1093 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1094 if (scval == nullptr) {
1095 const Type *t = spobj->bottom_type();
1096 ciKlass* cik = t->is_oopptr()->exact_klass();
1097 assert(cik->is_instance_klass() ||
1098 cik->is_array_klass(), "Not supported allocation.");
1099 ObjectValue* sv = new ObjectValue(spobj->_idx,
1100 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1101 PhaseOutput::set_sv_for_object_node(objs, sv);
1102
1103 uint first_ind = spobj->first_index(youngest_jvms);
1104 for (uint i = 0; i < spobj->n_fields(); i++) {
1105 Node* fld_node = sfn->in(first_ind+i);
1106 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1107 }
1108 scval = sv;
1109 }
1110 } else if (obj_node->is_SafePointScalarMerge()) {
1111 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1112 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1113
1114 if (mv == nullptr) {
1115 GrowableArray<ScopeValue*> deps;
1116
1117 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1118 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1119 assert(deps.length() == 1, "missing value");
1120
1121 int selector_idx = smerge->selector_idx(youngest_jvms);
1122 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1123 assert(deps.length() == 2, "missing value");
1124
1125 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1126 set_sv_for_object_node(objs, mv);
1127
1128 for (uint i = 1; i < smerge->req(); i++) {
1129 Node* obj_node = smerge->in(i);
1130 FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs);
1131 }
1132 }
1133 scval = mv;
1134 } else if (!obj_node->is_Con()) {
1135 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1136 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1137 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1138 } else {
1139 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1140 }
1141 } else {
1142 const TypePtr *tp = obj_node->get_ptr_type();
1143 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1144 }
1145
1146 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1147 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1148 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1149 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1150 }
1151
1152 // Mark ObjectValue nodes as root nodes if they are directly
1153 // referenced in the JVMS.
1154 for (int i = 0; i < objs->length(); i++) {
1155 ScopeValue* sv = objs->at(i);
1156 if (sv->is_object_merge()) {
1157 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1158
1159 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1160 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1161 bool is_root = locarray->contains(ov) ||
1162 exparray->contains(ov) ||
1163 contains_as_owner(monarray, ov) ||
1164 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1165 ov->set_root(is_root);
1166 }
1167 }
1168 }
1169
1170 // We dump the object pool first, since deoptimization reads it in first.
1171 C->debug_info()->dump_object_pool(objs);
1172
1173 // Build first class objects to pass to scope
1174 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1175 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1176 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1177
1178 // Make method available for all Safepoints
1179 ciMethod* scope_method = method ? method : C->method();
1180 // Describe the scope here
1181 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1182 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1183 // Now we can describe the scope.
1184 methodHandle null_mh;
1185 bool rethrow_exception = false;
1186 C->debug_info()->describe_scope(
1187 safepoint_pc_offset,
1188 null_mh,
1189 scope_method,
1190 jvms->bci(),
1191 jvms->should_reexecute(),
1192 rethrow_exception,
1193 is_method_handle_invoke,
1194 return_oop,
1195 has_ea_local_in_scope,
1196 arg_escape,
1197 locvals,
1198 expvals,
1199 monvals
1200 );
1201 } // End jvms loop
1202
1203 // Mark the end of the scope set.
1204 C->debug_info()->end_safepoint(safepoint_pc_offset);
1205 }
1206
1207
1208
1209 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1210 class NonSafepointEmitter {
1211 Compile* C;
1212 JVMState* _pending_jvms;
1213 int _pending_offset;
1214
1215 void emit_non_safepoint();
1216
1217 public:
1218 NonSafepointEmitter(Compile* compile) {
1219 this->C = compile;
1220 _pending_jvms = nullptr;
1221 _pending_offset = 0;
1222 }
1223
1224 void observe_instruction(Node* n, int pc_offset) {
1225 if (!C->debug_info()->recording_non_safepoints()) return;
1226
1227 Node_Notes* nn = C->node_notes_at(n->_idx);
1228 if (nn == nullptr || nn->jvms() == nullptr) return;
1229 if (_pending_jvms != nullptr &&
1230 _pending_jvms->same_calls_as(nn->jvms())) {
1231 // Repeated JVMS? Stretch it up here.
1232 _pending_offset = pc_offset;
1233 } else {
1234 if (_pending_jvms != nullptr &&
1235 _pending_offset < pc_offset) {
1236 emit_non_safepoint();
1237 }
1238 _pending_jvms = nullptr;
1239 if (pc_offset > C->debug_info()->last_pc_offset()) {
1240 // This is the only way _pending_jvms can become non-null:
1241 _pending_jvms = nn->jvms();
1242 _pending_offset = pc_offset;
1243 }
1244 }
1245 }
1246
1247 // Stay out of the way of real safepoints:
1248 void observe_safepoint(JVMState* jvms, int pc_offset) {
1249 if (_pending_jvms != nullptr &&
1250 !_pending_jvms->same_calls_as(jvms) &&
1251 _pending_offset < pc_offset) {
1252 emit_non_safepoint();
1253 }
1254 _pending_jvms = nullptr;
1255 }
1256
1257 void flush_at_end() {
1258 if (_pending_jvms != nullptr) {
1259 emit_non_safepoint();
1260 }
1261 _pending_jvms = nullptr;
1262 }
1263 };
1264
1265 void NonSafepointEmitter::emit_non_safepoint() {
1266 JVMState* youngest_jvms = _pending_jvms;
1267 int pc_offset = _pending_offset;
1268
1269 // Clear it now:
1270 _pending_jvms = nullptr;
1271
1272 DebugInformationRecorder* debug_info = C->debug_info();
1273 assert(debug_info->recording_non_safepoints(), "sanity");
1274
1275 debug_info->add_non_safepoint(pc_offset);
1276 int max_depth = youngest_jvms->depth();
1277
1278 // Visit scopes from oldest to youngest.
1279 for (int depth = 1; depth <= max_depth; depth++) {
1280 JVMState* jvms = youngest_jvms->of_depth(depth);
1281 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1282 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1283 methodHandle null_mh;
1284 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1285 }
1286
1287 // Mark the end of the scope set.
1288 debug_info->end_non_safepoint(pc_offset);
1289 }
1290
1291 //------------------------------init_buffer------------------------------------
1292 void PhaseOutput::estimate_buffer_size(int& const_req) {
1293
1294 // Set the initially allocated size
1295 const_req = initial_const_capacity;
1296
1297 // The extra spacing after the code is necessary on some platforms.
1298 // Sometimes we need to patch in a jump after the last instruction,
1299 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1300
1301 // Compute the byte offset where we can store the deopt pc.
1302 if (C->fixed_slots() != 0) {
1303 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1304 }
1305
1306 // Compute prolog code size
1307 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1308 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1309
1310 if (C->has_mach_constant_base_node()) {
1311 uint add_size = 0;
1312 // Fill the constant table.
1313 // Note: This must happen before shorten_branches.
1314 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1315 Block* b = C->cfg()->get_block(i);
1316
1317 for (uint j = 0; j < b->number_of_nodes(); j++) {
1318 Node* n = b->get_node(j);
1319
1320 // If the node is a MachConstantNode evaluate the constant
1321 // value section.
1322 if (n->is_MachConstant()) {
1323 MachConstantNode* machcon = n->as_MachConstant();
1324 machcon->eval_constant(C);
1325 } else if (n->is_Mach()) {
1326 // On Power there are more nodes that issue constants.
1327 add_size += (n->as_Mach()->ins_num_consts() * 8);
1328 }
1329 }
1330 }
1331
1332 // Calculate the offsets of the constants and the size of the
1333 // constant table (including the padding to the next section).
1334 constant_table().calculate_offsets_and_size();
1335 const_req = constant_table().size() + add_size;
1336 }
1337
1338 // Initialize the space for the BufferBlob used to find and verify
1339 // instruction size in MachNode::emit_size()
1340 init_scratch_buffer_blob(const_req);
1341 }
1342
1343 CodeBuffer* PhaseOutput::init_buffer() {
1344 int stub_req = _buf_sizes._stub;
1345 int code_req = _buf_sizes._code;
1346 int const_req = _buf_sizes._const;
1347
1348 int pad_req = NativeCall::byte_size();
1349
1350 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1351 stub_req += bs->estimate_stub_size();
1352
1353 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1354 // class HandlerImpl is platform-specific and defined in the *.ad files.
1355 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1356 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1357 stub_req += MAX_stubs_size; // ensure per-stub margin
1358 code_req += MAX_inst_size; // ensure per-instruction margin
1359
1360 if (StressCodeBuffers)
1361 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1362
1363 int total_req =
1364 const_req +
1365 code_req +
1366 pad_req +
1367 stub_req +
1368 exception_handler_req +
1369 deopt_handler_req; // deopt handler
1370
1371 if (C->has_method_handle_invokes())
1372 total_req += deopt_handler_req; // deopt MH handler
1373
1374 CodeBuffer* cb = code_buffer();
1375 cb->initialize(total_req, _buf_sizes._reloc);
1376
1377 // Have we run out of code space?
1378 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1379 C->record_failure("CodeCache is full");
1380 return nullptr;
1381 }
1382 // Configure the code buffer.
1383 cb->initialize_consts_size(const_req);
1384 cb->initialize_stubs_size(stub_req);
1385 cb->initialize_oop_recorder(C->env()->oop_recorder());
1386
1387 // fill in the nop array for bundling computations
1388 MachNode *_nop_list[Bundle::_nop_count];
1389 Bundle::initialize_nops(_nop_list);
1390
1391 return cb;
1392 }
1393
1394 //------------------------------fill_buffer------------------------------------
1395 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1396 // blk_starts[] contains offsets calculated during short branches processing,
1397 // offsets should not be increased during following steps.
1398
1399 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1400 // of a loop. It is used to determine the padding for loop alignment.
1401 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1402
1403 compute_loop_first_inst_sizes();
1404
1405 // Create oopmap set.
1406 _oop_map_set = new OopMapSet();
1407
1408 // !!!!! This preserves old handling of oopmaps for now
1409 C->debug_info()->set_oopmaps(_oop_map_set);
1410
1411 uint nblocks = C->cfg()->number_of_blocks();
1412 // Count and start of implicit null check instructions
1413 uint inct_cnt = 0;
1414 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1415
1416 // Count and start of calls
1417 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1418
1419 uint return_offset = 0;
1420 int nop_size = (new MachNopNode())->size(C->regalloc());
1421
1422 int previous_offset = 0;
1423 int current_offset = 0;
1424 int last_call_offset = -1;
1425 int last_avoid_back_to_back_offset = -1;
1426 #ifdef ASSERT
1427 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1428 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1429 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1430 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1431 #endif
1432
1433 // Create an array of unused labels, one for each basic block, if printing is enabled
1434 #if defined(SUPPORT_OPTO_ASSEMBLY)
1435 int* node_offsets = nullptr;
1436 uint node_offset_limit = C->unique();
1437
1438 if (C->print_assembly()) {
1439 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1440 }
1441 if (node_offsets != nullptr) {
1442 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1443 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1444 }
1445 #endif
1446
1447 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1448
1449 // Emit the constant table.
1450 if (C->has_mach_constant_base_node()) {
1451 if (!constant_table().emit(masm)) {
1452 C->record_failure("consts section overflow");
1453 return;
1454 }
1455 }
1456
1457 // Create an array of labels, one for each basic block
1458 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1459 for (uint i = 0; i <= nblocks; i++) {
1460 blk_labels[i].init();
1461 }
1462
1463 // Now fill in the code buffer
1464 Node* delay_slot = nullptr;
1465 for (uint i = 0; i < nblocks; i++) {
1466 Block* block = C->cfg()->get_block(i);
1467 _block = block;
1468 Node* head = block->head();
1469
1470 // If this block needs to start aligned (i.e, can be reached other
1471 // than by falling-thru from the previous block), then force the
1472 // start of a new bundle.
1473 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1474 masm->code()->flush_bundle(true);
1475 }
1476
1477 #ifdef ASSERT
1478 if (!block->is_connector()) {
1479 stringStream st;
1480 block->dump_head(C->cfg(), &st);
1481 masm->block_comment(st.freeze());
1482 }
1483 jmp_target[i] = 0;
1484 jmp_offset[i] = 0;
1485 jmp_size[i] = 0;
1486 jmp_rule[i] = 0;
1487 #endif
1488 int blk_offset = current_offset;
1489
1490 // Define the label at the beginning of the basic block
1491 masm->bind(blk_labels[block->_pre_order]);
1492
1493 uint last_inst = block->number_of_nodes();
1494
1495 // Emit block normally, except for last instruction.
1496 // Emit means "dump code bits into code buffer".
1497 for (uint j = 0; j<last_inst; j++) {
1498 _index = j;
1499
1500 // Get the node
1501 Node* n = block->get_node(j);
1502
1503 // See if delay slots are supported
1504 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1505 assert(delay_slot == nullptr, "no use of delay slot node");
1506 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1507
1508 delay_slot = n;
1509 continue;
1510 }
1511
1512 // If this starts a new instruction group, then flush the current one
1513 // (but allow split bundles)
1514 if (Pipeline::requires_bundling() && starts_bundle(n))
1515 masm->code()->flush_bundle(false);
1516
1517 // Special handling for SafePoint/Call Nodes
1518 bool is_mcall = false;
1519 if (n->is_Mach()) {
1520 MachNode *mach = n->as_Mach();
1521 is_mcall = n->is_MachCall();
1522 bool is_sfn = n->is_MachSafePoint();
1523
1524 // If this requires all previous instructions be flushed, then do so
1525 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1526 masm->code()->flush_bundle(true);
1527 current_offset = masm->offset();
1528 }
1529
1530 // A padding may be needed again since a previous instruction
1531 // could be moved to delay slot.
1532
1533 // align the instruction if necessary
1534 int padding = mach->compute_padding(current_offset);
1535 // Make sure safepoint node for polling is distinct from a call's
1536 // return by adding a nop if needed.
1537 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1538 padding = nop_size;
1539 }
1540 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1541 current_offset == last_avoid_back_to_back_offset) {
1542 // Avoid back to back some instructions.
1543 padding = nop_size;
1544 }
1545
1546 if (padding > 0) {
1547 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1548 int nops_cnt = padding / nop_size;
1549 MachNode *nop = new MachNopNode(nops_cnt);
1550 block->insert_node(nop, j++);
1551 last_inst++;
1552 C->cfg()->map_node_to_block(nop, block);
1553 // Ensure enough space.
1554 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1555 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1556 C->record_failure("CodeCache is full");
1557 return;
1558 }
1559 nop->emit(masm, C->regalloc());
1560 masm->code()->flush_bundle(true);
1561 current_offset = masm->offset();
1562 }
1563
1564 bool observe_safepoint = is_sfn;
1565 // Remember the start of the last call in a basic block
1566 if (is_mcall) {
1567 MachCallNode *mcall = mach->as_MachCall();
1568
1569 // This destination address is NOT PC-relative
1570 mcall->method_set((intptr_t)mcall->entry_point());
1571
1572 // Save the return address
1573 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1574
1575 observe_safepoint = mcall->guaranteed_safepoint();
1576 }
1577
1578 // sfn will be valid whenever mcall is valid now because of inheritance
1579 if (observe_safepoint) {
1580 // Handle special safepoint nodes for synchronization
1581 if (!is_mcall) {
1582 MachSafePointNode *sfn = mach->as_MachSafePoint();
1583 // !!!!! Stubs only need an oopmap right now, so bail out
1584 if (sfn->jvms()->method() == nullptr) {
1585 // Write the oopmap directly to the code blob??!!
1586 continue;
1587 }
1588 } // End synchronization
1589
1590 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1591 current_offset);
1592 Process_OopMap_Node(mach, current_offset);
1593 } // End if safepoint
1594
1595 // If this is a null check, then add the start of the previous instruction to the list
1596 else if( mach->is_MachNullCheck() ) {
1597 inct_starts[inct_cnt++] = previous_offset;
1598 }
1599
1600 // If this is a branch, then fill in the label with the target BB's label
1601 else if (mach->is_MachBranch()) {
1602 // This requires the TRUE branch target be in succs[0]
1603 uint block_num = block->non_connector_successor(0)->_pre_order;
1604
1605 // Try to replace long branch if delay slot is not used,
1606 // it is mostly for back branches since forward branch's
1607 // distance is not updated yet.
1608 bool delay_slot_is_used = valid_bundle_info(n) &&
1609 C->output()->node_bundling(n)->use_unconditional_delay();
1610 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1611 assert(delay_slot == nullptr, "not expecting delay slot node");
1612 int br_size = n->size(C->regalloc());
1613 int offset = blk_starts[block_num] - current_offset;
1614 if (block_num >= i) {
1615 // Current and following block's offset are not
1616 // finalized yet, adjust distance by the difference
1617 // between calculated and final offsets of current block.
1618 offset -= (blk_starts[i] - blk_offset);
1619 }
1620 // In the following code a nop could be inserted before
1621 // the branch which will increase the backward distance.
1622 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1623 if (needs_padding && offset <= 0)
1624 offset -= nop_size;
1625
1626 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1627 // We've got a winner. Replace this branch.
1628 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1629
1630 // Update the jmp_size.
1631 int new_size = replacement->size(C->regalloc());
1632 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1633 // Insert padding between avoid_back_to_back branches.
1634 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1635 MachNode *nop = new MachNopNode();
1636 block->insert_node(nop, j++);
1637 C->cfg()->map_node_to_block(nop, block);
1638 last_inst++;
1639 nop->emit(masm, C->regalloc());
1640 masm->code()->flush_bundle(true);
1641 current_offset = masm->offset();
1642 }
1643 #ifdef ASSERT
1644 jmp_target[i] = block_num;
1645 jmp_offset[i] = current_offset - blk_offset;
1646 jmp_size[i] = new_size;
1647 jmp_rule[i] = mach->rule();
1648 #endif
1649 block->map_node(replacement, j);
1650 mach->subsume_by(replacement, C);
1651 n = replacement;
1652 mach = replacement;
1653 }
1654 }
1655 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1656 } else if (mach->ideal_Opcode() == Op_Jump) {
1657 for (uint h = 0; h < block->_num_succs; h++) {
1658 Block* succs_block = block->_succs[h];
1659 for (uint j = 1; j < succs_block->num_preds(); j++) {
1660 Node* jpn = succs_block->pred(j);
1661 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1662 uint block_num = succs_block->non_connector()->_pre_order;
1663 Label *blkLabel = &blk_labels[block_num];
1664 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1665 }
1666 }
1667 }
1668 } else if (!n->is_Proj()) {
1669 // Remember the beginning of the previous instruction, in case
1670 // it's followed by a flag-kill and a null-check. Happens on
1671 // Intel all the time, with add-to-memory kind of opcodes.
1672 previous_offset = current_offset;
1673 }
1674
1675 // Not an else-if!
1676 // If this is a trap based cmp then add its offset to the list.
1677 if (mach->is_TrapBasedCheckNode()) {
1678 inct_starts[inct_cnt++] = current_offset;
1679 }
1680 }
1681
1682 // Verify that there is sufficient space remaining
1683 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1684 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1685 C->record_failure("CodeCache is full");
1686 return;
1687 }
1688
1689 // Save the offset for the listing
1690 #if defined(SUPPORT_OPTO_ASSEMBLY)
1691 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1692 node_offsets[n->_idx] = masm->offset();
1693 }
1694 #endif
1695 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1696
1697 // "Normal" instruction case
1698 DEBUG_ONLY(uint instr_offset = masm->offset());
1699 n->emit(masm, C->regalloc());
1700 current_offset = masm->offset();
1701
1702 // Above we only verified that there is enough space in the instruction section.
1703 // However, the instruction may emit stubs that cause code buffer expansion.
1704 // Bail out here if expansion failed due to a lack of code cache space.
1705 if (C->failing()) {
1706 return;
1707 }
1708
1709 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1710 "ret_addr_offset() not within emitted code");
1711
1712 #ifdef ASSERT
1713 uint n_size = n->size(C->regalloc());
1714 if (n_size < (current_offset-instr_offset)) {
1715 MachNode* mach = n->as_Mach();
1716 n->dump();
1717 mach->dump_format(C->regalloc(), tty);
1718 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1719 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1720 tty->print_cr(" ------------------- ");
1721 BufferBlob* blob = this->scratch_buffer_blob();
1722 address blob_begin = blob->content_begin();
1723 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1724 assert(false, "wrong size of mach node");
1725 }
1726 #endif
1727 non_safepoints.observe_instruction(n, current_offset);
1728
1729 // mcall is last "call" that can be a safepoint
1730 // record it so we can see if a poll will directly follow it
1731 // in which case we'll need a pad to make the PcDesc sites unique
1732 // see 5010568. This can be slightly inaccurate but conservative
1733 // in the case that return address is not actually at current_offset.
1734 // This is a small price to pay.
1735
1736 if (is_mcall) {
1737 last_call_offset = current_offset;
1738 }
1739
1740 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1741 // Avoid back to back some instructions.
1742 last_avoid_back_to_back_offset = current_offset;
1743 }
1744
1745 // See if this instruction has a delay slot
1746 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1747 guarantee(delay_slot != nullptr, "expecting delay slot node");
1748
1749 // Back up 1 instruction
1750 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1751
1752 // Save the offset for the listing
1753 #if defined(SUPPORT_OPTO_ASSEMBLY)
1754 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1755 node_offsets[delay_slot->_idx] = masm->offset();
1756 }
1757 #endif
1758
1759 // Support a SafePoint in the delay slot
1760 if (delay_slot->is_MachSafePoint()) {
1761 MachNode *mach = delay_slot->as_Mach();
1762 // !!!!! Stubs only need an oopmap right now, so bail out
1763 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1764 // Write the oopmap directly to the code blob??!!
1765 delay_slot = nullptr;
1766 continue;
1767 }
1768
1769 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1770 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1771 adjusted_offset);
1772 // Generate an OopMap entry
1773 Process_OopMap_Node(mach, adjusted_offset);
1774 }
1775
1776 // Insert the delay slot instruction
1777 delay_slot->emit(masm, C->regalloc());
1778
1779 // Don't reuse it
1780 delay_slot = nullptr;
1781 }
1782
1783 } // End for all instructions in block
1784
1785 // If the next block is the top of a loop, pad this block out to align
1786 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1787 if (i < nblocks-1) {
1788 Block *nb = C->cfg()->get_block(i + 1);
1789 int padding = nb->alignment_padding(current_offset);
1790 if( padding > 0 ) {
1791 MachNode *nop = new MachNopNode(padding / nop_size);
1792 block->insert_node(nop, block->number_of_nodes());
1793 C->cfg()->map_node_to_block(nop, block);
1794 nop->emit(masm, C->regalloc());
1795 current_offset = masm->offset();
1796 }
1797 }
1798 // Verify that the distance for generated before forward
1799 // short branches is still valid.
1800 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1801
1802 // Save new block start offset
1803 blk_starts[i] = blk_offset;
1804 } // End of for all blocks
1805 blk_starts[nblocks] = current_offset;
1806
1807 non_safepoints.flush_at_end();
1808
1809 // Offset too large?
1810 if (C->failing()) return;
1811
1812 // Define a pseudo-label at the end of the code
1813 masm->bind( blk_labels[nblocks] );
1814
1815 // Compute the size of the first block
1816 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1817
1818 #ifdef ASSERT
1819 for (uint i = 0; i < nblocks; i++) { // For all blocks
1820 if (jmp_target[i] != 0) {
1821 int br_size = jmp_size[i];
1822 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1823 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1824 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]);
1825 assert(false, "Displacement too large for short jmp");
1826 }
1827 }
1828 }
1829 #endif
1830
1831 if (!masm->code()->finalize_stubs()) {
1832 C->record_failure("CodeCache is full");
1833 return;
1834 }
1835
1836 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1837 bs->emit_stubs(*masm->code());
1838 if (C->failing()) return;
1839
1840 // Fill in stubs.
1841 assert(masm->inst_mark() == nullptr, "should be.");
1842 _stub_list.emit(*masm);
1843 if (C->failing()) return;
1844
1845 #ifndef PRODUCT
1846 // Information on the size of the method, without the extraneous code
1847 Scheduling::increment_method_size(masm->offset());
1848 #endif
1849
1850 // ------------------
1851 // Fill in exception table entries.
1852 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1853
1854 // Only java methods have exception handlers and deopt handlers
1855 // class HandlerImpl is platform-specific and defined in the *.ad files.
1856 if (C->method()) {
1857 // Emit the exception handler code.
1858 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1859 if (C->failing()) {
1860 return; // CodeBuffer::expand failed
1861 }
1862 // Emit the deopt handler code.
1863 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1864
1865 // Emit the MethodHandle deopt handler code (if required).
1866 if (C->has_method_handle_invokes() && !C->failing()) {
1867 // We can use the same code as for the normal deopt handler, we
1868 // just need a different entry point address.
1869 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1870 }
1871 }
1872
1873 // One last check for failed CodeBuffer::expand:
1874 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1875 C->record_failure("CodeCache is full");
1876 return;
1877 }
1878
1879 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1880 if (C->print_assembly()) {
1881 tty->cr();
1882 tty->print_cr("============================= C2-compiled nmethod ==============================");
1883 }
1884 #endif
1885
1886 #if defined(SUPPORT_OPTO_ASSEMBLY)
1887 // Dump the assembly code, including basic-block numbers
1888 if (C->print_assembly()) {
1889 ttyLocker ttyl; // keep the following output all in one block
1890 if (!VMThread::should_terminate()) { // test this under the tty lock
1891 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1892 // We call them first and write to a stringStream, then we retake the lock to
1893 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1894 ResourceMark rm;
1895 stringStream method_metadata_str;
1896 if (C->method() != nullptr) {
1897 C->method()->print_metadata(&method_metadata_str);
1898 }
1899 stringStream dump_asm_str;
1900 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1901
1902 NoSafepointVerifier nsv;
1903 ttyLocker ttyl2;
1904 // This output goes directly to the tty, not the compiler log.
1905 // To enable tools to match it up with the compilation activity,
1906 // be sure to tag this tty output with the compile ID.
1907 if (xtty != nullptr) {
1908 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1909 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1910 }
1911 if (C->method() != nullptr) {
1912 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1913 tty->print_raw(method_metadata_str.freeze());
1914 } else if (C->stub_name() != nullptr) {
1915 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1916 }
1917 tty->cr();
1918 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1919 tty->print_raw(dump_asm_str.freeze());
1920 tty->print_cr("--------------------------------------------------------------------------------");
1921 if (xtty != nullptr) {
1922 xtty->tail("opto_assembly");
1923 }
1924 }
1925 }
1926 #endif
1927 }
1928
1929 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1930 _inc_table.set_size(cnt);
1931
1932 uint inct_cnt = 0;
1933 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1934 Block* block = C->cfg()->get_block(i);
1935 Node *n = nullptr;
1936 int j;
1937
1938 // Find the branch; ignore trailing NOPs.
1939 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1940 n = block->get_node(j);
1941 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1942 break;
1943 }
1944 }
1945
1946 // If we didn't find anything, continue
1947 if (j < 0) {
1948 continue;
1949 }
1950
1951 // Compute ExceptionHandlerTable subtable entry and add it
1952 // (skip empty blocks)
1953 if (n->is_Catch()) {
1954
1955 // Get the offset of the return from the call
1956 uint call_return = call_returns[block->_pre_order];
1957 #ifdef ASSERT
1958 assert( call_return > 0, "no call seen for this basic block" );
1959 while (block->get_node(--j)->is_MachProj()) ;
1960 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1961 #endif
1962 // last instruction is a CatchNode, find it's CatchProjNodes
1963 int nof_succs = block->_num_succs;
1964 // allocate space
1965 GrowableArray<intptr_t> handler_bcis(nof_succs);
1966 GrowableArray<intptr_t> handler_pcos(nof_succs);
1967 // iterate through all successors
1968 for (int j = 0; j < nof_succs; j++) {
1969 Block* s = block->_succs[j];
1970 bool found_p = false;
1971 for (uint k = 1; k < s->num_preds(); k++) {
1972 Node* pk = s->pred(k);
1973 if (pk->is_CatchProj() && pk->in(0) == n) {
1974 const CatchProjNode* p = pk->as_CatchProj();
1975 found_p = true;
1976 // add the corresponding handler bci & pco information
1977 if (p->_con != CatchProjNode::fall_through_index) {
1978 // p leads to an exception handler (and is not fall through)
1979 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1980 // no duplicates, please
1981 if (!handler_bcis.contains(p->handler_bci())) {
1982 uint block_num = s->non_connector()->_pre_order;
1983 handler_bcis.append(p->handler_bci());
1984 handler_pcos.append(blk_labels[block_num].loc_pos());
1985 }
1986 }
1987 }
1988 }
1989 assert(found_p, "no matching predecessor found");
1990 // Note: Due to empty block removal, one block may have
1991 // several CatchProj inputs, from the same Catch.
1992 }
1993
1994 // Set the offset of the return from the call
1995 assert(handler_bcis.find(-1) != -1, "must have default handler");
1996 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
1997 continue;
1998 }
1999
2000 // Handle implicit null exception table updates
2001 if (n->is_MachNullCheck()) {
2002 assert(n->in(1)->as_Mach()->barrier_data() == 0,
2003 "Implicit null checks on memory accesses with barriers are not yet supported");
2004 uint block_num = block->non_connector_successor(0)->_pre_order;
2005 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2006 continue;
2007 }
2008 // Handle implicit exception table updates: trap instructions.
2009 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2010 uint block_num = block->non_connector_successor(0)->_pre_order;
2011 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2012 continue;
2013 }
2014 } // End of for all blocks fill in exception table entries
2015 }
2016
2017 // Static Variables
2018 #ifndef PRODUCT
2019 uint Scheduling::_total_nop_size = 0;
2020 uint Scheduling::_total_method_size = 0;
2021 uint Scheduling::_total_branches = 0;
2022 uint Scheduling::_total_unconditional_delays = 0;
2023 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2024 #endif
2025
2026 // Initializer for class Scheduling
2027
2028 Scheduling::Scheduling(Arena *arena, Compile &compile)
2029 : _arena(arena),
2030 _cfg(compile.cfg()),
2031 _regalloc(compile.regalloc()),
2032 _scheduled(arena),
2033 _available(arena),
2034 _reg_node(arena),
2035 _pinch_free_list(arena),
2036 _next_node(nullptr),
2037 _bundle_instr_count(0),
2038 _bundle_cycle_number(0),
2039 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2040 #ifndef PRODUCT
2041 , _branches(0)
2042 , _unconditional_delays(0)
2043 #endif
2044 {
2045 // Create a MachNopNode
2046 _nop = new MachNopNode();
2047
2048 // Now that the nops are in the array, save the count
2049 // (but allow entries for the nops)
2050 _node_bundling_limit = compile.unique();
2051 uint node_max = _regalloc->node_regs_max_index();
2052
2053 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2054
2055 // This one is persistent within the Compile class
2056 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2057
2058 // Allocate space for fixed-size arrays
2059 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2060 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2061
2062 // Clear the arrays
2063 for (uint i = 0; i < node_max; i++) {
2064 ::new (&_node_bundling_base[i]) Bundle();
2065 }
2066 memset(_uses, 0, node_max * sizeof(short));
2067 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2068
2069 // Clear the bundling information
2070 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2071
2072 // Get the last node
2073 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2074
2075 _next_node = block->get_node(block->number_of_nodes() - 1);
2076 }
2077
2078 #ifndef PRODUCT
2079 // Scheduling destructor
2080 Scheduling::~Scheduling() {
2081 _total_branches += _branches;
2082 _total_unconditional_delays += _unconditional_delays;
2083 }
2084 #endif
2085
2086 // Step ahead "i" cycles
2087 void Scheduling::step(uint i) {
2088
2089 Bundle *bundle = node_bundling(_next_node);
2090 bundle->set_starts_bundle();
2091
2092 // Update the bundle record, but leave the flags information alone
2093 if (_bundle_instr_count > 0) {
2094 bundle->set_instr_count(_bundle_instr_count);
2095 bundle->set_resources_used(_bundle_use.resourcesUsed());
2096 }
2097
2098 // Update the state information
2099 _bundle_instr_count = 0;
2100 _bundle_cycle_number += i;
2101 _bundle_use.step(i);
2102 }
2103
2104 void Scheduling::step_and_clear() {
2105 Bundle *bundle = node_bundling(_next_node);
2106 bundle->set_starts_bundle();
2107
2108 // Update the bundle record
2109 if (_bundle_instr_count > 0) {
2110 bundle->set_instr_count(_bundle_instr_count);
2111 bundle->set_resources_used(_bundle_use.resourcesUsed());
2112
2113 _bundle_cycle_number += 1;
2114 }
2115
2116 // Clear the bundling information
2117 _bundle_instr_count = 0;
2118 _bundle_use.reset();
2119
2120 memcpy(_bundle_use_elements,
2121 Pipeline_Use::elaborated_elements,
2122 sizeof(Pipeline_Use::elaborated_elements));
2123 }
2124
2125 // Perform instruction scheduling and bundling over the sequence of
2126 // instructions in backwards order.
2127 void PhaseOutput::ScheduleAndBundle() {
2128
2129 // Don't optimize this if it isn't a method
2130 if (!C->method())
2131 return;
2132
2133 // Don't optimize this if scheduling is disabled
2134 if (!C->do_scheduling())
2135 return;
2136
2137 // Scheduling code works only with pairs (8 bytes) maximum.
2138 // And when the scalable vector register is used, we may spill/unspill
2139 // the whole reg regardless of the max vector size.
2140 if (C->max_vector_size() > 8 ||
2141 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2142 return;
2143 }
2144
2145 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2146
2147 // Create a data structure for all the scheduling information
2148 Scheduling scheduling(Thread::current()->resource_area(), *C);
2149
2150 // Walk backwards over each basic block, computing the needed alignment
2151 // Walk over all the basic blocks
2152 scheduling.DoScheduling();
2153
2154 #ifndef PRODUCT
2155 if (C->trace_opto_output()) {
2156 // Buffer and print all at once
2157 ResourceMark rm;
2158 stringStream ss;
2159 ss.print("\n---- After ScheduleAndBundle ----\n");
2160 print_scheduling(&ss);
2161 tty->print("%s", ss.as_string());
2162 }
2163 #endif
2164 }
2165
2166 #ifndef PRODUCT
2167 // Separated out so that it can be called directly from debugger
2168 void PhaseOutput::print_scheduling() {
2169 print_scheduling(tty);
2170 }
2171
2172 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2173 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2174 output_stream->print("\nBB#%03d:\n", i);
2175 Block* block = C->cfg()->get_block(i);
2176 for (uint j = 0; j < block->number_of_nodes(); j++) {
2177 Node* n = block->get_node(j);
2178 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2179 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2180 n->dump("\n", false, output_stream);
2181 }
2182 }
2183 }
2184 #endif
2185
2186 // See if this node fits into the present instruction bundle
2187 bool Scheduling::NodeFitsInBundle(Node *n) {
2188 uint n_idx = n->_idx;
2189
2190 // If this is the unconditional delay instruction, then it fits
2191 if (n == _unconditional_delay_slot) {
2192 #ifndef PRODUCT
2193 if (_cfg->C->trace_opto_output())
2194 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2195 #endif
2196 return (true);
2197 }
2198
2199 // If the node cannot be scheduled this cycle, skip it
2200 if (_current_latency[n_idx] > _bundle_cycle_number) {
2201 #ifndef PRODUCT
2202 if (_cfg->C->trace_opto_output())
2203 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2204 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2205 #endif
2206 return (false);
2207 }
2208
2209 const Pipeline *node_pipeline = n->pipeline();
2210
2211 uint instruction_count = node_pipeline->instructionCount();
2212 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2213 instruction_count = 0;
2214 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2215 instruction_count++;
2216
2217 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2218 #ifndef PRODUCT
2219 if (_cfg->C->trace_opto_output())
2220 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2221 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2222 #endif
2223 return (false);
2224 }
2225
2226 // Don't allow non-machine nodes to be handled this way
2227 if (!n->is_Mach() && instruction_count == 0)
2228 return (false);
2229
2230 // See if there is any overlap
2231 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2232
2233 if (delay > 0) {
2234 #ifndef PRODUCT
2235 if (_cfg->C->trace_opto_output())
2236 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2237 #endif
2238 return false;
2239 }
2240
2241 #ifndef PRODUCT
2242 if (_cfg->C->trace_opto_output())
2243 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2244 #endif
2245
2246 return true;
2247 }
2248
2249 Node * Scheduling::ChooseNodeToBundle() {
2250 uint siz = _available.size();
2251
2252 if (siz == 0) {
2253
2254 #ifndef PRODUCT
2255 if (_cfg->C->trace_opto_output())
2256 tty->print("# ChooseNodeToBundle: null\n");
2257 #endif
2258 return (nullptr);
2259 }
2260
2261 // Fast path, if only 1 instruction in the bundle
2262 if (siz == 1) {
2263 #ifndef PRODUCT
2264 if (_cfg->C->trace_opto_output()) {
2265 tty->print("# ChooseNodeToBundle (only 1): ");
2266 _available[0]->dump();
2267 }
2268 #endif
2269 return (_available[0]);
2270 }
2271
2272 // Don't bother, if the bundle is already full
2273 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2274 for ( uint i = 0; i < siz; i++ ) {
2275 Node *n = _available[i];
2276
2277 // Skip projections, we'll handle them another way
2278 if (n->is_Proj())
2279 continue;
2280
2281 // This presupposed that instructions are inserted into the
2282 // available list in a legality order; i.e. instructions that
2283 // must be inserted first are at the head of the list
2284 if (NodeFitsInBundle(n)) {
2285 #ifndef PRODUCT
2286 if (_cfg->C->trace_opto_output()) {
2287 tty->print("# ChooseNodeToBundle: ");
2288 n->dump();
2289 }
2290 #endif
2291 return (n);
2292 }
2293 }
2294 }
2295
2296 // Nothing fits in this bundle, choose the highest priority
2297 #ifndef PRODUCT
2298 if (_cfg->C->trace_opto_output()) {
2299 tty->print("# ChooseNodeToBundle: ");
2300 _available[0]->dump();
2301 }
2302 #endif
2303
2304 return _available[0];
2305 }
2306
2307 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2308 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2309
2310 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2311 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2312 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2313 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2314
2315 // Comparison between stack -> reg and stack -> reg
2316 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2317 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2318 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2319 }
2320
2321 // Comparison between reg -> stack and reg -> stack
2322 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2323 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2324 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2325 }
2326
2327 return 0; // Not comparable
2328 }
2329
2330 void Scheduling::AddNodeToAvailableList(Node *n) {
2331 assert( !n->is_Proj(), "projections never directly made available" );
2332 #ifndef PRODUCT
2333 if (_cfg->C->trace_opto_output()) {
2334 tty->print("# AddNodeToAvailableList: ");
2335 n->dump();
2336 }
2337 #endif
2338
2339 int latency = _current_latency[n->_idx];
2340
2341 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2342 // for stack spilling or unspilling have the same latency, we sort
2343 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2344 // have more opportunities to do ld/st merging
2345 uint i;
2346 for (i = 0; i < _available.size(); i++) {
2347 if (_current_latency[_available[i]->_idx] > latency) {
2348 break;
2349 } else if (_current_latency[_available[i]->_idx] == latency &&
2350 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2351 compare_two_spill_nodes(n, _available[i]) > 0) {
2352 break;
2353 }
2354 }
2355
2356 // Special Check for compares following branches
2357 if( n->is_Mach() && _scheduled.size() > 0 ) {
2358 int op = n->as_Mach()->ideal_Opcode();
2359 Node *last = _scheduled[0];
2360 if( last->is_MachIf() && last->in(1) == n &&
2361 ( op == Op_CmpI ||
2362 op == Op_CmpU ||
2363 op == Op_CmpUL ||
2364 op == Op_CmpP ||
2365 op == Op_CmpF ||
2366 op == Op_CmpD ||
2367 op == Op_CmpL ) ) {
2368
2369 // Recalculate position, moving to front of same latency
2370 for ( i=0 ; i < _available.size(); i++ )
2371 if (_current_latency[_available[i]->_idx] >= latency)
2372 break;
2373 }
2374 }
2375
2376 // Insert the node in the available list
2377 _available.insert(i, n);
2378
2379 #ifndef PRODUCT
2380 if (_cfg->C->trace_opto_output())
2381 dump_available();
2382 #endif
2383 }
2384
2385 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2386 for ( uint i=0; i < n->len(); i++ ) {
2387 Node *def = n->in(i);
2388 if (!def) continue;
2389 if( def->is_Proj() ) // If this is a machine projection, then
2390 def = def->in(0); // propagate usage thru to the base instruction
2391
2392 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2393 continue;
2394 }
2395
2396 // Compute the latency
2397 uint l = _bundle_cycle_number + n->latency(i);
2398 if (_current_latency[def->_idx] < l)
2399 _current_latency[def->_idx] = l;
2400
2401 // If this does not have uses then schedule it
2402 if ((--_uses[def->_idx]) == 0)
2403 AddNodeToAvailableList(def);
2404 }
2405 }
2406
2407 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2408 #ifndef PRODUCT
2409 if (_cfg->C->trace_opto_output()) {
2410 tty->print("# AddNodeToBundle: ");
2411 n->dump();
2412 }
2413 #endif
2414
2415 // Remove this from the available list
2416 uint i;
2417 for (i = 0; i < _available.size(); i++)
2418 if (_available[i] == n)
2419 break;
2420 assert(i < _available.size(), "entry in _available list not found");
2421 _available.remove(i);
2422
2423 // See if this fits in the current bundle
2424 const Pipeline *node_pipeline = n->pipeline();
2425 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2426
2427 // Check for instructions to be placed in the delay slot. We
2428 // do this before we actually schedule the current instruction,
2429 // because the delay slot follows the current instruction.
2430 if (Pipeline::_branch_has_delay_slot &&
2431 node_pipeline->hasBranchDelay() &&
2432 !_unconditional_delay_slot) {
2433
2434 uint siz = _available.size();
2435
2436 // Conditional branches can support an instruction that
2437 // is unconditionally executed and not dependent by the
2438 // branch, OR a conditionally executed instruction if
2439 // the branch is taken. In practice, this means that
2440 // the first instruction at the branch target is
2441 // copied to the delay slot, and the branch goes to
2442 // the instruction after that at the branch target
2443 if ( n->is_MachBranch() ) {
2444
2445 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2446 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2447
2448 #ifndef PRODUCT
2449 _branches++;
2450 #endif
2451
2452 // At least 1 instruction is on the available list
2453 // that is not dependent on the branch
2454 for (uint i = 0; i < siz; i++) {
2455 Node *d = _available[i];
2456 const Pipeline *avail_pipeline = d->pipeline();
2457
2458 // Don't allow safepoints in the branch shadow, that will
2459 // cause a number of difficulties
2460 if ( avail_pipeline->instructionCount() == 1 &&
2461 !avail_pipeline->hasMultipleBundles() &&
2462 !avail_pipeline->hasBranchDelay() &&
2463 Pipeline::instr_has_unit_size() &&
2464 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2465 NodeFitsInBundle(d) &&
2466 !node_bundling(d)->used_in_delay()) {
2467
2468 if (d->is_Mach() && !d->is_MachSafePoint()) {
2469 // A node that fits in the delay slot was found, so we need to
2470 // set the appropriate bits in the bundle pipeline information so
2471 // that it correctly indicates resource usage. Later, when we
2472 // attempt to add this instruction to the bundle, we will skip
2473 // setting the resource usage.
2474 _unconditional_delay_slot = d;
2475 node_bundling(n)->set_use_unconditional_delay();
2476 node_bundling(d)->set_used_in_unconditional_delay();
2477 _bundle_use.add_usage(avail_pipeline->resourceUse());
2478 _current_latency[d->_idx] = _bundle_cycle_number;
2479 _next_node = d;
2480 ++_bundle_instr_count;
2481 #ifndef PRODUCT
2482 _unconditional_delays++;
2483 #endif
2484 break;
2485 }
2486 }
2487 }
2488 }
2489
2490 // No delay slot, add a nop to the usage
2491 if (!_unconditional_delay_slot) {
2492 // See if adding an instruction in the delay slot will overflow
2493 // the bundle.
2494 if (!NodeFitsInBundle(_nop)) {
2495 #ifndef PRODUCT
2496 if (_cfg->C->trace_opto_output())
2497 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2498 #endif
2499 step(1);
2500 }
2501
2502 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2503 _next_node = _nop;
2504 ++_bundle_instr_count;
2505 }
2506
2507 // See if the instruction in the delay slot requires a
2508 // step of the bundles
2509 if (!NodeFitsInBundle(n)) {
2510 #ifndef PRODUCT
2511 if (_cfg->C->trace_opto_output())
2512 tty->print("# *** STEP(branch won't fit) ***\n");
2513 #endif
2514 // Update the state information
2515 _bundle_instr_count = 0;
2516 _bundle_cycle_number += 1;
2517 _bundle_use.step(1);
2518 }
2519 }
2520
2521 // Get the number of instructions
2522 uint instruction_count = node_pipeline->instructionCount();
2523 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2524 instruction_count = 0;
2525
2526 // Compute the latency information
2527 uint delay = 0;
2528
2529 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2530 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2531 if (relative_latency < 0)
2532 relative_latency = 0;
2533
2534 delay = _bundle_use.full_latency(relative_latency, node_usage);
2535
2536 // Does not fit in this bundle, start a new one
2537 if (delay > 0) {
2538 step(delay);
2539
2540 #ifndef PRODUCT
2541 if (_cfg->C->trace_opto_output())
2542 tty->print("# *** STEP(%d) ***\n", delay);
2543 #endif
2544 }
2545 }
2546
2547 // If this was placed in the delay slot, ignore it
2548 if (n != _unconditional_delay_slot) {
2549
2550 if (delay == 0) {
2551 if (node_pipeline->hasMultipleBundles()) {
2552 #ifndef PRODUCT
2553 if (_cfg->C->trace_opto_output())
2554 tty->print("# *** STEP(multiple instructions) ***\n");
2555 #endif
2556 step(1);
2557 }
2558
2559 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2560 #ifndef PRODUCT
2561 if (_cfg->C->trace_opto_output())
2562 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2563 instruction_count + _bundle_instr_count,
2564 Pipeline::_max_instrs_per_cycle);
2565 #endif
2566 step(1);
2567 }
2568 }
2569
2570 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2571 _bundle_instr_count++;
2572
2573 // Set the node's latency
2574 _current_latency[n->_idx] = _bundle_cycle_number;
2575
2576 // Now merge the functional unit information
2577 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2578 _bundle_use.add_usage(node_usage);
2579
2580 // Increment the number of instructions in this bundle
2581 _bundle_instr_count += instruction_count;
2582
2583 // Remember this node for later
2584 if (n->is_Mach())
2585 _next_node = n;
2586 }
2587
2588 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2589 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2590 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2591 // into the block. All other scheduled nodes get put in the schedule here.
2592 int op = n->Opcode();
2593 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2594 (op != Op_Node && // Not an unused antidepedence node and
2595 // not an unallocated boxlock
2596 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2597
2598 // Push any trailing projections
2599 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2600 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2601 Node *foi = n->fast_out(i);
2602 if( foi->is_Proj() )
2603 _scheduled.push(foi);
2604 }
2605 }
2606
2607 // Put the instruction in the schedule list
2608 _scheduled.push(n);
2609 }
2610
2611 #ifndef PRODUCT
2612 if (_cfg->C->trace_opto_output())
2613 dump_available();
2614 #endif
2615
2616 // Walk all the definitions, decrementing use counts, and
2617 // if a definition has a 0 use count, place it in the available list.
2618 DecrementUseCounts(n,bb);
2619 }
2620
2621 // This method sets the use count within a basic block. We will ignore all
2622 // uses outside the current basic block. As we are doing a backwards walk,
2623 // any node we reach that has a use count of 0 may be scheduled. This also
2624 // avoids the problem of cyclic references from phi nodes, as long as phi
2625 // nodes are at the front of the basic block. This method also initializes
2626 // the available list to the set of instructions that have no uses within this
2627 // basic block.
2628 void Scheduling::ComputeUseCount(const Block *bb) {
2629 #ifndef PRODUCT
2630 if (_cfg->C->trace_opto_output())
2631 tty->print("# -> ComputeUseCount\n");
2632 #endif
2633
2634 // Clear the list of available and scheduled instructions, just in case
2635 _available.clear();
2636 _scheduled.clear();
2637
2638 // No delay slot specified
2639 _unconditional_delay_slot = nullptr;
2640
2641 #ifdef ASSERT
2642 for( uint i=0; i < bb->number_of_nodes(); i++ )
2643 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2644 #endif
2645
2646 // Force the _uses count to never go to zero for unscheduable pieces
2647 // of the block
2648 for( uint k = 0; k < _bb_start; k++ )
2649 _uses[bb->get_node(k)->_idx] = 1;
2650 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2651 _uses[bb->get_node(l)->_idx] = 1;
2652
2653 // Iterate backwards over the instructions in the block. Don't count the
2654 // branch projections at end or the block header instructions.
2655 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2656 Node *n = bb->get_node(j);
2657 if( n->is_Proj() ) continue; // Projections handled another way
2658
2659 // Account for all uses
2660 for ( uint k = 0; k < n->len(); k++ ) {
2661 Node *inp = n->in(k);
2662 if (!inp) continue;
2663 assert(inp != n, "no cycles allowed" );
2664 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2665 if (inp->is_Proj()) { // Skip through Proj's
2666 inp = inp->in(0);
2667 }
2668 ++_uses[inp->_idx]; // Count 1 block-local use
2669 }
2670 }
2671
2672 // If this instruction has a 0 use count, then it is available
2673 if (!_uses[n->_idx]) {
2674 _current_latency[n->_idx] = _bundle_cycle_number;
2675 AddNodeToAvailableList(n);
2676 }
2677
2678 #ifndef PRODUCT
2679 if (_cfg->C->trace_opto_output()) {
2680 tty->print("# uses: %3d: ", _uses[n->_idx]);
2681 n->dump();
2682 }
2683 #endif
2684 }
2685
2686 #ifndef PRODUCT
2687 if (_cfg->C->trace_opto_output())
2688 tty->print("# <- ComputeUseCount\n");
2689 #endif
2690 }
2691
2692 // This routine performs scheduling on each basic block in reverse order,
2693 // using instruction latencies and taking into account function unit
2694 // availability.
2695 void Scheduling::DoScheduling() {
2696 #ifndef PRODUCT
2697 if (_cfg->C->trace_opto_output())
2698 tty->print("# -> DoScheduling\n");
2699 #endif
2700
2701 Block *succ_bb = nullptr;
2702 Block *bb;
2703 Compile* C = Compile::current();
2704
2705 // Walk over all the basic blocks in reverse order
2706 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2707 bb = _cfg->get_block(i);
2708
2709 #ifndef PRODUCT
2710 if (_cfg->C->trace_opto_output()) {
2711 tty->print("# Schedule BB#%03d (initial)\n", i);
2712 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2713 bb->get_node(j)->dump();
2714 }
2715 }
2716 #endif
2717
2718 // On the head node, skip processing
2719 if (bb == _cfg->get_root_block()) {
2720 continue;
2721 }
2722
2723 // Skip empty, connector blocks
2724 if (bb->is_connector())
2725 continue;
2726
2727 // If the following block is not the sole successor of
2728 // this one, then reset the pipeline information
2729 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2730 #ifndef PRODUCT
2731 if (_cfg->C->trace_opto_output()) {
2732 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2733 _next_node->_idx, _bundle_instr_count);
2734 }
2735 #endif
2736 step_and_clear();
2737 }
2738
2739 // Leave untouched the starting instruction, any Phis, a CreateEx node
2740 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2741 _bb_end = bb->number_of_nodes()-1;
2742 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2743 Node *n = bb->get_node(_bb_start);
2744 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2745 // Also, MachIdealNodes do not get scheduled
2746 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2747 MachNode *mach = n->as_Mach();
2748 int iop = mach->ideal_Opcode();
2749 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2750 if( iop == Op_Con ) continue; // Do not schedule Top
2751 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2752 mach->pipeline() == MachNode::pipeline_class() &&
2753 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2754 continue;
2755 break; // Funny loop structure to be sure...
2756 }
2757 // Compute last "interesting" instruction in block - last instruction we
2758 // might schedule. _bb_end points just after last schedulable inst. We
2759 // normally schedule conditional branches (despite them being forced last
2760 // in the block), because they have delay slots we can fill. Calls all
2761 // have their delay slots filled in the template expansions, so we don't
2762 // bother scheduling them.
2763 Node *last = bb->get_node(_bb_end);
2764 // Ignore trailing NOPs.
2765 while (_bb_end > 0 && last->is_Mach() &&
2766 last->as_Mach()->ideal_Opcode() == Op_Con) {
2767 last = bb->get_node(--_bb_end);
2768 }
2769 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2770 if( last->is_Catch() ||
2771 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2772 // There might be a prior call. Skip it.
2773 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2774 } else if( last->is_MachNullCheck() ) {
2775 // Backup so the last null-checked memory instruction is
2776 // outside the schedulable range. Skip over the nullcheck,
2777 // projection, and the memory nodes.
2778 Node *mem = last->in(1);
2779 do {
2780 _bb_end--;
2781 } while (mem != bb->get_node(_bb_end));
2782 } else {
2783 // Set _bb_end to point after last schedulable inst.
2784 _bb_end++;
2785 }
2786
2787 assert( _bb_start <= _bb_end, "inverted block ends" );
2788
2789 // Compute the register antidependencies for the basic block
2790 ComputeRegisterAntidependencies(bb);
2791 if (C->failing()) return; // too many D-U pinch points
2792
2793 // Compute the usage within the block, and set the list of all nodes
2794 // in the block that have no uses within the block.
2795 ComputeUseCount(bb);
2796
2797 // Schedule the remaining instructions in the block
2798 while ( _available.size() > 0 ) {
2799 Node *n = ChooseNodeToBundle();
2800 guarantee(n != nullptr, "no nodes available");
2801 AddNodeToBundle(n,bb);
2802 }
2803
2804 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2805 #ifdef ASSERT
2806 for( uint l = _bb_start; l < _bb_end; l++ ) {
2807 Node *n = bb->get_node(l);
2808 uint m;
2809 for( m = 0; m < _bb_end-_bb_start; m++ )
2810 if( _scheduled[m] == n )
2811 break;
2812 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2813 }
2814 #endif
2815
2816 // Now copy the instructions (in reverse order) back to the block
2817 for ( uint k = _bb_start; k < _bb_end; k++ )
2818 bb->map_node(_scheduled[_bb_end-k-1], k);
2819
2820 #ifndef PRODUCT
2821 if (_cfg->C->trace_opto_output()) {
2822 tty->print("# Schedule BB#%03d (final)\n", i);
2823 uint current = 0;
2824 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2825 Node *n = bb->get_node(j);
2826 if( valid_bundle_info(n) ) {
2827 Bundle *bundle = node_bundling(n);
2828 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2829 tty->print("*** Bundle: ");
2830 bundle->dump();
2831 }
2832 n->dump();
2833 }
2834 }
2835 }
2836 #endif
2837 #ifdef ASSERT
2838 verify_good_schedule(bb,"after block local scheduling");
2839 #endif
2840 }
2841
2842 #ifndef PRODUCT
2843 if (_cfg->C->trace_opto_output())
2844 tty->print("# <- DoScheduling\n");
2845 #endif
2846
2847 // Record final node-bundling array location
2848 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2849
2850 } // end DoScheduling
2851
2852 // Verify that no live-range used in the block is killed in the block by a
2853 // wrong DEF. This doesn't verify live-ranges that span blocks.
2854
2855 // Check for edge existence. Used to avoid adding redundant precedence edges.
2856 static bool edge_from_to( Node *from, Node *to ) {
2857 for( uint i=0; i<from->len(); i++ )
2858 if( from->in(i) == to )
2859 return true;
2860 return false;
2861 }
2862
2863 #ifdef ASSERT
2864 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2865 // Check for bad kills
2866 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2867 Node *prior_use = _reg_node[def];
2868 if( prior_use && !edge_from_to(prior_use,n) ) {
2869 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2870 n->dump();
2871 tty->print_cr("...");
2872 prior_use->dump();
2873 assert(edge_from_to(prior_use,n), "%s", msg);
2874 }
2875 _reg_node.map(def,nullptr); // Kill live USEs
2876 }
2877 }
2878
2879 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2880
2881 // Zap to something reasonable for the verify code
2882 _reg_node.clear();
2883
2884 // Walk over the block backwards. Check to make sure each DEF doesn't
2885 // kill a live value (other than the one it's supposed to). Add each
2886 // USE to the live set.
2887 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2888 Node *n = b->get_node(i);
2889 int n_op = n->Opcode();
2890 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2891 // Fat-proj kills a slew of registers
2892 RegMaskIterator rmi(n->out_RegMask());
2893 while (rmi.has_next()) {
2894 OptoReg::Name kill = rmi.next();
2895 verify_do_def(n, kill, msg);
2896 }
2897 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2898 // Get DEF'd registers the normal way
2899 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2900 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2901 }
2902
2903 // Now make all USEs live
2904 for( uint i=1; i<n->req(); i++ ) {
2905 Node *def = n->in(i);
2906 assert(def != nullptr, "input edge required");
2907 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2908 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2909 if( OptoReg::is_valid(reg_lo) ) {
2910 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2911 _reg_node.map(reg_lo,n);
2912 }
2913 if( OptoReg::is_valid(reg_hi) ) {
2914 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2915 _reg_node.map(reg_hi,n);
2916 }
2917 }
2918
2919 }
2920
2921 // Zap to something reasonable for the Antidependence code
2922 _reg_node.clear();
2923 }
2924 #endif
2925
2926 // Conditionally add precedence edges. Avoid putting edges on Projs.
2927 static void add_prec_edge_from_to( Node *from, Node *to ) {
2928 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2929 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2930 from = from->in(0);
2931 }
2932 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2933 !edge_from_to( from, to ) ) // Avoid duplicate edge
2934 from->add_prec(to);
2935 }
2936
2937 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2938 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2939 return;
2940
2941 if (OptoReg::is_reg(def_reg)) {
2942 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2943 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2944 // This is one of the high slots of a vector register.
2945 // ScheduleAndBundle already checked there are no live wide
2946 // vectors in this method so it can be safely ignored.
2947 return;
2948 }
2949 }
2950
2951 Node *pinch = _reg_node[def_reg]; // Get pinch point
2952 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2953 is_def ) { // Check for a true def (not a kill)
2954 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2955 return;
2956 }
2957
2958 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2959 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2960
2961 // After some number of kills there _may_ be a later def
2962 Node *later_def = nullptr;
2963
2964 Compile* C = Compile::current();
2965
2966 // Finding a kill requires a real pinch-point.
2967 // Check for not already having a pinch-point.
2968 // Pinch points are Op_Node's.
2969 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2970 later_def = pinch; // Must be def/kill as optimistic pinch-point
2971 if ( _pinch_free_list.size() > 0) {
2972 pinch = _pinch_free_list.pop();
2973 } else {
2974 pinch = new Node(1); // Pinch point to-be
2975 }
2976 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2977 DEBUG_ONLY( pinch->dump(); );
2978 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2979 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2980 return;
2981 }
2982 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2983 _reg_node.map(def_reg,pinch); // Record pinch-point
2984 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2985 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2986 pinch->init_req(0, C->top()); // set not null for the next call
2987 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2988 later_def = nullptr; // and no later def
2989 }
2990 pinch->set_req(0,later_def); // Hook later def so we can find it
2991 } else { // Else have valid pinch point
2992 if( pinch->in(0) ) // If there is a later-def
2993 later_def = pinch->in(0); // Get it
2994 }
2995
2996 // Add output-dependence edge from later def to kill
2997 if( later_def ) // If there is some original def
2998 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2999
3000 // See if current kill is also a use, and so is forced to be the pinch-point.
3001 if( pinch->Opcode() == Op_Node ) {
3002 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3003 for( uint i=1; i<uses->req(); i++ ) {
3004 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3005 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3006 // Yes, found a use/kill pinch-point
3007 pinch->set_req(0,nullptr); //
3008 pinch->replace_by(kill); // Move anti-dep edges up
3009 pinch = kill;
3010 _reg_node.map(def_reg,pinch);
3011 return;
3012 }
3013 }
3014 }
3015
3016 // Add edge from kill to pinch-point
3017 add_prec_edge_from_to(kill,pinch);
3018 }
3019
3020 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3021 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3022 return;
3023 Node *pinch = _reg_node[use_reg]; // Get pinch point
3024 // Check for no later def_reg/kill in block
3025 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3026 // Use has to be block-local as well
3027 _cfg->get_block_for_node(use) == b) {
3028 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3029 pinch->req() == 1 ) { // pinch not yet in block?
3030 pinch->del_req(0); // yank pointer to later-def, also set flag
3031 // Insert the pinch-point in the block just after the last use
3032 b->insert_node(pinch, b->find_node(use) + 1);
3033 _bb_end++; // Increase size scheduled region in block
3034 }
3035
3036 add_prec_edge_from_to(pinch,use);
3037 }
3038 }
3039
3040 // We insert antidependences between the reads and following write of
3041 // allocated registers to prevent illegal code motion. Hopefully, the
3042 // number of added references should be fairly small, especially as we
3043 // are only adding references within the current basic block.
3044 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3045
3046 #ifdef ASSERT
3047 verify_good_schedule(b,"before block local scheduling");
3048 #endif
3049
3050 // A valid schedule, for each register independently, is an endless cycle
3051 // of: a def, then some uses (connected to the def by true dependencies),
3052 // then some kills (defs with no uses), finally the cycle repeats with a new
3053 // def. The uses are allowed to float relative to each other, as are the
3054 // kills. No use is allowed to slide past a kill (or def). This requires
3055 // antidependencies between all uses of a single def and all kills that
3056 // follow, up to the next def. More edges are redundant, because later defs
3057 // & kills are already serialized with true or antidependencies. To keep
3058 // the edge count down, we add a 'pinch point' node if there's more than
3059 // one use or more than one kill/def.
3060
3061 // We add dependencies in one bottom-up pass.
3062
3063 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3064
3065 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3066 // register. If not, we record the DEF/KILL in _reg_node, the
3067 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3068 // "pinch point", a new Node that's in the graph but not in the block.
3069 // We put edges from the prior and current DEF/KILLs to the pinch point.
3070 // We put the pinch point in _reg_node. If there's already a pinch point
3071 // we merely add an edge from the current DEF/KILL to the pinch point.
3072
3073 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3074 // put an edge from the pinch point to the USE.
3075
3076 // To be expedient, the _reg_node array is pre-allocated for the whole
3077 // compilation. _reg_node is lazily initialized; it either contains a null,
3078 // or a valid def/kill/pinch-point, or a leftover node from some prior
3079 // block. Leftover node from some prior block is treated like a null (no
3080 // prior def, so no anti-dependence needed). Valid def is distinguished by
3081 // it being in the current block.
3082 bool fat_proj_seen = false;
3083 uint last_safept = _bb_end-1;
3084 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3085 Node* last_safept_node = end_node;
3086 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3087 Node *n = b->get_node(i);
3088 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3089 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3090 // Fat-proj kills a slew of registers
3091 // This can add edges to 'n' and obscure whether or not it was a def,
3092 // hence the is_def flag.
3093 fat_proj_seen = true;
3094 RegMaskIterator rmi(n->out_RegMask());
3095 while (rmi.has_next()) {
3096 OptoReg::Name kill = rmi.next();
3097 anti_do_def(b, n, kill, is_def);
3098 }
3099 } else {
3100 // Get DEF'd registers the normal way
3101 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3102 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3103 }
3104
3105 // Kill projections on a branch should appear to occur on the
3106 // branch, not afterwards, so grab the masks from the projections
3107 // and process them.
3108 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3109 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3110 Node* use = n->fast_out(i);
3111 if (use->is_Proj()) {
3112 RegMaskIterator rmi(use->out_RegMask());
3113 while (rmi.has_next()) {
3114 OptoReg::Name kill = rmi.next();
3115 anti_do_def(b, n, kill, false);
3116 }
3117 }
3118 }
3119 }
3120
3121 // Check each register used by this instruction for a following DEF/KILL
3122 // that must occur afterward and requires an anti-dependence edge.
3123 for( uint j=0; j<n->req(); j++ ) {
3124 Node *def = n->in(j);
3125 if( def ) {
3126 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3127 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3128 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3129 }
3130 }
3131 // Do not allow defs of new derived values to float above GC
3132 // points unless the base is definitely available at the GC point.
3133
3134 Node *m = b->get_node(i);
3135
3136 // Add precedence edge from following safepoint to use of derived pointer
3137 if( last_safept_node != end_node &&
3138 m != last_safept_node) {
3139 for (uint k = 1; k < m->req(); k++) {
3140 const Type *t = m->in(k)->bottom_type();
3141 if( t->isa_oop_ptr() &&
3142 t->is_ptr()->offset() != 0 ) {
3143 last_safept_node->add_prec( m );
3144 break;
3145 }
3146 }
3147 }
3148
3149 if( n->jvms() ) { // Precedence edge from derived to safept
3150 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3151 if( b->get_node(last_safept) != last_safept_node ) {
3152 last_safept = b->find_node(last_safept_node);
3153 }
3154 for( uint j=last_safept; j > i; j-- ) {
3155 Node *mach = b->get_node(j);
3156 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3157 mach->add_prec( n );
3158 }
3159 last_safept = i;
3160 last_safept_node = m;
3161 }
3162 }
3163
3164 if (fat_proj_seen) {
3165 // Garbage collect pinch nodes that were not consumed.
3166 // They are usually created by a fat kill MachProj for a call.
3167 garbage_collect_pinch_nodes();
3168 }
3169 }
3170
3171 // Garbage collect pinch nodes for reuse by other blocks.
3172 //
3173 // The block scheduler's insertion of anti-dependence
3174 // edges creates many pinch nodes when the block contains
3175 // 2 or more Calls. A pinch node is used to prevent a
3176 // combinatorial explosion of edges. If a set of kills for a
3177 // register is anti-dependent on a set of uses (or defs), rather
3178 // than adding an edge in the graph between each pair of kill
3179 // and use (or def), a pinch is inserted between them:
3180 //
3181 // use1 use2 use3
3182 // \ | /
3183 // \ | /
3184 // pinch
3185 // / | \
3186 // / | \
3187 // kill1 kill2 kill3
3188 //
3189 // One pinch node is created per register killed when
3190 // the second call is encountered during a backwards pass
3191 // over the block. Most of these pinch nodes are never
3192 // wired into the graph because the register is never
3193 // used or def'ed in the block.
3194 //
3195 void Scheduling::garbage_collect_pinch_nodes() {
3196 #ifndef PRODUCT
3197 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3198 #endif
3199 int trace_cnt = 0;
3200 for (uint k = 0; k < _reg_node.max(); k++) {
3201 Node* pinch = _reg_node[k];
3202 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3203 // no predecence input edges
3204 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3205 cleanup_pinch(pinch);
3206 _pinch_free_list.push(pinch);
3207 _reg_node.map(k, nullptr);
3208 #ifndef PRODUCT
3209 if (_cfg->C->trace_opto_output()) {
3210 trace_cnt++;
3211 if (trace_cnt > 40) {
3212 tty->print("\n");
3213 trace_cnt = 0;
3214 }
3215 tty->print(" %d", pinch->_idx);
3216 }
3217 #endif
3218 }
3219 }
3220 #ifndef PRODUCT
3221 if (_cfg->C->trace_opto_output()) tty->print("\n");
3222 #endif
3223 }
3224
3225 // Clean up a pinch node for reuse.
3226 void Scheduling::cleanup_pinch( Node *pinch ) {
3227 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3228
3229 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3230 Node* use = pinch->last_out(i);
3231 uint uses_found = 0;
3232 for (uint j = use->req(); j < use->len(); j++) {
3233 if (use->in(j) == pinch) {
3234 use->rm_prec(j);
3235 uses_found++;
3236 }
3237 }
3238 assert(uses_found > 0, "must be a precedence edge");
3239 i -= uses_found; // we deleted 1 or more copies of this edge
3240 }
3241 // May have a later_def entry
3242 pinch->set_req(0, nullptr);
3243 }
3244
3245 #ifndef PRODUCT
3246
3247 void Scheduling::dump_available() const {
3248 tty->print("#Availist ");
3249 for (uint i = 0; i < _available.size(); i++)
3250 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3251 tty->cr();
3252 }
3253
3254 // Print Scheduling Statistics
3255 void Scheduling::print_statistics() {
3256 // Print the size added by nops for bundling
3257 tty->print("Nops added %d bytes to total of %d bytes",
3258 _total_nop_size, _total_method_size);
3259 if (_total_method_size > 0)
3260 tty->print(", for %.2f%%",
3261 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3262 tty->print("\n");
3263
3264 // Print the number of branch shadows filled
3265 if (Pipeline::_branch_has_delay_slot) {
3266 tty->print("Of %d branches, %d had unconditional delay slots filled",
3267 _total_branches, _total_unconditional_delays);
3268 if (_total_branches > 0)
3269 tty->print(", for %.2f%%",
3270 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3271 tty->print("\n");
3272 }
3273
3274 uint total_instructions = 0, total_bundles = 0;
3275
3276 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3277 uint bundle_count = _total_instructions_per_bundle[i];
3278 total_instructions += bundle_count * i;
3279 total_bundles += bundle_count;
3280 }
3281
3282 if (total_bundles > 0)
3283 tty->print("Average ILP (excluding nops) is %.2f\n",
3284 ((double)total_instructions) / ((double)total_bundles));
3285 }
3286 #endif
3287
3288 //-----------------------init_scratch_buffer_blob------------------------------
3289 // Construct a temporary BufferBlob and cache it for this compile.
3290 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3291 // If there is already a scratch buffer blob allocated and the
3292 // constant section is big enough, use it. Otherwise free the
3293 // current and allocate a new one.
3294 BufferBlob* blob = scratch_buffer_blob();
3295 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3296 // Use the current blob.
3297 } else {
3298 if (blob != nullptr) {
3299 BufferBlob::free(blob);
3300 }
3301
3302 ResourceMark rm;
3303 _scratch_const_size = const_size;
3304 int size = C2Compiler::initial_code_buffer_size(const_size);
3305 blob = BufferBlob::create("Compile::scratch_buffer", size);
3306 // Record the buffer blob for next time.
3307 set_scratch_buffer_blob(blob);
3308 // Have we run out of code space?
3309 if (scratch_buffer_blob() == nullptr) {
3310 // Let CompilerBroker disable further compilations.
3311 C->record_failure("Not enough space for scratch buffer in CodeCache");
3312 return;
3313 }
3314 }
3315
3316 // Initialize the relocation buffers
3317 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3318 set_scratch_locs_memory(locs_buf);
3319 }
3320
3321
3322 //-----------------------scratch_emit_size-------------------------------------
3323 // Helper function that computes size by emitting code
3324 uint PhaseOutput::scratch_emit_size(const Node* n) {
3325 // Start scratch_emit_size section.
3326 set_in_scratch_emit_size(true);
3327
3328 // Emit into a trash buffer and count bytes emitted.
3329 // This is a pretty expensive way to compute a size,
3330 // but it works well enough if seldom used.
3331 // All common fixed-size instructions are given a size
3332 // method by the AD file.
3333 // Note that the scratch buffer blob and locs memory are
3334 // allocated at the beginning of the compile task, and
3335 // may be shared by several calls to scratch_emit_size.
3336 // The allocation of the scratch buffer blob is particularly
3337 // expensive, since it has to grab the code cache lock.
3338 BufferBlob* blob = this->scratch_buffer_blob();
3339 assert(blob != nullptr, "Initialize BufferBlob at start");
3340 assert(blob->size() > MAX_inst_size, "sanity");
3341 relocInfo* locs_buf = scratch_locs_memory();
3342 address blob_begin = blob->content_begin();
3343 address blob_end = (address)locs_buf;
3344 assert(blob->contains(blob_end), "sanity");
3345 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3346 buf.initialize_consts_size(_scratch_const_size);
3347 buf.initialize_stubs_size(MAX_stubs_size);
3348 assert(locs_buf != nullptr, "sanity");
3349 int lsize = MAX_locs_size / 3;
3350 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3351 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3352 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3353 // Mark as scratch buffer.
3354 buf.consts()->set_scratch_emit();
3355 buf.insts()->set_scratch_emit();
3356 buf.stubs()->set_scratch_emit();
3357
3358 // Do the emission.
3359
3360 Label fakeL; // Fake label for branch instructions.
3361 Label* saveL = nullptr;
3362 uint save_bnum = 0;
3363 bool is_branch = n->is_MachBranch();
3364 C2_MacroAssembler masm(&buf);
3365 masm.bind(fakeL);
3366 if (is_branch) {
3367 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3368 n->as_MachBranch()->label_set(&fakeL, 0);
3369 }
3370 n->emit(&masm, C->regalloc());
3371
3372 // Emitting into the scratch buffer should not fail
3373 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3374
3375 if (is_branch) // Restore label.
3376 n->as_MachBranch()->label_set(saveL, save_bnum);
3377
3378 // End scratch_emit_size section.
3379 set_in_scratch_emit_size(false);
3380
3381 return buf.insts_size();
3382 }
3383
3384 void PhaseOutput::install() {
3385 if (!C->should_install_code()) {
3386 return;
3387 } else if (C->stub_function() != nullptr) {
3388 install_stub(C->stub_name());
3389 } else {
3390 install_code(C->method(),
3391 C->entry_bci(),
3392 CompileBroker::compiler2(),
3393 C->has_unsafe_access(),
3394 SharedRuntime::is_wide_vector(C->max_vector_size()));
3395 }
3396 }
3397
3398 void PhaseOutput::install_code(ciMethod* target,
3399 int entry_bci,
3400 AbstractCompiler* compiler,
3401 bool has_unsafe_access,
3402 bool has_wide_vectors) {
3403 // Check if we want to skip execution of all compiled code.
3404 {
3405 #ifndef PRODUCT
3406 if (OptoNoExecute) {
3407 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3408 return;
3409 }
3410 #endif
3411 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3412
3413 if (C->is_osr_compilation()) {
3414 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3415 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3416 } else {
3417 if (!target->is_static()) {
3418 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3419 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3420 // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3421 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3422 }
3423 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3424 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3425 }
3426
3427 C->env()->register_method(target,
3428 entry_bci,
3429 &_code_offsets,
3430 _orig_pc_slot_offset_in_bytes,
3431 code_buffer(),
3432 frame_size_in_words(),
3433 oop_map_set(),
3434 &_handler_table,
3435 inc_table(),
3436 compiler,
3437 has_unsafe_access,
3438 SharedRuntime::is_wide_vector(C->max_vector_size()),
3439 C->has_monitors(),
3440 C->has_scoped_access(),
3441 0);
3442
3443 if (C->log() != nullptr) { // Print code cache state into compiler log
3444 C->log()->code_cache_state();
3445 }
3446 }
3447 }
3448 void PhaseOutput::install_stub(const char* stub_name) {
3449 // Entry point will be accessed using stub_entry_point();
3450 if (code_buffer() == nullptr) {
3451 Matcher::soft_match_failure();
3452 } else {
3453 if (PrintAssembly && (WizardMode || Verbose))
3454 tty->print_cr("### Stub::%s", stub_name);
3455
3456 if (!C->failing()) {
3457 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3458
3459 // Make the NMethod
3460 // For now we mark the frame as never safe for profile stackwalking
3461 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3462 code_buffer(),
3463 CodeOffsets::frame_never_safe,
3464 // _code_offsets.value(CodeOffsets::Frame_Complete),
3465 frame_size_in_words(),
3466 oop_map_set(),
3467 false);
3468 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3469
3470 C->set_stub_entry_point(rs->entry_point());
3471 }
3472 }
3473 }
3474
3475 // Support for bundling info
3476 Bundle* PhaseOutput::node_bundling(const Node *n) {
3477 assert(valid_bundle_info(n), "oob");
3478 return &_node_bundling_base[n->_idx];
3479 }
3480
3481 bool PhaseOutput::valid_bundle_info(const Node *n) {
3482 return (_node_bundling_limit > n->_idx);
3483 }
3484
3485 //------------------------------frame_size_in_words-----------------------------
3486 // frame_slots in units of words
3487 int PhaseOutput::frame_size_in_words() const {
3488 // shift is 0 in LP32 and 1 in LP64
3489 const int shift = (LogBytesPerWord - LogBytesPerInt);
3490 int words = _frame_slots >> shift;
3491 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3492 return words;
3493 }
3494
3495 // To bang the stack of this compiled method we use the stack size
3496 // that the interpreter would need in case of a deoptimization. This
3497 // removes the need to bang the stack in the deoptimization blob which
3498 // in turn simplifies stack overflow handling.
3499 int PhaseOutput::bang_size_in_bytes() const {
3500 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3501 }
3502
3503 //------------------------------dump_asm---------------------------------------
3504 // Dump formatted assembly
3505 #if defined(SUPPORT_OPTO_ASSEMBLY)
3506 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3507
3508 int pc_digits = 3; // #chars required for pc
3509 int sb_chars = 3; // #chars for "start bundle" indicator
3510 int tab_size = 8;
3511 if (pcs != nullptr) {
3512 int max_pc = 0;
3513 for (uint i = 0; i < pc_limit; i++) {
3514 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3515 }
3516 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3517 }
3518 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3519
3520 bool cut_short = false;
3521 st->print_cr("#");
3522 st->print("# "); C->tf()->dump_on(st); st->cr();
3523 st->print_cr("#");
3524
3525 // For all blocks
3526 int pc = 0x0; // Program counter
3527 char starts_bundle = ' ';
3528 C->regalloc()->dump_frame();
3529
3530 Node *n = nullptr;
3531 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3532 if (VMThread::should_terminate()) {
3533 cut_short = true;
3534 break;
3535 }
3536 Block* block = C->cfg()->get_block(i);
3537 if (block->is_connector() && !Verbose) {
3538 continue;
3539 }
3540 n = block->head();
3541 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3542 pc = pcs[n->_idx];
3543 st->print("%*.*x", pc_digits, pc_digits, pc);
3544 }
3545 st->fill_to(prefix_len);
3546 block->dump_head(C->cfg(), st);
3547 if (block->is_connector()) {
3548 st->fill_to(prefix_len);
3549 st->print_cr("# Empty connector block");
3550 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3551 st->fill_to(prefix_len);
3552 st->print_cr("# Block is sole successor of call");
3553 }
3554
3555 // For all instructions
3556 Node *delay = nullptr;
3557 for (uint j = 0; j < block->number_of_nodes(); j++) {
3558 if (VMThread::should_terminate()) {
3559 cut_short = true;
3560 break;
3561 }
3562 n = block->get_node(j);
3563 if (valid_bundle_info(n)) {
3564 Bundle* bundle = node_bundling(n);
3565 if (bundle->used_in_unconditional_delay()) {
3566 delay = n;
3567 continue;
3568 }
3569 if (bundle->starts_bundle()) {
3570 starts_bundle = '+';
3571 }
3572 }
3573
3574 if (WizardMode) {
3575 n->dump();
3576 }
3577
3578 if( !n->is_Region() && // Dont print in the Assembly
3579 !n->is_Phi() && // a few noisely useless nodes
3580 !n->is_Proj() &&
3581 !n->is_MachTemp() &&
3582 !n->is_SafePointScalarObject() &&
3583 !n->is_Catch() && // Would be nice to print exception table targets
3584 !n->is_MergeMem() && // Not very interesting
3585 !n->is_top() && // Debug info table constants
3586 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3587 ) {
3588 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3589 pc = pcs[n->_idx];
3590 st->print("%*.*x", pc_digits, pc_digits, pc);
3591 } else {
3592 st->fill_to(pc_digits);
3593 }
3594 st->print(" %c ", starts_bundle);
3595 starts_bundle = ' ';
3596 st->fill_to(prefix_len);
3597 n->format(C->regalloc(), st);
3598 st->cr();
3599 }
3600
3601 // If we have an instruction with a delay slot, and have seen a delay,
3602 // then back up and print it
3603 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3604 // Coverity finding - Explicit null dereferenced.
3605 guarantee(delay != nullptr, "no unconditional delay instruction");
3606 if (WizardMode) delay->dump();
3607
3608 if (node_bundling(delay)->starts_bundle())
3609 starts_bundle = '+';
3610 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3611 pc = pcs[n->_idx];
3612 st->print("%*.*x", pc_digits, pc_digits, pc);
3613 } else {
3614 st->fill_to(pc_digits);
3615 }
3616 st->print(" %c ", starts_bundle);
3617 starts_bundle = ' ';
3618 st->fill_to(prefix_len);
3619 delay->format(C->regalloc(), st);
3620 st->cr();
3621 delay = nullptr;
3622 }
3623
3624 // Dump the exception table as well
3625 if( n->is_Catch() && (Verbose || WizardMode) ) {
3626 // Print the exception table for this offset
3627 _handler_table.print_subtable_for(pc);
3628 }
3629 st->bol(); // Make sure we start on a new line
3630 }
3631 st->cr(); // one empty line between blocks
3632 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3633 } // End of per-block dump
3634
3635 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3636 }
3637 #endif
3638
3639 #ifndef PRODUCT
3640 void PhaseOutput::print_statistics() {
3641 Scheduling::print_statistics();
3642 }
3643 #endif