1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/assembler.inline.hpp"
26 #include "code/compiledIC.hpp"
27 #include "code/debugInfo.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compilerDirectives.hpp"
31 #include "compiler/disassembler.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "memory/allocation.hpp"
36 #include "opto/ad.hpp"
37 #include "opto/block.hpp"
38 #include "opto/c2compiler.hpp"
39 #include "opto/c2_MacroAssembler.hpp"
40 #include "opto/callnode.hpp"
41 #include "opto/cfgnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/machnode.hpp"
44 #include "opto/node.hpp"
45 #include "opto/optoreg.hpp"
46 #include "opto/output.hpp"
47 #include "opto/regalloc.hpp"
48 #include "opto/type.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/xmlstream.hpp"
53
54 #ifndef PRODUCT
55 #define DEBUG_ARG(x) , x
56 #else
57 #define DEBUG_ARG(x)
58 #endif
59
60 //------------------------------Scheduling----------------------------------
61 // This class contains all the information necessary to implement instruction
62 // scheduling and bundling.
63 class Scheduling {
64
65 private:
66 // Arena to use
67 Arena *_arena;
68
69 // Control-Flow Graph info
70 PhaseCFG *_cfg;
71
72 // Register Allocation info
73 PhaseRegAlloc *_regalloc;
74
75 // Number of nodes in the method
76 uint _node_bundling_limit;
77
78 // List of scheduled nodes. Generated in reverse order
79 Node_List _scheduled;
80
81 // List of nodes currently available for choosing for scheduling
82 Node_List _available;
83
84 // For each instruction beginning a bundle, the number of following
85 // nodes to be bundled with it.
86 Bundle *_node_bundling_base;
87
88 // Mapping from register to Node
89 Node_List _reg_node;
90
91 // Free list for pinch nodes.
92 Node_List _pinch_free_list;
93
94 // Number of uses of this node within the containing basic block.
95 short *_uses;
96
97 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
98 // front, branch+proj at end. Also skips Catch/CProj (same as
99 // branch-at-end), plus just-prior exception-throwing call.
100 uint _bb_start, _bb_end;
101
102 // Latency from the end of the basic block as scheduled
103 unsigned short *_current_latency;
104
105 // Remember the next node
106 Node *_next_node;
107
108 // Use this for an unconditional branch delay slot
109 Node *_unconditional_delay_slot;
110
111 // Pointer to a Nop
112 MachNopNode *_nop;
113
114 // Length of the current bundle, in instructions
115 uint _bundle_instr_count;
116
117 // Current Cycle number, for computing latencies and bundling
118 uint _bundle_cycle_number;
119
120 // Bundle information
121 Pipeline_Use_Element _bundle_use_elements[resource_count];
122 Pipeline_Use _bundle_use;
123
124 // Dump the available list
125 void dump_available() const;
126
127 public:
128 Scheduling(Arena *arena, Compile &compile);
129
130 // Destructor
131 NOT_PRODUCT( ~Scheduling(); )
132
133 // Step ahead "i" cycles
134 void step(uint i);
135
136 // Step ahead 1 cycle, and clear the bundle state (for example,
137 // at a branch target)
138 void step_and_clear();
139
140 Bundle* node_bundling(const Node *n) {
141 assert(valid_bundle_info(n), "oob");
142 return (&_node_bundling_base[n->_idx]);
143 }
144
145 bool valid_bundle_info(const Node *n) const {
146 return (_node_bundling_limit > n->_idx);
147 }
148
149 bool starts_bundle(const Node *n) const {
150 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
151 }
152
153 // Do the scheduling
154 void DoScheduling();
155
156 // Compute the register antidependencies within a basic block
157 void ComputeRegisterAntidependencies(Block *bb);
158 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
159 void verify_good_schedule( Block *b, const char *msg );
160 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
161 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
162
163 // Add a node to the current bundle
164 void AddNodeToBundle(Node *n, const Block *bb);
165
166 // Return an integer less than, equal to, or greater than zero
167 // if the stack offset of the first argument is respectively
168 // less than, equal to, or greater than the second.
169 int compare_two_spill_nodes(Node* first, Node* second);
170
171 // Add a node to the list of available nodes
172 void AddNodeToAvailableList(Node *n);
173
174 // Compute the local use count for the nodes in a block, and compute
175 // the list of instructions with no uses in the block as available
176 void ComputeUseCount(const Block *bb);
177
178 // Choose an instruction from the available list to add to the bundle
179 Node * ChooseNodeToBundle();
180
181 // See if this Node fits into the currently accumulating bundle
182 bool NodeFitsInBundle(Node *n);
183
184 // Decrement the use count for a node
185 void DecrementUseCounts(Node *n, const Block *bb);
186
187 // Garbage collect pinch nodes for reuse by other blocks.
188 void garbage_collect_pinch_nodes();
189 // Clean up a pinch node for reuse (helper for above).
190 void cleanup_pinch( Node *pinch );
191
192 // Information for statistics gathering
193 #ifndef PRODUCT
194 private:
195 // Gather information on size of nops relative to total
196 uint _branches, _unconditional_delays;
197
198 static uint _total_nop_size, _total_method_size;
199 static uint _total_branches, _total_unconditional_delays;
200 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
201
202 public:
203 static void print_statistics();
204
205 static void increment_instructions_per_bundle(uint i) {
206 _total_instructions_per_bundle[i]++;
207 }
208
209 static void increment_nop_size(uint s) {
210 _total_nop_size += s;
211 }
212
213 static void increment_method_size(uint s) {
214 _total_method_size += s;
215 }
216 #endif
217
218 };
219
220 PhaseOutput::PhaseOutput()
221 : Phase(Phase::Output),
222 _code_buffer("Compile::Fill_buffer"),
223 _first_block_size(0),
224 _handler_table(),
225 _inc_table(),
226 _stub_list(),
227 _oop_map_set(nullptr),
228 _scratch_buffer_blob(nullptr),
229 _scratch_locs_memory(nullptr),
230 _scratch_const_size(-1),
231 _in_scratch_emit_size(false),
232 _frame_slots(0),
233 _code_offsets(),
234 _node_bundling_limit(0),
235 _node_bundling_base(nullptr),
236 _orig_pc_slot(0),
237 _orig_pc_slot_offset_in_bytes(0),
238 _buf_sizes(),
239 _block(nullptr),
240 _index(0) {
241 C->set_output(this);
242 if (C->stub_name() == nullptr) {
243 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
244 }
245 }
246
247 PhaseOutput::~PhaseOutput() {
248 C->set_output(nullptr);
249 if (_scratch_buffer_blob != nullptr) {
250 BufferBlob::free(_scratch_buffer_blob);
251 }
252 }
253
254 void PhaseOutput::perform_mach_node_analysis() {
255 // Late barrier analysis must be done after schedule and bundle
256 // Otherwise liveness based spilling will fail
257 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
258 bs->late_barrier_analysis();
259
260 pd_perform_mach_node_analysis();
261
262 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
263 }
264
265 // Convert Nodes to instruction bits and pass off to the VM
266 void PhaseOutput::Output() {
267 // RootNode goes
268 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
269
270 // The number of new nodes (mostly MachNop) is proportional to
271 // the number of java calls and inner loops which are aligned.
272 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
273 C->inner_loops()*(OptoLoopAlignment-1)),
274 "out of nodes before code generation" ) ) {
275 return;
276 }
277 // Make sure I can find the Start Node
278 Block *entry = C->cfg()->get_block(1);
279 Block *broot = C->cfg()->get_root_block();
280
281 const StartNode *start = entry->head()->as_Start();
282
283 // Replace StartNode with prolog
284 MachPrologNode *prolog = new MachPrologNode();
285 entry->map_node(prolog, 0);
286 C->cfg()->map_node_to_block(prolog, entry);
287 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
288
289 // Virtual methods need an unverified entry point
290
291 if( C->is_osr_compilation() ) {
292 if( PoisonOSREntry ) {
293 // TODO: Should use a ShouldNotReachHereNode...
294 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
295 }
296 } else {
297 if( C->method() && !C->method()->flags().is_static() ) {
298 // Insert unvalidated entry point
299 C->cfg()->insert( broot, 0, new MachUEPNode() );
300 }
301
302 }
303
304 // Break before main entry point
305 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
306 (OptoBreakpoint && C->is_method_compilation()) ||
307 (OptoBreakpointOSR && C->is_osr_compilation()) ||
308 (OptoBreakpointC2R && !C->method()) ) {
309 // checking for C->method() means that OptoBreakpoint does not apply to
310 // runtime stubs or frame converters
311 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
312 }
313
314 // Insert epilogs before every return
315 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
316 Block* block = C->cfg()->get_block(i);
317 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
318 Node* m = block->end();
319 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
320 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
321 block->add_inst(epilog);
322 C->cfg()->map_node_to_block(epilog, block);
323 }
324 }
325 }
326
327 // Keeper of sizing aspects
328 _buf_sizes = BufferSizingData();
329
330 // Initialize code buffer
331 estimate_buffer_size(_buf_sizes._const);
332 if (C->failing()) return;
333
334 // Pre-compute the length of blocks and replace
335 // long branches with short if machine supports it.
336 // Must be done before ScheduleAndBundle due to SPARC delay slots
337 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
338 blk_starts[0] = 0;
339 shorten_branches(blk_starts);
340
341 ScheduleAndBundle();
342 if (C->failing()) {
343 return;
344 }
345
346 perform_mach_node_analysis();
347
348 // Complete sizing of codebuffer
349 CodeBuffer* cb = init_buffer();
350 if (cb == nullptr || C->failing()) {
351 return;
352 }
353
354 BuildOopMaps();
355
356 if (C->failing()) {
357 return;
358 }
359
360 C2_MacroAssembler masm(cb);
361 fill_buffer(&masm, blk_starts);
362 }
363
364 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
365 // Determine if we need to generate a stack overflow check.
366 // Do it if the method is not a stub function and
367 // has java calls or has frame size > vm_page_size/8.
368 // The debug VM checks that deoptimization doesn't trigger an
369 // unexpected stack overflow (compiled method stack banging should
370 // guarantee it doesn't happen) so we always need the stack bang in
371 // a debug VM.
372 return (C->stub_function() == nullptr &&
373 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
374 DEBUG_ONLY(|| true)));
375 }
376
377 bool PhaseOutput::need_register_stack_bang() const {
378 // Determine if we need to generate a register stack overflow check.
379 // This is only used on architectures which have split register
380 // and memory stacks.
381 // Bang if the method is not a stub function and has java calls
382 return (C->stub_function() == nullptr && C->has_java_calls());
383 }
384
385
386 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
387 // of a loop. When aligning a loop we need to provide enough instructions
388 // in cpu's fetch buffer to feed decoders. The loop alignment could be
389 // avoided if we have enough instructions in fetch buffer at the head of a loop.
390 // By default, the size is set to 999999 by Block's constructor so that
391 // a loop will be aligned if the size is not reset here.
392 //
393 // Note: Mach instructions could contain several HW instructions
394 // so the size is estimated only.
395 //
396 void PhaseOutput::compute_loop_first_inst_sizes() {
397 // The next condition is used to gate the loop alignment optimization.
398 // Don't aligned a loop if there are enough instructions at the head of a loop
399 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
400 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
401 // equal to 11 bytes which is the largest address NOP instruction.
402 if (MaxLoopPad < OptoLoopAlignment - 1) {
403 uint last_block = C->cfg()->number_of_blocks() - 1;
404 for (uint i = 1; i <= last_block; i++) {
405 Block* block = C->cfg()->get_block(i);
406 // Check the first loop's block which requires an alignment.
407 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
408 uint sum_size = 0;
409 uint inst_cnt = NumberOfLoopInstrToAlign;
410 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
411
412 // Check subsequent fallthrough blocks if the loop's first
413 // block(s) does not have enough instructions.
414 Block *nb = block;
415 while(inst_cnt > 0 &&
416 i < last_block &&
417 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
418 !nb->has_successor(block)) {
419 i++;
420 nb = C->cfg()->get_block(i);
421 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
422 } // while( inst_cnt > 0 && i < last_block )
423
424 block->set_first_inst_size(sum_size);
425 } // f( b->head()->is_Loop() )
426 } // for( i <= last_block )
427 } // if( MaxLoopPad < OptoLoopAlignment-1 )
428 }
429
430 // The architecture description provides short branch variants for some long
431 // branch instructions. Replace eligible long branches with short branches.
432 void PhaseOutput::shorten_branches(uint* blk_starts) {
433
434 Compile::TracePhase tp(_t_shortenBranches);
435
436 // Compute size of each block, method size, and relocation information size
437 uint nblocks = C->cfg()->number_of_blocks();
438
439 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
440 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
441 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
442
443 // Collect worst case block paddings
444 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
445 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
446
447 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
448 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
449
450 bool has_short_branch_candidate = false;
451
452 // Initialize the sizes to 0
453 int code_size = 0; // Size in bytes of generated code
454 int stub_size = 0; // Size in bytes of all stub entries
455 // Size in bytes of all relocation entries, including those in local stubs.
456 // Start with 2-bytes of reloc info for the unvalidated entry point
457 int reloc_size = 1; // Number of relocation entries
458
459 // Make three passes. The first computes pessimistic blk_starts,
460 // relative jmp_offset and reloc_size information. The second performs
461 // short branch substitution using the pessimistic sizing. The
462 // third inserts nops where needed.
463
464 // Step one, perform a pessimistic sizing pass.
465 uint last_call_adr = max_juint;
466 uint last_avoid_back_to_back_adr = max_juint;
467 uint nop_size = (new MachNopNode())->size(C->regalloc());
468 for (uint i = 0; i < nblocks; i++) { // For all blocks
469 Block* block = C->cfg()->get_block(i);
470 _block = block;
471
472 // During short branch replacement, we store the relative (to blk_starts)
473 // offset of jump in jmp_offset, rather than the absolute offset of jump.
474 // This is so that we do not need to recompute sizes of all nodes when
475 // we compute correct blk_starts in our next sizing pass.
476 jmp_offset[i] = 0;
477 jmp_size[i] = 0;
478 jmp_nidx[i] = -1;
479 DEBUG_ONLY( jmp_target[i] = 0; )
480 DEBUG_ONLY( jmp_rule[i] = 0; )
481
482 // Sum all instruction sizes to compute block size
483 uint last_inst = block->number_of_nodes();
484 uint blk_size = 0;
485 for (uint j = 0; j < last_inst; j++) {
486 _index = j;
487 Node* nj = block->get_node(_index);
488 // Handle machine instruction nodes
489 if (nj->is_Mach()) {
490 MachNode* mach = nj->as_Mach();
491 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
492 reloc_size += mach->reloc();
493 if (mach->is_MachCall()) {
494 // add size information for trampoline stub
495 // class CallStubImpl is platform-specific and defined in the *.ad files.
496 stub_size += CallStubImpl::size_call_trampoline();
497 reloc_size += CallStubImpl::reloc_call_trampoline();
498
499 MachCallNode *mcall = mach->as_MachCall();
500 // This destination address is NOT PC-relative
501
502 mcall->method_set((intptr_t)mcall->entry_point());
503
504 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
505 stub_size += CompiledDirectCall::to_interp_stub_size();
506 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
507 }
508 } else if (mach->is_MachSafePoint()) {
509 // If call/safepoint are adjacent, account for possible
510 // nop to disambiguate the two safepoints.
511 // ScheduleAndBundle() can rearrange nodes in a block,
512 // check for all offsets inside this block.
513 if (last_call_adr >= blk_starts[i]) {
514 blk_size += nop_size;
515 }
516 }
517 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
518 // Nop is inserted between "avoid back to back" instructions.
519 // ScheduleAndBundle() can rearrange nodes in a block,
520 // check for all offsets inside this block.
521 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
522 blk_size += nop_size;
523 }
524 }
525 if (mach->may_be_short_branch()) {
526 if (!nj->is_MachBranch()) {
527 #ifndef PRODUCT
528 nj->dump(3);
529 #endif
530 Unimplemented();
531 }
532 assert(jmp_nidx[i] == -1, "block should have only one branch");
533 jmp_offset[i] = blk_size;
534 jmp_size[i] = nj->size(C->regalloc());
535 jmp_nidx[i] = j;
536 has_short_branch_candidate = true;
537 }
538 }
539 blk_size += nj->size(C->regalloc());
540 // Remember end of call offset
541 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
542 last_call_adr = blk_starts[i]+blk_size;
543 }
544 // Remember end of avoid_back_to_back offset
545 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
546 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
547 }
548 }
549
550 // When the next block starts a loop, we may insert pad NOP
551 // instructions. Since we cannot know our future alignment,
552 // assume the worst.
553 if (i < nblocks - 1) {
554 Block* nb = C->cfg()->get_block(i + 1);
555 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
556 if (max_loop_pad > 0) {
557 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
558 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
559 // If either is the last instruction in this block, bump by
560 // max_loop_pad in lock-step with blk_size, so sizing
561 // calculations in subsequent blocks still can conservatively
562 // detect that it may the last instruction in this block.
563 if (last_call_adr == blk_starts[i]+blk_size) {
564 last_call_adr += max_loop_pad;
565 }
566 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
567 last_avoid_back_to_back_adr += max_loop_pad;
568 }
569 blk_size += max_loop_pad;
570 block_worst_case_pad[i + 1] = max_loop_pad;
571 }
572 }
573
574 // Save block size; update total method size
575 blk_starts[i+1] = blk_starts[i]+blk_size;
576 }
577
578 // Step two, replace eligible long jumps.
579 bool progress = true;
580 uint last_may_be_short_branch_adr = max_juint;
581 while (has_short_branch_candidate && progress) {
582 progress = false;
583 has_short_branch_candidate = false;
584 int adjust_block_start = 0;
585 for (uint i = 0; i < nblocks; i++) {
586 Block* block = C->cfg()->get_block(i);
587 int idx = jmp_nidx[i];
588 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
589 if (mach != nullptr && mach->may_be_short_branch()) {
590 #ifdef ASSERT
591 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
592 int j;
593 // Find the branch; ignore trailing NOPs.
594 for (j = block->number_of_nodes()-1; j>=0; j--) {
595 Node* n = block->get_node(j);
596 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
597 break;
598 }
599 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
600 #endif
601 int br_size = jmp_size[i];
602 int br_offs = blk_starts[i] + jmp_offset[i];
603
604 // This requires the TRUE branch target be in succs[0]
605 uint bnum = block->non_connector_successor(0)->_pre_order;
606 int offset = blk_starts[bnum] - br_offs;
607 if (bnum > i) { // adjust following block's offset
608 offset -= adjust_block_start;
609 }
610
611 // This block can be a loop header, account for the padding
612 // in the previous block.
613 int block_padding = block_worst_case_pad[i];
614 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
615 // In the following code a nop could be inserted before
616 // the branch which will increase the backward distance.
617 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
618 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
619
620 if (needs_padding && offset <= 0)
621 offset -= nop_size;
622
623 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
624 // We've got a winner. Replace this branch.
625 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
626
627 // Update the jmp_size.
628 int new_size = replacement->size(C->regalloc());
629 int diff = br_size - new_size;
630 assert(diff >= (int)nop_size, "short_branch size should be smaller");
631 // Conservatively take into account padding between
632 // avoid_back_to_back branches. Previous branch could be
633 // converted into avoid_back_to_back branch during next
634 // rounds.
635 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
636 jmp_offset[i] += nop_size;
637 diff -= nop_size;
638 }
639 adjust_block_start += diff;
640 block->map_node(replacement, idx);
641 mach->subsume_by(replacement, C);
642 mach = replacement;
643 progress = true;
644
645 jmp_size[i] = new_size;
646 DEBUG_ONLY( jmp_target[i] = bnum; );
647 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
648 } else {
649 // The jump distance is not short, try again during next iteration.
650 has_short_branch_candidate = true;
651 }
652 } // (mach->may_be_short_branch())
653 if (mach != nullptr && (mach->may_be_short_branch() ||
654 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
655 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
656 }
657 blk_starts[i+1] -= adjust_block_start;
658 }
659 }
660
661 #ifdef ASSERT
662 for (uint i = 0; i < nblocks; i++) { // For all blocks
663 if (jmp_target[i] != 0) {
664 int br_size = jmp_size[i];
665 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
666 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
667 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]);
668 }
669 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
670 }
671 }
672 #endif
673
674 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
675 // after ScheduleAndBundle().
676
677 // ------------------
678 // Compute size for code buffer
679 code_size = blk_starts[nblocks];
680
681 // Relocation records
682 reloc_size += 1; // Relo entry for exception handler
683
684 // Adjust reloc_size to number of record of relocation info
685 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
686 // a relocation index.
687 // The CodeBuffer will expand the locs array if this estimate is too low.
688 reloc_size *= 10 / sizeof(relocInfo);
689
690 _buf_sizes._reloc = reloc_size;
691 _buf_sizes._code = code_size;
692 _buf_sizes._stub = stub_size;
693 }
694
695 //------------------------------FillLocArray-----------------------------------
696 // Create a bit of debug info and append it to the array. The mapping is from
697 // Java local or expression stack to constant, register or stack-slot. For
698 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
699 // entry has been taken care of and caller should skip it).
700 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
701 // This should never have accepted Bad before
702 assert(OptoReg::is_valid(regnum), "location must be valid");
703 return (OptoReg::is_reg(regnum))
704 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
705 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
706 }
707
708
709 ObjectValue*
710 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
711 for (int i = 0; i < objs->length(); i++) {
712 assert(objs->at(i)->is_object(), "corrupt object cache");
713 ObjectValue* sv = objs->at(i)->as_ObjectValue();
714 if (sv->id() == id) {
715 return sv;
716 }
717 }
718 // Otherwise..
719 return nullptr;
720 }
721
722 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
723 ObjectValue* sv ) {
724 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
725 objs->append(sv);
726 }
727
728
729 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
730 GrowableArray<ScopeValue*> *array,
731 GrowableArray<ScopeValue*> *objs ) {
732 assert( local, "use _top instead of null" );
733 if (array->length() != idx) {
734 assert(array->length() == idx + 1, "Unexpected array count");
735 // Old functionality:
736 // return
737 // New functionality:
738 // Assert if the local is not top. In product mode let the new node
739 // override the old entry.
740 assert(local == C->top(), "LocArray collision");
741 if (local == C->top()) {
742 return;
743 }
744 array->pop();
745 }
746 const Type *t = local->bottom_type();
747
748 // Is it a safepoint scalar object node?
749 if (local->is_SafePointScalarObject()) {
750 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
751
752 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
753 if (sv == nullptr) {
754 ciKlass* cik = t->is_oopptr()->exact_klass();
755 assert(cik->is_instance_klass() ||
756 cik->is_array_klass(), "Not supported allocation.");
757 sv = new ObjectValue(spobj->_idx,
758 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
759 set_sv_for_object_node(objs, sv);
760
761 uint first_ind = spobj->first_index(sfpt->jvms());
762 for (uint i = 0; i < spobj->n_fields(); i++) {
763 Node* fld_node = sfpt->in(first_ind+i);
764 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
765 }
766 }
767 array->append(sv);
768 return;
769 } else if (local->is_SafePointScalarMerge()) {
770 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
771 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
772
773 if (mv == nullptr) {
774 GrowableArray<ScopeValue*> deps;
775
776 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
777 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
778 assert(deps.length() == 1, "missing value");
779
780 int selector_idx = smerge->selector_idx(sfpt->jvms());
781 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
782 assert(deps.length() == 2, "missing value");
783
784 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
785 set_sv_for_object_node(objs, mv);
786
787 for (uint i = 1; i < smerge->req(); i++) {
788 Node* obj_node = smerge->in(i);
789 int idx = mv->possible_objects()->length();
790 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
791
792 // By default ObjectValues that are in 'possible_objects' are not root objects.
793 // They will be marked as root later if they are directly referenced in a JVMS.
794 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
795 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
796 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
797 }
798 }
799 array->append(mv);
800 return;
801 }
802
803 // Grab the register number for the local
804 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
805 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
806 // Record the double as two float registers.
807 // The register mask for such a value always specifies two adjacent
808 // float registers, with the lower register number even.
809 // Normally, the allocation of high and low words to these registers
810 // is irrelevant, because nearly all operations on register pairs
811 // (e.g., StoreD) treat them as a single unit.
812 // Here, we assume in addition that the words in these two registers
813 // stored "naturally" (by operations like StoreD and double stores
814 // within the interpreter) such that the lower-numbered register
815 // is written to the lower memory address. This may seem like
816 // a machine dependency, but it is not--it is a requirement on
817 // the author of the <arch>.ad file to ensure that, for every
818 // even/odd double-register pair to which a double may be allocated,
819 // the word in the even single-register is stored to the first
820 // memory word. (Note that register numbers are completely
821 // arbitrary, and are not tied to any machine-level encodings.)
822 #ifdef _LP64
823 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
824 array->append(new ConstantIntValue((jint)0));
825 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
826 } else if ( t->base() == Type::Long ) {
827 array->append(new ConstantIntValue((jint)0));
828 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
829 } else if ( t->base() == Type::RawPtr ) {
830 // jsr/ret return address which must be restored into the full
831 // width 64-bit stack slot.
832 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
833 }
834 #else //_LP64
835 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
836 // Repack the double/long as two jints.
837 // The convention the interpreter uses is that the second local
838 // holds the first raw word of the native double representation.
839 // This is actually reasonable, since locals and stack arrays
840 // grow downwards in all implementations.
841 // (If, on some machine, the interpreter's Java locals or stack
842 // were to grow upwards, the embedded doubles would be word-swapped.)
843 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
844 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
845 }
846 #endif //_LP64
847 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
848 OptoReg::is_reg(regnum) ) {
849 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
850 ? Location::float_in_dbl : Location::normal ));
851 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
852 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
853 ? Location::int_in_long : Location::normal ));
854 } else if( t->base() == Type::NarrowOop ) {
855 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
856 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
857 t->base() == Type::VectorD || t->base() == Type::VectorX ||
858 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
859 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
860 } else if (C->regalloc()->is_oop(local)) {
861 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
862 t->base() == Type::AryPtr,
863 "Unexpected type: %s", t->msg());
864 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
865 } else {
866 assert(t->base() == Type::Int || t->base() == Type::Half ||
867 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
868 "Unexpected type: %s", t->msg());
869 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
870 }
871 return;
872 }
873
874 // No register. It must be constant data.
875 switch (t->base()) {
876 case Type::Half: // Second half of a double
877 ShouldNotReachHere(); // Caller should skip 2nd halves
878 break;
879 case Type::AnyPtr:
880 array->append(new ConstantOopWriteValue(nullptr));
881 break;
882 case Type::AryPtr:
883 case Type::InstPtr: // fall through
884 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
885 break;
886 case Type::NarrowOop:
887 if (t == TypeNarrowOop::NULL_PTR) {
888 array->append(new ConstantOopWriteValue(nullptr));
889 } else {
890 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
891 }
892 break;
893 case Type::Int:
894 array->append(new ConstantIntValue(t->is_int()->get_con()));
895 break;
896 case Type::RawPtr:
897 // A return address (T_ADDRESS).
898 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
899 #ifdef _LP64
900 // Must be restored to the full-width 64-bit stack slot.
901 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
902 #else
903 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
904 #endif
905 break;
906 case Type::FloatCon: {
907 float f = t->is_float_constant()->getf();
908 array->append(new ConstantIntValue(jint_cast(f)));
909 break;
910 }
911 case Type::DoubleCon: {
912 jdouble d = t->is_double_constant()->getd();
913 #ifdef _LP64
914 array->append(new ConstantIntValue((jint)0));
915 array->append(new ConstantDoubleValue(d));
916 #else
917 // Repack the double as two jints.
918 // The convention the interpreter uses is that the second local
919 // holds the first raw word of the native double representation.
920 // This is actually reasonable, since locals and stack arrays
921 // grow downwards in all implementations.
922 // (If, on some machine, the interpreter's Java locals or stack
923 // were to grow upwards, the embedded doubles would be word-swapped.)
924 jlong_accessor acc;
925 acc.long_value = jlong_cast(d);
926 array->append(new ConstantIntValue(acc.words[1]));
927 array->append(new ConstantIntValue(acc.words[0]));
928 #endif
929 break;
930 }
931 case Type::Long: {
932 jlong d = t->is_long()->get_con();
933 #ifdef _LP64
934 array->append(new ConstantIntValue((jint)0));
935 array->append(new ConstantLongValue(d));
936 #else
937 // Repack the long as two jints.
938 // The convention the interpreter uses is that the second local
939 // holds the first raw word of the native double representation.
940 // This is actually reasonable, since locals and stack arrays
941 // grow downwards in all implementations.
942 // (If, on some machine, the interpreter's Java locals or stack
943 // were to grow upwards, the embedded doubles would be word-swapped.)
944 jlong_accessor acc;
945 acc.long_value = d;
946 array->append(new ConstantIntValue(acc.words[1]));
947 array->append(new ConstantIntValue(acc.words[0]));
948 #endif
949 break;
950 }
951 case Type::Top: // Add an illegal value here
952 array->append(new LocationValue(Location()));
953 break;
954 default:
955 ShouldNotReachHere();
956 break;
957 }
958 }
959
960 // Determine if this node starts a bundle
961 bool PhaseOutput::starts_bundle(const Node *n) const {
962 return (_node_bundling_limit > n->_idx &&
963 _node_bundling_base[n->_idx].starts_bundle());
964 }
965
966 // Determine if there is a monitor that has 'ov' as its owner.
967 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
968 for (int k = 0; k < monarray->length(); k++) {
969 MonitorValue* mv = monarray->at(k);
970 if (mv->owner() == ov) {
971 return true;
972 }
973 }
974
975 return false;
976 }
977
978 // Determine if there is a scalar replaced object description represented by 'ov'.
979 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
980 GrowableArray<ScopeValue*>* objs,
981 ObjectValue* ov) const {
982 for (int i = 0; i < jvms->scl_size(); i++) {
983 Node* n = sfn->scalarized_obj(jvms, i);
984 // Other kinds of nodes that we may encounter here, for instance constants
985 // representing values of fields of objects scalarized, aren't relevant for
986 // us, since they don't map to ObjectValue.
987 if (!n->is_SafePointScalarObject()) {
988 continue;
989 }
990
991 ObjectValue* other = sv_for_node_id(objs, n->_idx);
992 if (ov == other) {
993 return true;
994 }
995 }
996 return false;
997 }
998
999 //--------------------------Process_OopMap_Node--------------------------------
1000 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1001 // Handle special safepoint nodes for synchronization
1002 MachSafePointNode *sfn = mach->as_MachSafePoint();
1003 MachCallNode *mcall;
1004
1005 int safepoint_pc_offset = current_offset;
1006 bool is_method_handle_invoke = false;
1007 bool return_oop = false;
1008 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1009 bool arg_escape = false;
1010
1011 // Add the safepoint in the DebugInfoRecorder
1012 if( !mach->is_MachCall() ) {
1013 mcall = nullptr;
1014 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1015 } else {
1016 mcall = mach->as_MachCall();
1017
1018 // Is the call a MethodHandle call?
1019 if (mcall->is_MachCallJava()) {
1020 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1021 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1022 is_method_handle_invoke = true;
1023 }
1024 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1025 }
1026
1027 // Check if a call returns an object.
1028 if (mcall->returns_pointer()) {
1029 return_oop = true;
1030 }
1031 safepoint_pc_offset += mcall->ret_addr_offset();
1032 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1033 }
1034
1035 // Loop over the JVMState list to add scope information
1036 // Do not skip safepoints with a null method, they need monitor info
1037 JVMState* youngest_jvms = sfn->jvms();
1038 int max_depth = youngest_jvms->depth();
1039
1040 // Allocate the object pool for scalar-replaced objects -- the map from
1041 // small-integer keys (which can be recorded in the local and ostack
1042 // arrays) to descriptions of the object state.
1043 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1044
1045 // Visit scopes from oldest to youngest.
1046 for (int depth = 1; depth <= max_depth; depth++) {
1047 JVMState* jvms = youngest_jvms->of_depth(depth);
1048 int idx;
1049 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1050 // Safepoints that do not have method() set only provide oop-map and monitor info
1051 // to support GC; these do not support deoptimization.
1052 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1053 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1054 int num_mon = jvms->nof_monitors();
1055 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1056 "JVMS local count must match that of the method");
1057
1058 // Add Local and Expression Stack Information
1059
1060 // Insert locals into the locarray
1061 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1062 for( idx = 0; idx < num_locs; idx++ ) {
1063 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1064 }
1065
1066 // Insert expression stack entries into the exparray
1067 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1068 for( idx = 0; idx < num_exps; idx++ ) {
1069 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1070 }
1071
1072 // Add in mappings of the monitors
1073 assert( !method ||
1074 !method->is_synchronized() ||
1075 method->is_native() ||
1076 num_mon > 0 ||
1077 !GenerateSynchronizationCode,
1078 "monitors must always exist for synchronized methods");
1079
1080 // Build the growable array of ScopeValues for exp stack
1081 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1082
1083 // Loop over monitors and insert into array
1084 for (idx = 0; idx < num_mon; idx++) {
1085 // Grab the node that defines this monitor
1086 Node* box_node = sfn->monitor_box(jvms, idx);
1087 Node* obj_node = sfn->monitor_obj(jvms, idx);
1088
1089 // Create ScopeValue for object
1090 ScopeValue *scval = nullptr;
1091
1092 if (obj_node->is_SafePointScalarObject()) {
1093 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1094 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1095 if (scval == nullptr) {
1096 const Type *t = spobj->bottom_type();
1097 ciKlass* cik = t->is_oopptr()->exact_klass();
1098 assert(cik->is_instance_klass() ||
1099 cik->is_array_klass(), "Not supported allocation.");
1100 ObjectValue* sv = new ObjectValue(spobj->_idx,
1101 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1102 PhaseOutput::set_sv_for_object_node(objs, sv);
1103
1104 uint first_ind = spobj->first_index(youngest_jvms);
1105 for (uint i = 0; i < spobj->n_fields(); i++) {
1106 Node* fld_node = sfn->in(first_ind+i);
1107 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1108 }
1109 scval = sv;
1110 }
1111 } else if (obj_node->is_SafePointScalarMerge()) {
1112 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1113 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1114
1115 if (mv == nullptr) {
1116 GrowableArray<ScopeValue*> deps;
1117
1118 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1119 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1120 assert(deps.length() == 1, "missing value");
1121
1122 int selector_idx = smerge->selector_idx(youngest_jvms);
1123 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1124 assert(deps.length() == 2, "missing value");
1125
1126 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1127 set_sv_for_object_node(objs, mv);
1128
1129 for (uint i = 1; i < smerge->req(); i++) {
1130 Node* obj_node = smerge->in(i);
1131 int idx = mv->possible_objects()->length();
1132 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1133
1134 // By default ObjectValues that are in 'possible_objects' are not root objects.
1135 // They will be marked as root later if they are directly referenced in a JVMS.
1136 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1137 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1138 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1139 }
1140 }
1141 scval = mv;
1142 } else if (!obj_node->is_Con()) {
1143 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1144 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1145 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1146 } else {
1147 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1148 }
1149 } else {
1150 const TypePtr *tp = obj_node->get_ptr_type();
1151 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1152 }
1153
1154 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1155 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1156 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1157 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1158 }
1159
1160 // Mark ObjectValue nodes as root nodes if they are directly
1161 // referenced in the JVMS.
1162 for (int i = 0; i < objs->length(); i++) {
1163 ScopeValue* sv = objs->at(i);
1164 if (sv->is_object_merge()) {
1165 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1166
1167 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1168 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1169 if (ov->is_root()) {
1170 // Already flagged as 'root' by something else. We shouldn't change it
1171 // to non-root in a younger JVMS because it may need to be alive in
1172 // a younger JVMS.
1173 } else {
1174 bool is_root = locarray->contains(ov) ||
1175 exparray->contains(ov) ||
1176 contains_as_owner(monarray, ov) ||
1177 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1178 ov->set_root(is_root);
1179 }
1180 }
1181 }
1182 }
1183
1184 // We dump the object pool first, since deoptimization reads it in first.
1185 C->debug_info()->dump_object_pool(objs);
1186
1187 // Build first class objects to pass to scope
1188 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1189 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1190 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1191
1192 // Make method available for all Safepoints
1193 ciMethod* scope_method = method ? method : C->method();
1194 // Describe the scope here
1195 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1196 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1197 // Now we can describe the scope.
1198 methodHandle null_mh;
1199 bool rethrow_exception = false;
1200 C->debug_info()->describe_scope(
1201 safepoint_pc_offset,
1202 null_mh,
1203 scope_method,
1204 jvms->bci(),
1205 jvms->should_reexecute(),
1206 rethrow_exception,
1207 is_method_handle_invoke,
1208 return_oop,
1209 has_ea_local_in_scope,
1210 arg_escape,
1211 locvals,
1212 expvals,
1213 monvals
1214 );
1215 } // End jvms loop
1216
1217 // Mark the end of the scope set.
1218 C->debug_info()->end_safepoint(safepoint_pc_offset);
1219 }
1220
1221
1222
1223 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1224 class NonSafepointEmitter {
1225 Compile* C;
1226 JVMState* _pending_jvms;
1227 int _pending_offset;
1228
1229 void emit_non_safepoint();
1230
1231 public:
1232 NonSafepointEmitter(Compile* compile) {
1233 this->C = compile;
1234 _pending_jvms = nullptr;
1235 _pending_offset = 0;
1236 }
1237
1238 void observe_instruction(Node* n, int pc_offset) {
1239 if (!C->debug_info()->recording_non_safepoints()) return;
1240
1241 Node_Notes* nn = C->node_notes_at(n->_idx);
1242 if (nn == nullptr || nn->jvms() == nullptr) return;
1243 if (_pending_jvms != nullptr &&
1244 _pending_jvms->same_calls_as(nn->jvms())) {
1245 // Repeated JVMS? Stretch it up here.
1246 _pending_offset = pc_offset;
1247 } else {
1248 if (_pending_jvms != nullptr &&
1249 _pending_offset < pc_offset) {
1250 emit_non_safepoint();
1251 }
1252 _pending_jvms = nullptr;
1253 if (pc_offset > C->debug_info()->last_pc_offset()) {
1254 // This is the only way _pending_jvms can become non-null:
1255 _pending_jvms = nn->jvms();
1256 _pending_offset = pc_offset;
1257 }
1258 }
1259 }
1260
1261 // Stay out of the way of real safepoints:
1262 void observe_safepoint(JVMState* jvms, int pc_offset) {
1263 if (_pending_jvms != nullptr &&
1264 !_pending_jvms->same_calls_as(jvms) &&
1265 _pending_offset < pc_offset) {
1266 emit_non_safepoint();
1267 }
1268 _pending_jvms = nullptr;
1269 }
1270
1271 void flush_at_end() {
1272 if (_pending_jvms != nullptr) {
1273 emit_non_safepoint();
1274 }
1275 _pending_jvms = nullptr;
1276 }
1277 };
1278
1279 void NonSafepointEmitter::emit_non_safepoint() {
1280 JVMState* youngest_jvms = _pending_jvms;
1281 int pc_offset = _pending_offset;
1282
1283 // Clear it now:
1284 _pending_jvms = nullptr;
1285
1286 DebugInformationRecorder* debug_info = C->debug_info();
1287 assert(debug_info->recording_non_safepoints(), "sanity");
1288
1289 debug_info->add_non_safepoint(pc_offset);
1290 int max_depth = youngest_jvms->depth();
1291
1292 // Visit scopes from oldest to youngest.
1293 for (int depth = 1; depth <= max_depth; depth++) {
1294 JVMState* jvms = youngest_jvms->of_depth(depth);
1295 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1296 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1297 methodHandle null_mh;
1298 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1299 }
1300
1301 // Mark the end of the scope set.
1302 debug_info->end_non_safepoint(pc_offset);
1303 }
1304
1305 //------------------------------init_buffer------------------------------------
1306 void PhaseOutput::estimate_buffer_size(int& const_req) {
1307
1308 // Set the initially allocated size
1309 const_req = initial_const_capacity;
1310
1311 // The extra spacing after the code is necessary on some platforms.
1312 // Sometimes we need to patch in a jump after the last instruction,
1313 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1314
1315 // Compute the byte offset where we can store the deopt pc.
1316 if (C->fixed_slots() != 0) {
1317 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1318 }
1319
1320 // Compute prolog code size
1321 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1322 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1323
1324 if (C->has_mach_constant_base_node()) {
1325 uint add_size = 0;
1326 // Fill the constant table.
1327 // Note: This must happen before shorten_branches.
1328 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1329 Block* b = C->cfg()->get_block(i);
1330
1331 for (uint j = 0; j < b->number_of_nodes(); j++) {
1332 Node* n = b->get_node(j);
1333
1334 // If the node is a MachConstantNode evaluate the constant
1335 // value section.
1336 if (n->is_MachConstant()) {
1337 MachConstantNode* machcon = n->as_MachConstant();
1338 machcon->eval_constant(C);
1339 } else if (n->is_Mach()) {
1340 // On Power there are more nodes that issue constants.
1341 add_size += (n->as_Mach()->ins_num_consts() * 8);
1342 }
1343 }
1344 }
1345
1346 // Calculate the offsets of the constants and the size of the
1347 // constant table (including the padding to the next section).
1348 constant_table().calculate_offsets_and_size();
1349 const_req = constant_table().alignment() + constant_table().size() + add_size;
1350 }
1351
1352 // Initialize the space for the BufferBlob used to find and verify
1353 // instruction size in MachNode::emit_size()
1354 init_scratch_buffer_blob(const_req);
1355 }
1356
1357 CodeBuffer* PhaseOutput::init_buffer() {
1358 int stub_req = _buf_sizes._stub;
1359 int code_req = _buf_sizes._code;
1360 int const_req = _buf_sizes._const;
1361
1362 int pad_req = NativeCall::byte_size();
1363
1364 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1365 stub_req += bs->estimate_stub_size();
1366
1367 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1368 // class HandlerImpl is platform-specific and defined in the *.ad files.
1369 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1370 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1371 stub_req += MAX_stubs_size; // ensure per-stub margin
1372 code_req += MAX_inst_size; // ensure per-instruction margin
1373
1374 if (StressCodeBuffers)
1375 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1376
1377 int total_req =
1378 const_req +
1379 code_req +
1380 pad_req +
1381 stub_req +
1382 exception_handler_req +
1383 deopt_handler_req; // deopt handler
1384
1385 if (C->has_method_handle_invokes())
1386 total_req += deopt_handler_req; // deopt MH handler
1387
1388 CodeBuffer* cb = code_buffer();
1389 cb->set_const_section_alignment(constant_table().alignment());
1390 cb->initialize(total_req, _buf_sizes._reloc);
1391
1392 // Have we run out of code space?
1393 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1394 C->record_failure("CodeCache is full");
1395 return nullptr;
1396 }
1397 // Configure the code buffer.
1398 cb->initialize_consts_size(const_req);
1399 cb->initialize_stubs_size(stub_req);
1400 cb->initialize_oop_recorder(C->env()->oop_recorder());
1401
1402 // fill in the nop array for bundling computations
1403 MachNode *_nop_list[Bundle::_nop_count];
1404 Bundle::initialize_nops(_nop_list);
1405
1406 return cb;
1407 }
1408
1409 //------------------------------fill_buffer------------------------------------
1410 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1411 // blk_starts[] contains offsets calculated during short branches processing,
1412 // offsets should not be increased during following steps.
1413
1414 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1415 // of a loop. It is used to determine the padding for loop alignment.
1416 Compile::TracePhase tp(_t_fillBuffer);
1417
1418 compute_loop_first_inst_sizes();
1419
1420 // Create oopmap set.
1421 _oop_map_set = new OopMapSet();
1422
1423 // !!!!! This preserves old handling of oopmaps for now
1424 C->debug_info()->set_oopmaps(_oop_map_set);
1425
1426 uint nblocks = C->cfg()->number_of_blocks();
1427 // Count and start of implicit null check instructions
1428 uint inct_cnt = 0;
1429 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1430
1431 // Count and start of calls
1432 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1433
1434 uint return_offset = 0;
1435 int nop_size = (new MachNopNode())->size(C->regalloc());
1436
1437 int previous_offset = 0;
1438 int current_offset = 0;
1439 int last_call_offset = -1;
1440 int last_avoid_back_to_back_offset = -1;
1441 #ifdef ASSERT
1442 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1443 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1444 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1445 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1446 #endif
1447
1448 // Create an array of unused labels, one for each basic block, if printing is enabled
1449 #if defined(SUPPORT_OPTO_ASSEMBLY)
1450 int* node_offsets = nullptr;
1451 uint node_offset_limit = C->unique();
1452
1453 if (C->print_assembly()) {
1454 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1455 }
1456 if (node_offsets != nullptr) {
1457 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1458 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1459 }
1460 #endif
1461
1462 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1463
1464 // Emit the constant table.
1465 if (C->has_mach_constant_base_node()) {
1466 if (!constant_table().emit(masm)) {
1467 C->record_failure("consts section overflow");
1468 return;
1469 }
1470 }
1471
1472 // Create an array of labels, one for each basic block
1473 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1474 for (uint i = 0; i <= nblocks; i++) {
1475 blk_labels[i].init();
1476 }
1477
1478 // Now fill in the code buffer
1479 Node* delay_slot = nullptr;
1480 for (uint i = 0; i < nblocks; i++) {
1481 Block* block = C->cfg()->get_block(i);
1482 _block = block;
1483 Node* head = block->head();
1484
1485 // If this block needs to start aligned (i.e, can be reached other
1486 // than by falling-thru from the previous block), then force the
1487 // start of a new bundle.
1488 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1489 masm->code()->flush_bundle(true);
1490 }
1491
1492 #ifdef ASSERT
1493 if (!block->is_connector()) {
1494 stringStream st;
1495 block->dump_head(C->cfg(), &st);
1496 masm->block_comment(st.freeze());
1497 }
1498 jmp_target[i] = 0;
1499 jmp_offset[i] = 0;
1500 jmp_size[i] = 0;
1501 jmp_rule[i] = 0;
1502 #endif
1503 int blk_offset = current_offset;
1504
1505 // Define the label at the beginning of the basic block
1506 masm->bind(blk_labels[block->_pre_order]);
1507
1508 uint last_inst = block->number_of_nodes();
1509
1510 // Emit block normally, except for last instruction.
1511 // Emit means "dump code bits into code buffer".
1512 for (uint j = 0; j<last_inst; j++) {
1513 _index = j;
1514
1515 // Get the node
1516 Node* n = block->get_node(j);
1517
1518 // See if delay slots are supported
1519 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1520 assert(delay_slot == nullptr, "no use of delay slot node");
1521 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1522
1523 delay_slot = n;
1524 continue;
1525 }
1526
1527 // If this starts a new instruction group, then flush the current one
1528 // (but allow split bundles)
1529 if (Pipeline::requires_bundling() && starts_bundle(n))
1530 masm->code()->flush_bundle(false);
1531
1532 // Special handling for SafePoint/Call Nodes
1533 bool is_mcall = false;
1534 if (n->is_Mach()) {
1535 MachNode *mach = n->as_Mach();
1536 is_mcall = n->is_MachCall();
1537 bool is_sfn = n->is_MachSafePoint();
1538
1539 // If this requires all previous instructions be flushed, then do so
1540 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1541 masm->code()->flush_bundle(true);
1542 current_offset = masm->offset();
1543 }
1544
1545 // A padding may be needed again since a previous instruction
1546 // could be moved to delay slot.
1547
1548 // align the instruction if necessary
1549 int padding = mach->compute_padding(current_offset);
1550 // Make sure safepoint node for polling is distinct from a call's
1551 // return by adding a nop if needed.
1552 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1553 padding = nop_size;
1554 }
1555 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1556 current_offset == last_avoid_back_to_back_offset) {
1557 // Avoid back to back some instructions.
1558 padding = nop_size;
1559 }
1560
1561 if (padding > 0) {
1562 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1563 int nops_cnt = padding / nop_size;
1564 MachNode *nop = new MachNopNode(nops_cnt);
1565 block->insert_node(nop, j++);
1566 last_inst++;
1567 C->cfg()->map_node_to_block(nop, block);
1568 // Ensure enough space.
1569 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1570 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1571 C->record_failure("CodeCache is full");
1572 return;
1573 }
1574 nop->emit(masm, C->regalloc());
1575 masm->code()->flush_bundle(true);
1576 current_offset = masm->offset();
1577 }
1578
1579 bool observe_safepoint = is_sfn;
1580 // Remember the start of the last call in a basic block
1581 if (is_mcall) {
1582 MachCallNode *mcall = mach->as_MachCall();
1583
1584 // This destination address is NOT PC-relative
1585 mcall->method_set((intptr_t)mcall->entry_point());
1586
1587 // Save the return address
1588 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1589
1590 observe_safepoint = mcall->guaranteed_safepoint();
1591 }
1592
1593 // sfn will be valid whenever mcall is valid now because of inheritance
1594 if (observe_safepoint) {
1595 // Handle special safepoint nodes for synchronization
1596 if (!is_mcall) {
1597 MachSafePointNode *sfn = mach->as_MachSafePoint();
1598 // !!!!! Stubs only need an oopmap right now, so bail out
1599 if (sfn->jvms()->method() == nullptr) {
1600 // Write the oopmap directly to the code blob??!!
1601 continue;
1602 }
1603 } // End synchronization
1604
1605 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1606 current_offset);
1607 Process_OopMap_Node(mach, current_offset);
1608 } // End if safepoint
1609
1610 // If this is a null check, then add the start of the previous instruction to the list
1611 else if( mach->is_MachNullCheck() ) {
1612 inct_starts[inct_cnt++] = previous_offset;
1613 }
1614
1615 // If this is a branch, then fill in the label with the target BB's label
1616 else if (mach->is_MachBranch()) {
1617 // This requires the TRUE branch target be in succs[0]
1618 uint block_num = block->non_connector_successor(0)->_pre_order;
1619
1620 // Try to replace long branch if delay slot is not used,
1621 // it is mostly for back branches since forward branch's
1622 // distance is not updated yet.
1623 bool delay_slot_is_used = valid_bundle_info(n) &&
1624 C->output()->node_bundling(n)->use_unconditional_delay();
1625 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1626 assert(delay_slot == nullptr, "not expecting delay slot node");
1627 int br_size = n->size(C->regalloc());
1628 int offset = blk_starts[block_num] - current_offset;
1629 if (block_num >= i) {
1630 // Current and following block's offset are not
1631 // finalized yet, adjust distance by the difference
1632 // between calculated and final offsets of current block.
1633 offset -= (blk_starts[i] - blk_offset);
1634 }
1635 // In the following code a nop could be inserted before
1636 // the branch which will increase the backward distance.
1637 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1638 if (needs_padding && offset <= 0)
1639 offset -= nop_size;
1640
1641 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1642 // We've got a winner. Replace this branch.
1643 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1644
1645 // Update the jmp_size.
1646 int new_size = replacement->size(C->regalloc());
1647 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1648 // Insert padding between avoid_back_to_back branches.
1649 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1650 MachNode *nop = new MachNopNode();
1651 block->insert_node(nop, j++);
1652 C->cfg()->map_node_to_block(nop, block);
1653 last_inst++;
1654 nop->emit(masm, C->regalloc());
1655 masm->code()->flush_bundle(true);
1656 current_offset = masm->offset();
1657 }
1658 #ifdef ASSERT
1659 jmp_target[i] = block_num;
1660 jmp_offset[i] = current_offset - blk_offset;
1661 jmp_size[i] = new_size;
1662 jmp_rule[i] = mach->rule();
1663 #endif
1664 block->map_node(replacement, j);
1665 mach->subsume_by(replacement, C);
1666 n = replacement;
1667 mach = replacement;
1668 }
1669 }
1670 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1671 } else if (mach->ideal_Opcode() == Op_Jump) {
1672 for (uint h = 0; h < block->_num_succs; h++) {
1673 Block* succs_block = block->_succs[h];
1674 for (uint j = 1; j < succs_block->num_preds(); j++) {
1675 Node* jpn = succs_block->pred(j);
1676 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1677 uint block_num = succs_block->non_connector()->_pre_order;
1678 Label *blkLabel = &blk_labels[block_num];
1679 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1680 }
1681 }
1682 }
1683 } else if (!n->is_Proj()) {
1684 // Remember the beginning of the previous instruction, in case
1685 // it's followed by a flag-kill and a null-check. Happens on
1686 // Intel all the time, with add-to-memory kind of opcodes.
1687 previous_offset = current_offset;
1688 }
1689
1690 // Not an else-if!
1691 // If this is a trap based cmp then add its offset to the list.
1692 if (mach->is_TrapBasedCheckNode()) {
1693 inct_starts[inct_cnt++] = current_offset;
1694 }
1695 }
1696
1697 // Verify that there is sufficient space remaining
1698 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1699 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1700 C->record_failure("CodeCache is full");
1701 return;
1702 }
1703
1704 // Save the offset for the listing
1705 #if defined(SUPPORT_OPTO_ASSEMBLY)
1706 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1707 node_offsets[n->_idx] = masm->offset();
1708 }
1709 #endif
1710 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1711
1712 // "Normal" instruction case
1713 DEBUG_ONLY(uint instr_offset = masm->offset());
1714 n->emit(masm, C->regalloc());
1715 current_offset = masm->offset();
1716
1717 // Above we only verified that there is enough space in the instruction section.
1718 // However, the instruction may emit stubs that cause code buffer expansion.
1719 // Bail out here if expansion failed due to a lack of code cache space.
1720 if (C->failing()) {
1721 return;
1722 }
1723
1724 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1725 "ret_addr_offset() not within emitted code");
1726
1727 #ifdef ASSERT
1728 uint n_size = n->size(C->regalloc());
1729 if (n_size < (current_offset-instr_offset)) {
1730 MachNode* mach = n->as_Mach();
1731 n->dump();
1732 mach->dump_format(C->regalloc(), tty);
1733 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1734 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1735 tty->print_cr(" ------------------- ");
1736 BufferBlob* blob = this->scratch_buffer_blob();
1737 address blob_begin = blob->content_begin();
1738 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1739 assert(false, "wrong size of mach node");
1740 }
1741 #endif
1742 non_safepoints.observe_instruction(n, current_offset);
1743
1744 // mcall is last "call" that can be a safepoint
1745 // record it so we can see if a poll will directly follow it
1746 // in which case we'll need a pad to make the PcDesc sites unique
1747 // see 5010568. This can be slightly inaccurate but conservative
1748 // in the case that return address is not actually at current_offset.
1749 // This is a small price to pay.
1750
1751 if (is_mcall) {
1752 last_call_offset = current_offset;
1753 }
1754
1755 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1756 // Avoid back to back some instructions.
1757 last_avoid_back_to_back_offset = current_offset;
1758 }
1759
1760 // See if this instruction has a delay slot
1761 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1762 guarantee(delay_slot != nullptr, "expecting delay slot node");
1763
1764 // Back up 1 instruction
1765 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1766
1767 // Save the offset for the listing
1768 #if defined(SUPPORT_OPTO_ASSEMBLY)
1769 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1770 node_offsets[delay_slot->_idx] = masm->offset();
1771 }
1772 #endif
1773
1774 // Support a SafePoint in the delay slot
1775 if (delay_slot->is_MachSafePoint()) {
1776 MachNode *mach = delay_slot->as_Mach();
1777 // !!!!! Stubs only need an oopmap right now, so bail out
1778 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1779 // Write the oopmap directly to the code blob??!!
1780 delay_slot = nullptr;
1781 continue;
1782 }
1783
1784 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1785 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1786 adjusted_offset);
1787 // Generate an OopMap entry
1788 Process_OopMap_Node(mach, adjusted_offset);
1789 }
1790
1791 // Insert the delay slot instruction
1792 delay_slot->emit(masm, C->regalloc());
1793
1794 // Don't reuse it
1795 delay_slot = nullptr;
1796 }
1797
1798 } // End for all instructions in block
1799
1800 // If the next block is the top of a loop, pad this block out to align
1801 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1802 if (i < nblocks-1) {
1803 Block *nb = C->cfg()->get_block(i + 1);
1804 int padding = nb->alignment_padding(current_offset);
1805 if( padding > 0 ) {
1806 MachNode *nop = new MachNopNode(padding / nop_size);
1807 block->insert_node(nop, block->number_of_nodes());
1808 C->cfg()->map_node_to_block(nop, block);
1809 nop->emit(masm, C->regalloc());
1810 current_offset = masm->offset();
1811 }
1812 }
1813 // Verify that the distance for generated before forward
1814 // short branches is still valid.
1815 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1816
1817 // Save new block start offset
1818 blk_starts[i] = blk_offset;
1819 } // End of for all blocks
1820 blk_starts[nblocks] = current_offset;
1821
1822 non_safepoints.flush_at_end();
1823
1824 // Offset too large?
1825 if (C->failing()) return;
1826
1827 // Define a pseudo-label at the end of the code
1828 masm->bind( blk_labels[nblocks] );
1829
1830 // Compute the size of the first block
1831 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1832
1833 #ifdef ASSERT
1834 for (uint i = 0; i < nblocks; i++) { // For all blocks
1835 if (jmp_target[i] != 0) {
1836 int br_size = jmp_size[i];
1837 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1838 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1839 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]);
1840 assert(false, "Displacement too large for short jmp");
1841 }
1842 }
1843 }
1844 #endif
1845
1846 if (!masm->code()->finalize_stubs()) {
1847 C->record_failure("CodeCache is full");
1848 return;
1849 }
1850
1851 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1852 bs->emit_stubs(*masm->code());
1853 if (C->failing()) return;
1854
1855 // Fill in stubs.
1856 assert(masm->inst_mark() == nullptr, "should be.");
1857 _stub_list.emit(*masm);
1858 if (C->failing()) return;
1859
1860 #ifndef PRODUCT
1861 // Information on the size of the method, without the extraneous code
1862 Scheduling::increment_method_size(masm->offset());
1863 #endif
1864
1865 // ------------------
1866 // Fill in exception table entries.
1867 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1868
1869 // Only java methods have exception handlers and deopt handlers
1870 // class HandlerImpl is platform-specific and defined in the *.ad files.
1871 if (C->method()) {
1872 // Emit the exception handler code.
1873 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1874 if (C->failing()) {
1875 return; // CodeBuffer::expand failed
1876 }
1877 // Emit the deopt handler code.
1878 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1879
1880 // Emit the MethodHandle deopt handler code (if required).
1881 if (C->has_method_handle_invokes() && !C->failing()) {
1882 // We can use the same code as for the normal deopt handler, we
1883 // just need a different entry point address.
1884 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1885 }
1886 }
1887
1888 // One last check for failed CodeBuffer::expand:
1889 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1890 C->record_failure("CodeCache is full");
1891 return;
1892 }
1893
1894 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1895 if (C->print_assembly()) {
1896 tty->cr();
1897 tty->print_cr("============================= C2-compiled nmethod ==============================");
1898 }
1899 #endif
1900
1901 #if defined(SUPPORT_OPTO_ASSEMBLY)
1902 // Dump the assembly code, including basic-block numbers
1903 if (C->print_assembly()) {
1904 ttyLocker ttyl; // keep the following output all in one block
1905 if (!VMThread::should_terminate()) { // test this under the tty lock
1906 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1907 // We call them first and write to a stringStream, then we retake the lock to
1908 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1909 ResourceMark rm;
1910 stringStream method_metadata_str;
1911 if (C->method() != nullptr) {
1912 C->method()->print_metadata(&method_metadata_str);
1913 }
1914 stringStream dump_asm_str;
1915 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1916
1917 NoSafepointVerifier nsv;
1918 ttyLocker ttyl2;
1919 // This output goes directly to the tty, not the compiler log.
1920 // To enable tools to match it up with the compilation activity,
1921 // be sure to tag this tty output with the compile ID.
1922 if (xtty != nullptr) {
1923 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1924 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1925 }
1926 if (C->method() != nullptr) {
1927 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1928 tty->print_raw(method_metadata_str.freeze());
1929 } else if (C->stub_name() != nullptr) {
1930 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1931 }
1932 tty->cr();
1933 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1934 tty->print_raw(dump_asm_str.freeze());
1935 tty->print_cr("--------------------------------------------------------------------------------");
1936 if (xtty != nullptr) {
1937 xtty->tail("opto_assembly");
1938 }
1939 }
1940 }
1941 #endif
1942 }
1943
1944 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1945 _inc_table.set_size(cnt);
1946
1947 uint inct_cnt = 0;
1948 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1949 Block* block = C->cfg()->get_block(i);
1950 Node *n = nullptr;
1951 int j;
1952
1953 // Find the branch; ignore trailing NOPs.
1954 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1955 n = block->get_node(j);
1956 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1957 break;
1958 }
1959 }
1960
1961 // If we didn't find anything, continue
1962 if (j < 0) {
1963 continue;
1964 }
1965
1966 // Compute ExceptionHandlerTable subtable entry and add it
1967 // (skip empty blocks)
1968 if (n->is_Catch()) {
1969
1970 // Get the offset of the return from the call
1971 uint call_return = call_returns[block->_pre_order];
1972 #ifdef ASSERT
1973 assert( call_return > 0, "no call seen for this basic block" );
1974 while (block->get_node(--j)->is_MachProj()) ;
1975 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1976 #endif
1977 // last instruction is a CatchNode, find it's CatchProjNodes
1978 int nof_succs = block->_num_succs;
1979 // allocate space
1980 GrowableArray<intptr_t> handler_bcis(nof_succs);
1981 GrowableArray<intptr_t> handler_pcos(nof_succs);
1982 // iterate through all successors
1983 for (int j = 0; j < nof_succs; j++) {
1984 Block* s = block->_succs[j];
1985 bool found_p = false;
1986 for (uint k = 1; k < s->num_preds(); k++) {
1987 Node* pk = s->pred(k);
1988 if (pk->is_CatchProj() && pk->in(0) == n) {
1989 const CatchProjNode* p = pk->as_CatchProj();
1990 found_p = true;
1991 // add the corresponding handler bci & pco information
1992 if (p->_con != CatchProjNode::fall_through_index) {
1993 // p leads to an exception handler (and is not fall through)
1994 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1995 // no duplicates, please
1996 if (!handler_bcis.contains(p->handler_bci())) {
1997 uint block_num = s->non_connector()->_pre_order;
1998 handler_bcis.append(p->handler_bci());
1999 handler_pcos.append(blk_labels[block_num].loc_pos());
2000 }
2001 }
2002 }
2003 }
2004 assert(found_p, "no matching predecessor found");
2005 // Note: Due to empty block removal, one block may have
2006 // several CatchProj inputs, from the same Catch.
2007 }
2008
2009 // Set the offset of the return from the call
2010 assert(handler_bcis.find(-1) != -1, "must have default handler");
2011 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2012 continue;
2013 }
2014
2015 // Handle implicit null exception table updates
2016 if (n->is_MachNullCheck()) {
2017 assert(n->in(1)->as_Mach()->barrier_data() == 0,
2018 "Implicit null checks on memory accesses with barriers are not yet supported");
2019 uint block_num = block->non_connector_successor(0)->_pre_order;
2020 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2021 continue;
2022 }
2023 // Handle implicit exception table updates: trap instructions.
2024 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2025 uint block_num = block->non_connector_successor(0)->_pre_order;
2026 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2027 continue;
2028 }
2029 } // End of for all blocks fill in exception table entries
2030 }
2031
2032 // Static Variables
2033 #ifndef PRODUCT
2034 uint Scheduling::_total_nop_size = 0;
2035 uint Scheduling::_total_method_size = 0;
2036 uint Scheduling::_total_branches = 0;
2037 uint Scheduling::_total_unconditional_delays = 0;
2038 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2039 #endif
2040
2041 // Initializer for class Scheduling
2042
2043 Scheduling::Scheduling(Arena *arena, Compile &compile)
2044 : _arena(arena),
2045 _cfg(compile.cfg()),
2046 _regalloc(compile.regalloc()),
2047 _scheduled(arena),
2048 _available(arena),
2049 _reg_node(arena),
2050 _pinch_free_list(arena),
2051 _next_node(nullptr),
2052 _bundle_instr_count(0),
2053 _bundle_cycle_number(0),
2054 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2055 #ifndef PRODUCT
2056 , _branches(0)
2057 , _unconditional_delays(0)
2058 #endif
2059 {
2060 // Create a MachNopNode
2061 _nop = new MachNopNode();
2062
2063 // Now that the nops are in the array, save the count
2064 // (but allow entries for the nops)
2065 _node_bundling_limit = compile.unique();
2066 uint node_max = _regalloc->node_regs_max_index();
2067
2068 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2069
2070 // This one is persistent within the Compile class
2071 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2072
2073 // Allocate space for fixed-size arrays
2074 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2075 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2076
2077 // Clear the arrays
2078 for (uint i = 0; i < node_max; i++) {
2079 ::new (&_node_bundling_base[i]) Bundle();
2080 }
2081 memset(_uses, 0, node_max * sizeof(short));
2082 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2083
2084 // Clear the bundling information
2085 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2086
2087 // Get the last node
2088 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2089
2090 _next_node = block->get_node(block->number_of_nodes() - 1);
2091 }
2092
2093 #ifndef PRODUCT
2094 // Scheduling destructor
2095 Scheduling::~Scheduling() {
2096 _total_branches += _branches;
2097 _total_unconditional_delays += _unconditional_delays;
2098 }
2099 #endif
2100
2101 // Step ahead "i" cycles
2102 void Scheduling::step(uint i) {
2103
2104 Bundle *bundle = node_bundling(_next_node);
2105 bundle->set_starts_bundle();
2106
2107 // Update the bundle record, but leave the flags information alone
2108 if (_bundle_instr_count > 0) {
2109 bundle->set_instr_count(_bundle_instr_count);
2110 bundle->set_resources_used(_bundle_use.resourcesUsed());
2111 }
2112
2113 // Update the state information
2114 _bundle_instr_count = 0;
2115 _bundle_cycle_number += i;
2116 _bundle_use.step(i);
2117 }
2118
2119 void Scheduling::step_and_clear() {
2120 Bundle *bundle = node_bundling(_next_node);
2121 bundle->set_starts_bundle();
2122
2123 // Update the bundle record
2124 if (_bundle_instr_count > 0) {
2125 bundle->set_instr_count(_bundle_instr_count);
2126 bundle->set_resources_used(_bundle_use.resourcesUsed());
2127
2128 _bundle_cycle_number += 1;
2129 }
2130
2131 // Clear the bundling information
2132 _bundle_instr_count = 0;
2133 _bundle_use.reset();
2134
2135 memcpy(_bundle_use_elements,
2136 Pipeline_Use::elaborated_elements,
2137 sizeof(Pipeline_Use::elaborated_elements));
2138 }
2139
2140 // Perform instruction scheduling and bundling over the sequence of
2141 // instructions in backwards order.
2142 void PhaseOutput::ScheduleAndBundle() {
2143
2144 // Don't optimize this if it isn't a method
2145 if (!C->method())
2146 return;
2147
2148 // Don't optimize this if scheduling is disabled
2149 if (!C->do_scheduling())
2150 return;
2151
2152 // Scheduling code works only with pairs (8 bytes) maximum.
2153 // And when the scalable vector register is used, we may spill/unspill
2154 // the whole reg regardless of the max vector size.
2155 if (C->max_vector_size() > 8 ||
2156 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2157 return;
2158 }
2159
2160 Compile::TracePhase tp(_t_instrSched);
2161
2162 // Create a data structure for all the scheduling information
2163 Scheduling scheduling(Thread::current()->resource_area(), *C);
2164
2165 // Walk backwards over each basic block, computing the needed alignment
2166 // Walk over all the basic blocks
2167 scheduling.DoScheduling();
2168
2169 #ifndef PRODUCT
2170 if (C->trace_opto_output()) {
2171 // Buffer and print all at once
2172 ResourceMark rm;
2173 stringStream ss;
2174 ss.print("\n---- After ScheduleAndBundle ----\n");
2175 print_scheduling(&ss);
2176 tty->print("%s", ss.as_string());
2177 }
2178 #endif
2179 }
2180
2181 #ifndef PRODUCT
2182 // Separated out so that it can be called directly from debugger
2183 void PhaseOutput::print_scheduling() {
2184 print_scheduling(tty);
2185 }
2186
2187 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2188 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2189 output_stream->print("\nBB#%03d:\n", i);
2190 Block* block = C->cfg()->get_block(i);
2191 for (uint j = 0; j < block->number_of_nodes(); j++) {
2192 Node* n = block->get_node(j);
2193 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2194 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2195 n->dump("\n", false, output_stream);
2196 }
2197 }
2198 }
2199 #endif
2200
2201 // See if this node fits into the present instruction bundle
2202 bool Scheduling::NodeFitsInBundle(Node *n) {
2203 uint n_idx = n->_idx;
2204
2205 // If this is the unconditional delay instruction, then it fits
2206 if (n == _unconditional_delay_slot) {
2207 #ifndef PRODUCT
2208 if (_cfg->C->trace_opto_output())
2209 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2210 #endif
2211 return (true);
2212 }
2213
2214 // If the node cannot be scheduled this cycle, skip it
2215 if (_current_latency[n_idx] > _bundle_cycle_number) {
2216 #ifndef PRODUCT
2217 if (_cfg->C->trace_opto_output())
2218 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2219 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2220 #endif
2221 return (false);
2222 }
2223
2224 const Pipeline *node_pipeline = n->pipeline();
2225
2226 uint instruction_count = node_pipeline->instructionCount();
2227 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2228 instruction_count = 0;
2229 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2230 instruction_count++;
2231
2232 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2233 #ifndef PRODUCT
2234 if (_cfg->C->trace_opto_output())
2235 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2236 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2237 #endif
2238 return (false);
2239 }
2240
2241 // Don't allow non-machine nodes to be handled this way
2242 if (!n->is_Mach() && instruction_count == 0)
2243 return (false);
2244
2245 // See if there is any overlap
2246 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2247
2248 if (delay > 0) {
2249 #ifndef PRODUCT
2250 if (_cfg->C->trace_opto_output())
2251 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2252 #endif
2253 return false;
2254 }
2255
2256 #ifndef PRODUCT
2257 if (_cfg->C->trace_opto_output())
2258 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2259 #endif
2260
2261 return true;
2262 }
2263
2264 Node * Scheduling::ChooseNodeToBundle() {
2265 uint siz = _available.size();
2266
2267 if (siz == 0) {
2268
2269 #ifndef PRODUCT
2270 if (_cfg->C->trace_opto_output())
2271 tty->print("# ChooseNodeToBundle: null\n");
2272 #endif
2273 return (nullptr);
2274 }
2275
2276 // Fast path, if only 1 instruction in the bundle
2277 if (siz == 1) {
2278 #ifndef PRODUCT
2279 if (_cfg->C->trace_opto_output()) {
2280 tty->print("# ChooseNodeToBundle (only 1): ");
2281 _available[0]->dump();
2282 }
2283 #endif
2284 return (_available[0]);
2285 }
2286
2287 // Don't bother, if the bundle is already full
2288 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2289 for ( uint i = 0; i < siz; i++ ) {
2290 Node *n = _available[i];
2291
2292 // Skip projections, we'll handle them another way
2293 if (n->is_Proj())
2294 continue;
2295
2296 // This presupposed that instructions are inserted into the
2297 // available list in a legality order; i.e. instructions that
2298 // must be inserted first are at the head of the list
2299 if (NodeFitsInBundle(n)) {
2300 #ifndef PRODUCT
2301 if (_cfg->C->trace_opto_output()) {
2302 tty->print("# ChooseNodeToBundle: ");
2303 n->dump();
2304 }
2305 #endif
2306 return (n);
2307 }
2308 }
2309 }
2310
2311 // Nothing fits in this bundle, choose the highest priority
2312 #ifndef PRODUCT
2313 if (_cfg->C->trace_opto_output()) {
2314 tty->print("# ChooseNodeToBundle: ");
2315 _available[0]->dump();
2316 }
2317 #endif
2318
2319 return _available[0];
2320 }
2321
2322 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2323 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2324
2325 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2326 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2327 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2328 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2329
2330 // Comparison between stack -> reg and stack -> reg
2331 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2332 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2333 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2334 }
2335
2336 // Comparison between reg -> stack and reg -> stack
2337 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2338 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2339 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2340 }
2341
2342 return 0; // Not comparable
2343 }
2344
2345 void Scheduling::AddNodeToAvailableList(Node *n) {
2346 assert( !n->is_Proj(), "projections never directly made available" );
2347 #ifndef PRODUCT
2348 if (_cfg->C->trace_opto_output()) {
2349 tty->print("# AddNodeToAvailableList: ");
2350 n->dump();
2351 }
2352 #endif
2353
2354 int latency = _current_latency[n->_idx];
2355
2356 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2357 // for stack spilling or unspilling have the same latency, we sort
2358 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2359 // have more opportunities to do ld/st merging
2360 uint i;
2361 for (i = 0; i < _available.size(); i++) {
2362 if (_current_latency[_available[i]->_idx] > latency) {
2363 break;
2364 } else if (_current_latency[_available[i]->_idx] == latency &&
2365 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2366 compare_two_spill_nodes(n, _available[i]) > 0) {
2367 break;
2368 }
2369 }
2370
2371 // Special Check for compares following branches
2372 if( n->is_Mach() && _scheduled.size() > 0 ) {
2373 int op = n->as_Mach()->ideal_Opcode();
2374 Node *last = _scheduled[0];
2375 if( last->is_MachIf() && last->in(1) == n &&
2376 ( op == Op_CmpI ||
2377 op == Op_CmpU ||
2378 op == Op_CmpUL ||
2379 op == Op_CmpP ||
2380 op == Op_CmpF ||
2381 op == Op_CmpD ||
2382 op == Op_CmpL ) ) {
2383
2384 // Recalculate position, moving to front of same latency
2385 for ( i=0 ; i < _available.size(); i++ )
2386 if (_current_latency[_available[i]->_idx] >= latency)
2387 break;
2388 }
2389 }
2390
2391 // Insert the node in the available list
2392 _available.insert(i, n);
2393
2394 #ifndef PRODUCT
2395 if (_cfg->C->trace_opto_output())
2396 dump_available();
2397 #endif
2398 }
2399
2400 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2401 for ( uint i=0; i < n->len(); i++ ) {
2402 Node *def = n->in(i);
2403 if (!def) continue;
2404 if( def->is_Proj() ) // If this is a machine projection, then
2405 def = def->in(0); // propagate usage thru to the base instruction
2406
2407 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2408 continue;
2409 }
2410
2411 // Compute the latency
2412 uint l = _bundle_cycle_number + n->latency(i);
2413 if (_current_latency[def->_idx] < l)
2414 _current_latency[def->_idx] = l;
2415
2416 // If this does not have uses then schedule it
2417 if ((--_uses[def->_idx]) == 0)
2418 AddNodeToAvailableList(def);
2419 }
2420 }
2421
2422 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2423 #ifndef PRODUCT
2424 if (_cfg->C->trace_opto_output()) {
2425 tty->print("# AddNodeToBundle: ");
2426 n->dump();
2427 }
2428 #endif
2429
2430 // Remove this from the available list
2431 uint i;
2432 for (i = 0; i < _available.size(); i++)
2433 if (_available[i] == n)
2434 break;
2435 assert(i < _available.size(), "entry in _available list not found");
2436 _available.remove(i);
2437
2438 // See if this fits in the current bundle
2439 const Pipeline *node_pipeline = n->pipeline();
2440 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2441
2442 // Check for instructions to be placed in the delay slot. We
2443 // do this before we actually schedule the current instruction,
2444 // because the delay slot follows the current instruction.
2445 if (Pipeline::_branch_has_delay_slot &&
2446 node_pipeline->hasBranchDelay() &&
2447 !_unconditional_delay_slot) {
2448
2449 uint siz = _available.size();
2450
2451 // Conditional branches can support an instruction that
2452 // is unconditionally executed and not dependent by the
2453 // branch, OR a conditionally executed instruction if
2454 // the branch is taken. In practice, this means that
2455 // the first instruction at the branch target is
2456 // copied to the delay slot, and the branch goes to
2457 // the instruction after that at the branch target
2458 if ( n->is_MachBranch() ) {
2459
2460 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2461 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2462
2463 #ifndef PRODUCT
2464 _branches++;
2465 #endif
2466
2467 // At least 1 instruction is on the available list
2468 // that is not dependent on the branch
2469 for (uint i = 0; i < siz; i++) {
2470 Node *d = _available[i];
2471 const Pipeline *avail_pipeline = d->pipeline();
2472
2473 // Don't allow safepoints in the branch shadow, that will
2474 // cause a number of difficulties
2475 if ( avail_pipeline->instructionCount() == 1 &&
2476 !avail_pipeline->hasMultipleBundles() &&
2477 !avail_pipeline->hasBranchDelay() &&
2478 Pipeline::instr_has_unit_size() &&
2479 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2480 NodeFitsInBundle(d) &&
2481 !node_bundling(d)->used_in_delay()) {
2482
2483 if (d->is_Mach() && !d->is_MachSafePoint()) {
2484 // A node that fits in the delay slot was found, so we need to
2485 // set the appropriate bits in the bundle pipeline information so
2486 // that it correctly indicates resource usage. Later, when we
2487 // attempt to add this instruction to the bundle, we will skip
2488 // setting the resource usage.
2489 _unconditional_delay_slot = d;
2490 node_bundling(n)->set_use_unconditional_delay();
2491 node_bundling(d)->set_used_in_unconditional_delay();
2492 _bundle_use.add_usage(avail_pipeline->resourceUse());
2493 _current_latency[d->_idx] = _bundle_cycle_number;
2494 _next_node = d;
2495 ++_bundle_instr_count;
2496 #ifndef PRODUCT
2497 _unconditional_delays++;
2498 #endif
2499 break;
2500 }
2501 }
2502 }
2503 }
2504
2505 // No delay slot, add a nop to the usage
2506 if (!_unconditional_delay_slot) {
2507 // See if adding an instruction in the delay slot will overflow
2508 // the bundle.
2509 if (!NodeFitsInBundle(_nop)) {
2510 #ifndef PRODUCT
2511 if (_cfg->C->trace_opto_output())
2512 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2513 #endif
2514 step(1);
2515 }
2516
2517 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2518 _next_node = _nop;
2519 ++_bundle_instr_count;
2520 }
2521
2522 // See if the instruction in the delay slot requires a
2523 // step of the bundles
2524 if (!NodeFitsInBundle(n)) {
2525 #ifndef PRODUCT
2526 if (_cfg->C->trace_opto_output())
2527 tty->print("# *** STEP(branch won't fit) ***\n");
2528 #endif
2529 // Update the state information
2530 _bundle_instr_count = 0;
2531 _bundle_cycle_number += 1;
2532 _bundle_use.step(1);
2533 }
2534 }
2535
2536 // Get the number of instructions
2537 uint instruction_count = node_pipeline->instructionCount();
2538 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2539 instruction_count = 0;
2540
2541 // Compute the latency information
2542 uint delay = 0;
2543
2544 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2545 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2546 if (relative_latency < 0)
2547 relative_latency = 0;
2548
2549 delay = _bundle_use.full_latency(relative_latency, node_usage);
2550
2551 // Does not fit in this bundle, start a new one
2552 if (delay > 0) {
2553 step(delay);
2554
2555 #ifndef PRODUCT
2556 if (_cfg->C->trace_opto_output())
2557 tty->print("# *** STEP(%d) ***\n", delay);
2558 #endif
2559 }
2560 }
2561
2562 // If this was placed in the delay slot, ignore it
2563 if (n != _unconditional_delay_slot) {
2564
2565 if (delay == 0) {
2566 if (node_pipeline->hasMultipleBundles()) {
2567 #ifndef PRODUCT
2568 if (_cfg->C->trace_opto_output())
2569 tty->print("# *** STEP(multiple instructions) ***\n");
2570 #endif
2571 step(1);
2572 }
2573
2574 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2575 #ifndef PRODUCT
2576 if (_cfg->C->trace_opto_output())
2577 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2578 instruction_count + _bundle_instr_count,
2579 Pipeline::_max_instrs_per_cycle);
2580 #endif
2581 step(1);
2582 }
2583 }
2584
2585 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2586 _bundle_instr_count++;
2587
2588 // Set the node's latency
2589 _current_latency[n->_idx] = _bundle_cycle_number;
2590
2591 // Now merge the functional unit information
2592 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2593 _bundle_use.add_usage(node_usage);
2594
2595 // Increment the number of instructions in this bundle
2596 _bundle_instr_count += instruction_count;
2597
2598 // Remember this node for later
2599 if (n->is_Mach())
2600 _next_node = n;
2601 }
2602
2603 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2604 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2605 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2606 // into the block. All other scheduled nodes get put in the schedule here.
2607 int op = n->Opcode();
2608 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2609 (op != Op_Node && // Not an unused antidepedence node and
2610 // not an unallocated boxlock
2611 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2612
2613 // Push any trailing projections
2614 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2615 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2616 Node *foi = n->fast_out(i);
2617 if( foi->is_Proj() )
2618 _scheduled.push(foi);
2619 }
2620 }
2621
2622 // Put the instruction in the schedule list
2623 _scheduled.push(n);
2624 }
2625
2626 #ifndef PRODUCT
2627 if (_cfg->C->trace_opto_output())
2628 dump_available();
2629 #endif
2630
2631 // Walk all the definitions, decrementing use counts, and
2632 // if a definition has a 0 use count, place it in the available list.
2633 DecrementUseCounts(n,bb);
2634 }
2635
2636 // This method sets the use count within a basic block. We will ignore all
2637 // uses outside the current basic block. As we are doing a backwards walk,
2638 // any node we reach that has a use count of 0 may be scheduled. This also
2639 // avoids the problem of cyclic references from phi nodes, as long as phi
2640 // nodes are at the front of the basic block. This method also initializes
2641 // the available list to the set of instructions that have no uses within this
2642 // basic block.
2643 void Scheduling::ComputeUseCount(const Block *bb) {
2644 #ifndef PRODUCT
2645 if (_cfg->C->trace_opto_output())
2646 tty->print("# -> ComputeUseCount\n");
2647 #endif
2648
2649 // Clear the list of available and scheduled instructions, just in case
2650 _available.clear();
2651 _scheduled.clear();
2652
2653 // No delay slot specified
2654 _unconditional_delay_slot = nullptr;
2655
2656 #ifdef ASSERT
2657 for( uint i=0; i < bb->number_of_nodes(); i++ )
2658 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2659 #endif
2660
2661 // Force the _uses count to never go to zero for unscheduable pieces
2662 // of the block
2663 for( uint k = 0; k < _bb_start; k++ )
2664 _uses[bb->get_node(k)->_idx] = 1;
2665 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2666 _uses[bb->get_node(l)->_idx] = 1;
2667
2668 // Iterate backwards over the instructions in the block. Don't count the
2669 // branch projections at end or the block header instructions.
2670 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2671 Node *n = bb->get_node(j);
2672 if( n->is_Proj() ) continue; // Projections handled another way
2673
2674 // Account for all uses
2675 for ( uint k = 0; k < n->len(); k++ ) {
2676 Node *inp = n->in(k);
2677 if (!inp) continue;
2678 assert(inp != n, "no cycles allowed" );
2679 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2680 if (inp->is_Proj()) { // Skip through Proj's
2681 inp = inp->in(0);
2682 }
2683 ++_uses[inp->_idx]; // Count 1 block-local use
2684 }
2685 }
2686
2687 // If this instruction has a 0 use count, then it is available
2688 if (!_uses[n->_idx]) {
2689 _current_latency[n->_idx] = _bundle_cycle_number;
2690 AddNodeToAvailableList(n);
2691 }
2692
2693 #ifndef PRODUCT
2694 if (_cfg->C->trace_opto_output()) {
2695 tty->print("# uses: %3d: ", _uses[n->_idx]);
2696 n->dump();
2697 }
2698 #endif
2699 }
2700
2701 #ifndef PRODUCT
2702 if (_cfg->C->trace_opto_output())
2703 tty->print("# <- ComputeUseCount\n");
2704 #endif
2705 }
2706
2707 // This routine performs scheduling on each basic block in reverse order,
2708 // using instruction latencies and taking into account function unit
2709 // availability.
2710 void Scheduling::DoScheduling() {
2711 #ifndef PRODUCT
2712 if (_cfg->C->trace_opto_output())
2713 tty->print("# -> DoScheduling\n");
2714 #endif
2715
2716 Block *succ_bb = nullptr;
2717 Block *bb;
2718 Compile* C = Compile::current();
2719
2720 // Walk over all the basic blocks in reverse order
2721 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2722 bb = _cfg->get_block(i);
2723
2724 #ifndef PRODUCT
2725 if (_cfg->C->trace_opto_output()) {
2726 tty->print("# Schedule BB#%03d (initial)\n", i);
2727 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2728 bb->get_node(j)->dump();
2729 }
2730 }
2731 #endif
2732
2733 // On the head node, skip processing
2734 if (bb == _cfg->get_root_block()) {
2735 continue;
2736 }
2737
2738 // Skip empty, connector blocks
2739 if (bb->is_connector())
2740 continue;
2741
2742 // If the following block is not the sole successor of
2743 // this one, then reset the pipeline information
2744 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2745 #ifndef PRODUCT
2746 if (_cfg->C->trace_opto_output()) {
2747 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2748 _next_node->_idx, _bundle_instr_count);
2749 }
2750 #endif
2751 step_and_clear();
2752 }
2753
2754 // Leave untouched the starting instruction, any Phis, a CreateEx node
2755 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2756 _bb_end = bb->number_of_nodes()-1;
2757 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2758 Node *n = bb->get_node(_bb_start);
2759 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2760 // Also, MachIdealNodes do not get scheduled
2761 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2762 MachNode *mach = n->as_Mach();
2763 int iop = mach->ideal_Opcode();
2764 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2765 if( iop == Op_Con ) continue; // Do not schedule Top
2766 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2767 mach->pipeline() == MachNode::pipeline_class() &&
2768 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2769 continue;
2770 break; // Funny loop structure to be sure...
2771 }
2772 // Compute last "interesting" instruction in block - last instruction we
2773 // might schedule. _bb_end points just after last schedulable inst. We
2774 // normally schedule conditional branches (despite them being forced last
2775 // in the block), because they have delay slots we can fill. Calls all
2776 // have their delay slots filled in the template expansions, so we don't
2777 // bother scheduling them.
2778 Node *last = bb->get_node(_bb_end);
2779 // Ignore trailing NOPs.
2780 while (_bb_end > 0 && last->is_Mach() &&
2781 last->as_Mach()->ideal_Opcode() == Op_Con) {
2782 last = bb->get_node(--_bb_end);
2783 }
2784 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2785 if( last->is_Catch() ||
2786 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2787 // There might be a prior call. Skip it.
2788 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2789 } else if( last->is_MachNullCheck() ) {
2790 // Backup so the last null-checked memory instruction is
2791 // outside the schedulable range. Skip over the nullcheck,
2792 // projection, and the memory nodes.
2793 Node *mem = last->in(1);
2794 do {
2795 _bb_end--;
2796 } while (mem != bb->get_node(_bb_end));
2797 } else {
2798 // Set _bb_end to point after last schedulable inst.
2799 _bb_end++;
2800 }
2801
2802 assert( _bb_start <= _bb_end, "inverted block ends" );
2803
2804 // Compute the register antidependencies for the basic block
2805 ComputeRegisterAntidependencies(bb);
2806 if (C->failing()) return; // too many D-U pinch points
2807
2808 // Compute the usage within the block, and set the list of all nodes
2809 // in the block that have no uses within the block.
2810 ComputeUseCount(bb);
2811
2812 // Schedule the remaining instructions in the block
2813 while ( _available.size() > 0 ) {
2814 Node *n = ChooseNodeToBundle();
2815 guarantee(n != nullptr, "no nodes available");
2816 AddNodeToBundle(n,bb);
2817 }
2818
2819 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2820 #ifdef ASSERT
2821 for( uint l = _bb_start; l < _bb_end; l++ ) {
2822 Node *n = bb->get_node(l);
2823 uint m;
2824 for( m = 0; m < _bb_end-_bb_start; m++ )
2825 if( _scheduled[m] == n )
2826 break;
2827 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2828 }
2829 #endif
2830
2831 // Now copy the instructions (in reverse order) back to the block
2832 for ( uint k = _bb_start; k < _bb_end; k++ )
2833 bb->map_node(_scheduled[_bb_end-k-1], k);
2834
2835 #ifndef PRODUCT
2836 if (_cfg->C->trace_opto_output()) {
2837 tty->print("# Schedule BB#%03d (final)\n", i);
2838 uint current = 0;
2839 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2840 Node *n = bb->get_node(j);
2841 if( valid_bundle_info(n) ) {
2842 Bundle *bundle = node_bundling(n);
2843 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2844 tty->print("*** Bundle: ");
2845 bundle->dump();
2846 }
2847 n->dump();
2848 }
2849 }
2850 }
2851 #endif
2852 #ifdef ASSERT
2853 verify_good_schedule(bb,"after block local scheduling");
2854 #endif
2855 }
2856
2857 #ifndef PRODUCT
2858 if (_cfg->C->trace_opto_output())
2859 tty->print("# <- DoScheduling\n");
2860 #endif
2861
2862 // Record final node-bundling array location
2863 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2864
2865 } // end DoScheduling
2866
2867 // Verify that no live-range used in the block is killed in the block by a
2868 // wrong DEF. This doesn't verify live-ranges that span blocks.
2869
2870 // Check for edge existence. Used to avoid adding redundant precedence edges.
2871 static bool edge_from_to( Node *from, Node *to ) {
2872 for( uint i=0; i<from->len(); i++ )
2873 if( from->in(i) == to )
2874 return true;
2875 return false;
2876 }
2877
2878 #ifdef ASSERT
2879 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2880 // Check for bad kills
2881 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2882 Node *prior_use = _reg_node[def];
2883 if( prior_use && !edge_from_to(prior_use,n) ) {
2884 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2885 n->dump();
2886 tty->print_cr("...");
2887 prior_use->dump();
2888 assert(edge_from_to(prior_use,n), "%s", msg);
2889 }
2890 _reg_node.map(def,nullptr); // Kill live USEs
2891 }
2892 }
2893
2894 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2895
2896 // Zap to something reasonable for the verify code
2897 _reg_node.clear();
2898
2899 // Walk over the block backwards. Check to make sure each DEF doesn't
2900 // kill a live value (other than the one it's supposed to). Add each
2901 // USE to the live set.
2902 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2903 Node *n = b->get_node(i);
2904 int n_op = n->Opcode();
2905 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2906 // Fat-proj kills a slew of registers
2907 RegMaskIterator rmi(n->out_RegMask());
2908 while (rmi.has_next()) {
2909 OptoReg::Name kill = rmi.next();
2910 verify_do_def(n, kill, msg);
2911 }
2912 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2913 // Get DEF'd registers the normal way
2914 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2915 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2916 }
2917
2918 // Now make all USEs live
2919 for( uint i=1; i<n->req(); i++ ) {
2920 Node *def = n->in(i);
2921 assert(def != nullptr, "input edge required");
2922 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2923 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2924 if( OptoReg::is_valid(reg_lo) ) {
2925 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2926 _reg_node.map(reg_lo,n);
2927 }
2928 if( OptoReg::is_valid(reg_hi) ) {
2929 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2930 _reg_node.map(reg_hi,n);
2931 }
2932 }
2933
2934 }
2935
2936 // Zap to something reasonable for the Antidependence code
2937 _reg_node.clear();
2938 }
2939 #endif
2940
2941 // Conditionally add precedence edges. Avoid putting edges on Projs.
2942 static void add_prec_edge_from_to( Node *from, Node *to ) {
2943 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2944 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2945 from = from->in(0);
2946 }
2947 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2948 !edge_from_to( from, to ) ) // Avoid duplicate edge
2949 from->add_prec(to);
2950 }
2951
2952 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2953 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2954 return;
2955
2956 if (OptoReg::is_reg(def_reg)) {
2957 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2958 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2959 // This is one of the high slots of a vector register.
2960 // ScheduleAndBundle already checked there are no live wide
2961 // vectors in this method so it can be safely ignored.
2962 return;
2963 }
2964 }
2965
2966 Node *pinch = _reg_node[def_reg]; // Get pinch point
2967 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2968 is_def ) { // Check for a true def (not a kill)
2969 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2970 return;
2971 }
2972
2973 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2974 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2975
2976 // After some number of kills there _may_ be a later def
2977 Node *later_def = nullptr;
2978
2979 Compile* C = Compile::current();
2980
2981 // Finding a kill requires a real pinch-point.
2982 // Check for not already having a pinch-point.
2983 // Pinch points are Op_Node's.
2984 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2985 later_def = pinch; // Must be def/kill as optimistic pinch-point
2986 if ( _pinch_free_list.size() > 0) {
2987 pinch = _pinch_free_list.pop();
2988 } else {
2989 pinch = new Node(1); // Pinch point to-be
2990 }
2991 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2992 DEBUG_ONLY( pinch->dump(); );
2993 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2994 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2995 return;
2996 }
2997 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2998 _reg_node.map(def_reg,pinch); // Record pinch-point
2999 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3000 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3001 pinch->init_req(0, C->top()); // set not null for the next call
3002 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3003 later_def = nullptr; // and no later def
3004 }
3005 pinch->set_req(0,later_def); // Hook later def so we can find it
3006 } else { // Else have valid pinch point
3007 if( pinch->in(0) ) // If there is a later-def
3008 later_def = pinch->in(0); // Get it
3009 }
3010
3011 // Add output-dependence edge from later def to kill
3012 if( later_def ) // If there is some original def
3013 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3014
3015 // See if current kill is also a use, and so is forced to be the pinch-point.
3016 if( pinch->Opcode() == Op_Node ) {
3017 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3018 for( uint i=1; i<uses->req(); i++ ) {
3019 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3020 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3021 // Yes, found a use/kill pinch-point
3022 pinch->set_req(0,nullptr); //
3023 pinch->replace_by(kill); // Move anti-dep edges up
3024 pinch = kill;
3025 _reg_node.map(def_reg,pinch);
3026 return;
3027 }
3028 }
3029 }
3030
3031 // Add edge from kill to pinch-point
3032 add_prec_edge_from_to(kill,pinch);
3033 }
3034
3035 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3036 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3037 return;
3038 Node *pinch = _reg_node[use_reg]; // Get pinch point
3039 // Check for no later def_reg/kill in block
3040 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3041 // Use has to be block-local as well
3042 _cfg->get_block_for_node(use) == b) {
3043 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3044 pinch->req() == 1 ) { // pinch not yet in block?
3045 pinch->del_req(0); // yank pointer to later-def, also set flag
3046 // Insert the pinch-point in the block just after the last use
3047 b->insert_node(pinch, b->find_node(use) + 1);
3048 _bb_end++; // Increase size scheduled region in block
3049 }
3050
3051 add_prec_edge_from_to(pinch,use);
3052 }
3053 }
3054
3055 // We insert antidependences between the reads and following write of
3056 // allocated registers to prevent illegal code motion. Hopefully, the
3057 // number of added references should be fairly small, especially as we
3058 // are only adding references within the current basic block.
3059 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3060
3061 #ifdef ASSERT
3062 verify_good_schedule(b,"before block local scheduling");
3063 #endif
3064
3065 // A valid schedule, for each register independently, is an endless cycle
3066 // of: a def, then some uses (connected to the def by true dependencies),
3067 // then some kills (defs with no uses), finally the cycle repeats with a new
3068 // def. The uses are allowed to float relative to each other, as are the
3069 // kills. No use is allowed to slide past a kill (or def). This requires
3070 // antidependencies between all uses of a single def and all kills that
3071 // follow, up to the next def. More edges are redundant, because later defs
3072 // & kills are already serialized with true or antidependencies. To keep
3073 // the edge count down, we add a 'pinch point' node if there's more than
3074 // one use or more than one kill/def.
3075
3076 // We add dependencies in one bottom-up pass.
3077
3078 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3079
3080 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3081 // register. If not, we record the DEF/KILL in _reg_node, the
3082 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3083 // "pinch point", a new Node that's in the graph but not in the block.
3084 // We put edges from the prior and current DEF/KILLs to the pinch point.
3085 // We put the pinch point in _reg_node. If there's already a pinch point
3086 // we merely add an edge from the current DEF/KILL to the pinch point.
3087
3088 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3089 // put an edge from the pinch point to the USE.
3090
3091 // To be expedient, the _reg_node array is pre-allocated for the whole
3092 // compilation. _reg_node is lazily initialized; it either contains a null,
3093 // or a valid def/kill/pinch-point, or a leftover node from some prior
3094 // block. Leftover node from some prior block is treated like a null (no
3095 // prior def, so no anti-dependence needed). Valid def is distinguished by
3096 // it being in the current block.
3097 bool fat_proj_seen = false;
3098 uint last_safept = _bb_end-1;
3099 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3100 Node* last_safept_node = end_node;
3101 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3102 Node *n = b->get_node(i);
3103 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3104 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3105 // Fat-proj kills a slew of registers
3106 // This can add edges to 'n' and obscure whether or not it was a def,
3107 // hence the is_def flag.
3108 fat_proj_seen = true;
3109 RegMaskIterator rmi(n->out_RegMask());
3110 while (rmi.has_next()) {
3111 OptoReg::Name kill = rmi.next();
3112 anti_do_def(b, n, kill, is_def);
3113 }
3114 } else {
3115 // Get DEF'd registers the normal way
3116 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3117 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3118 }
3119
3120 // Kill projections on a branch should appear to occur on the
3121 // branch, not afterwards, so grab the masks from the projections
3122 // and process them.
3123 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3124 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3125 Node* use = n->fast_out(i);
3126 if (use->is_Proj()) {
3127 RegMaskIterator rmi(use->out_RegMask());
3128 while (rmi.has_next()) {
3129 OptoReg::Name kill = rmi.next();
3130 anti_do_def(b, n, kill, false);
3131 }
3132 }
3133 }
3134 }
3135
3136 // Check each register used by this instruction for a following DEF/KILL
3137 // that must occur afterward and requires an anti-dependence edge.
3138 for( uint j=0; j<n->req(); j++ ) {
3139 Node *def = n->in(j);
3140 if( def ) {
3141 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3142 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3143 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3144 }
3145 }
3146 // Do not allow defs of new derived values to float above GC
3147 // points unless the base is definitely available at the GC point.
3148
3149 Node *m = b->get_node(i);
3150
3151 // Add precedence edge from following safepoint to use of derived pointer
3152 if( last_safept_node != end_node &&
3153 m != last_safept_node) {
3154 for (uint k = 1; k < m->req(); k++) {
3155 const Type *t = m->in(k)->bottom_type();
3156 if( t->isa_oop_ptr() &&
3157 t->is_ptr()->offset() != 0 ) {
3158 last_safept_node->add_prec( m );
3159 break;
3160 }
3161 }
3162 }
3163
3164 if( n->jvms() ) { // Precedence edge from derived to safept
3165 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3166 if( b->get_node(last_safept) != last_safept_node ) {
3167 last_safept = b->find_node(last_safept_node);
3168 }
3169 for( uint j=last_safept; j > i; j-- ) {
3170 Node *mach = b->get_node(j);
3171 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3172 mach->add_prec( n );
3173 }
3174 last_safept = i;
3175 last_safept_node = m;
3176 }
3177 }
3178
3179 if (fat_proj_seen) {
3180 // Garbage collect pinch nodes that were not consumed.
3181 // They are usually created by a fat kill MachProj for a call.
3182 garbage_collect_pinch_nodes();
3183 }
3184 }
3185
3186 // Garbage collect pinch nodes for reuse by other blocks.
3187 //
3188 // The block scheduler's insertion of anti-dependence
3189 // edges creates many pinch nodes when the block contains
3190 // 2 or more Calls. A pinch node is used to prevent a
3191 // combinatorial explosion of edges. If a set of kills for a
3192 // register is anti-dependent on a set of uses (or defs), rather
3193 // than adding an edge in the graph between each pair of kill
3194 // and use (or def), a pinch is inserted between them:
3195 //
3196 // use1 use2 use3
3197 // \ | /
3198 // \ | /
3199 // pinch
3200 // / | \
3201 // / | \
3202 // kill1 kill2 kill3
3203 //
3204 // One pinch node is created per register killed when
3205 // the second call is encountered during a backwards pass
3206 // over the block. Most of these pinch nodes are never
3207 // wired into the graph because the register is never
3208 // used or def'ed in the block.
3209 //
3210 void Scheduling::garbage_collect_pinch_nodes() {
3211 #ifndef PRODUCT
3212 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3213 #endif
3214 int trace_cnt = 0;
3215 for (uint k = 0; k < _reg_node.max(); k++) {
3216 Node* pinch = _reg_node[k];
3217 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3218 // no predecence input edges
3219 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3220 cleanup_pinch(pinch);
3221 _pinch_free_list.push(pinch);
3222 _reg_node.map(k, nullptr);
3223 #ifndef PRODUCT
3224 if (_cfg->C->trace_opto_output()) {
3225 trace_cnt++;
3226 if (trace_cnt > 40) {
3227 tty->print("\n");
3228 trace_cnt = 0;
3229 }
3230 tty->print(" %d", pinch->_idx);
3231 }
3232 #endif
3233 }
3234 }
3235 #ifndef PRODUCT
3236 if (_cfg->C->trace_opto_output()) tty->print("\n");
3237 #endif
3238 }
3239
3240 // Clean up a pinch node for reuse.
3241 void Scheduling::cleanup_pinch( Node *pinch ) {
3242 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3243
3244 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3245 Node* use = pinch->last_out(i);
3246 uint uses_found = 0;
3247 for (uint j = use->req(); j < use->len(); j++) {
3248 if (use->in(j) == pinch) {
3249 use->rm_prec(j);
3250 uses_found++;
3251 }
3252 }
3253 assert(uses_found > 0, "must be a precedence edge");
3254 i -= uses_found; // we deleted 1 or more copies of this edge
3255 }
3256 // May have a later_def entry
3257 pinch->set_req(0, nullptr);
3258 }
3259
3260 #ifndef PRODUCT
3261
3262 void Scheduling::dump_available() const {
3263 tty->print("#Availist ");
3264 for (uint i = 0; i < _available.size(); i++)
3265 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3266 tty->cr();
3267 }
3268
3269 // Print Scheduling Statistics
3270 void Scheduling::print_statistics() {
3271 // Print the size added by nops for bundling
3272 tty->print("Nops added %d bytes to total of %d bytes",
3273 _total_nop_size, _total_method_size);
3274 if (_total_method_size > 0)
3275 tty->print(", for %.2f%%",
3276 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3277 tty->print("\n");
3278
3279 // Print the number of branch shadows filled
3280 if (Pipeline::_branch_has_delay_slot) {
3281 tty->print("Of %d branches, %d had unconditional delay slots filled",
3282 _total_branches, _total_unconditional_delays);
3283 if (_total_branches > 0)
3284 tty->print(", for %.2f%%",
3285 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3286 tty->print("\n");
3287 }
3288
3289 uint total_instructions = 0, total_bundles = 0;
3290
3291 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3292 uint bundle_count = _total_instructions_per_bundle[i];
3293 total_instructions += bundle_count * i;
3294 total_bundles += bundle_count;
3295 }
3296
3297 if (total_bundles > 0)
3298 tty->print("Average ILP (excluding nops) is %.2f\n",
3299 ((double)total_instructions) / ((double)total_bundles));
3300 }
3301 #endif
3302
3303 //-----------------------init_scratch_buffer_blob------------------------------
3304 // Construct a temporary BufferBlob and cache it for this compile.
3305 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3306 // If there is already a scratch buffer blob allocated and the
3307 // constant section is big enough, use it. Otherwise free the
3308 // current and allocate a new one.
3309 BufferBlob* blob = scratch_buffer_blob();
3310 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3311 // Use the current blob.
3312 } else {
3313 if (blob != nullptr) {
3314 BufferBlob::free(blob);
3315 }
3316
3317 ResourceMark rm;
3318 _scratch_const_size = const_size;
3319 int size = C2Compiler::initial_code_buffer_size(const_size);
3320 blob = BufferBlob::create("Compile::scratch_buffer", size);
3321 // Record the buffer blob for next time.
3322 set_scratch_buffer_blob(blob);
3323 // Have we run out of code space?
3324 if (scratch_buffer_blob() == nullptr) {
3325 // Let CompilerBroker disable further compilations.
3326 C->record_failure("Not enough space for scratch buffer in CodeCache");
3327 return;
3328 }
3329 }
3330
3331 // Initialize the relocation buffers
3332 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3333 set_scratch_locs_memory(locs_buf);
3334 }
3335
3336
3337 //-----------------------scratch_emit_size-------------------------------------
3338 // Helper function that computes size by emitting code
3339 uint PhaseOutput::scratch_emit_size(const Node* n) {
3340 // Start scratch_emit_size section.
3341 set_in_scratch_emit_size(true);
3342
3343 // Emit into a trash buffer and count bytes emitted.
3344 // This is a pretty expensive way to compute a size,
3345 // but it works well enough if seldom used.
3346 // All common fixed-size instructions are given a size
3347 // method by the AD file.
3348 // Note that the scratch buffer blob and locs memory are
3349 // allocated at the beginning of the compile task, and
3350 // may be shared by several calls to scratch_emit_size.
3351 // The allocation of the scratch buffer blob is particularly
3352 // expensive, since it has to grab the code cache lock.
3353 BufferBlob* blob = this->scratch_buffer_blob();
3354 assert(blob != nullptr, "Initialize BufferBlob at start");
3355 assert(blob->size() > MAX_inst_size, "sanity");
3356 relocInfo* locs_buf = scratch_locs_memory();
3357 address blob_begin = blob->content_begin();
3358 address blob_end = (address)locs_buf;
3359 assert(blob->contains(blob_end), "sanity");
3360 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3361 buf.initialize_consts_size(_scratch_const_size);
3362 buf.initialize_stubs_size(MAX_stubs_size);
3363 assert(locs_buf != nullptr, "sanity");
3364 int lsize = MAX_locs_size / 3;
3365 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3366 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3367 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3368 // Mark as scratch buffer.
3369 buf.consts()->set_scratch_emit();
3370 buf.insts()->set_scratch_emit();
3371 buf.stubs()->set_scratch_emit();
3372
3373 // Do the emission.
3374
3375 Label fakeL; // Fake label for branch instructions.
3376 Label* saveL = nullptr;
3377 uint save_bnum = 0;
3378 bool is_branch = n->is_MachBranch();
3379 C2_MacroAssembler masm(&buf);
3380 masm.bind(fakeL);
3381 if (is_branch) {
3382 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3383 n->as_MachBranch()->label_set(&fakeL, 0);
3384 }
3385 n->emit(&masm, C->regalloc());
3386
3387 // Emitting into the scratch buffer should not fail
3388 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3389
3390 if (is_branch) // Restore label.
3391 n->as_MachBranch()->label_set(saveL, save_bnum);
3392
3393 // End scratch_emit_size section.
3394 set_in_scratch_emit_size(false);
3395
3396 return buf.insts_size();
3397 }
3398
3399 void PhaseOutput::install() {
3400 if (!C->should_install_code()) {
3401 return;
3402 } else if (C->stub_function() != nullptr) {
3403 install_stub(C->stub_name());
3404 } else {
3405 install_code(C->method(),
3406 C->entry_bci(),
3407 CompileBroker::compiler2(),
3408 C->has_unsafe_access(),
3409 SharedRuntime::is_wide_vector(C->max_vector_size()));
3410 }
3411 }
3412
3413 void PhaseOutput::install_code(ciMethod* target,
3414 int entry_bci,
3415 AbstractCompiler* compiler,
3416 bool has_unsafe_access,
3417 bool has_wide_vectors) {
3418 // Check if we want to skip execution of all compiled code.
3419 {
3420 #ifndef PRODUCT
3421 if (OptoNoExecute) {
3422 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3423 return;
3424 }
3425 #endif
3426 Compile::TracePhase tp(_t_registerMethod);
3427
3428 if (C->is_osr_compilation()) {
3429 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3430 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3431 } else {
3432 if (!target->is_static()) {
3433 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3434 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3435 // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3436 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3437 }
3438 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3439 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3440 }
3441
3442 C->env()->register_method(target,
3443 entry_bci,
3444 &_code_offsets,
3445 _orig_pc_slot_offset_in_bytes,
3446 code_buffer(),
3447 frame_size_in_words(),
3448 oop_map_set(),
3449 &_handler_table,
3450 inc_table(),
3451 compiler,
3452 has_unsafe_access,
3453 SharedRuntime::is_wide_vector(C->max_vector_size()),
3454 C->has_monitors(),
3455 C->has_scoped_access(),
3456 0);
3457
3458 if (C->log() != nullptr) { // Print code cache state into compiler log
3459 C->log()->code_cache_state();
3460 }
3461 }
3462 }
3463 void PhaseOutput::install_stub(const char* stub_name) {
3464 // Entry point will be accessed using stub_entry_point();
3465 if (code_buffer() == nullptr) {
3466 Matcher::soft_match_failure();
3467 } else {
3468 if (PrintAssembly && (WizardMode || Verbose))
3469 tty->print_cr("### Stub::%s", stub_name);
3470
3471 if (!C->failing()) {
3472 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3473
3474 // Make the NMethod
3475 // For now we mark the frame as never safe for profile stackwalking
3476 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3477 code_buffer(),
3478 CodeOffsets::frame_never_safe,
3479 // _code_offsets.value(CodeOffsets::Frame_Complete),
3480 frame_size_in_words(),
3481 oop_map_set(),
3482 false);
3483 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3484
3485 C->set_stub_entry_point(rs->entry_point());
3486 }
3487 }
3488 }
3489
3490 // Support for bundling info
3491 Bundle* PhaseOutput::node_bundling(const Node *n) {
3492 assert(valid_bundle_info(n), "oob");
3493 return &_node_bundling_base[n->_idx];
3494 }
3495
3496 bool PhaseOutput::valid_bundle_info(const Node *n) {
3497 return (_node_bundling_limit > n->_idx);
3498 }
3499
3500 //------------------------------frame_size_in_words-----------------------------
3501 // frame_slots in units of words
3502 int PhaseOutput::frame_size_in_words() const {
3503 // shift is 0 in LP32 and 1 in LP64
3504 const int shift = (LogBytesPerWord - LogBytesPerInt);
3505 int words = _frame_slots >> shift;
3506 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3507 return words;
3508 }
3509
3510 // To bang the stack of this compiled method we use the stack size
3511 // that the interpreter would need in case of a deoptimization. This
3512 // removes the need to bang the stack in the deoptimization blob which
3513 // in turn simplifies stack overflow handling.
3514 int PhaseOutput::bang_size_in_bytes() const {
3515 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3516 }
3517
3518 //------------------------------dump_asm---------------------------------------
3519 // Dump formatted assembly
3520 #if defined(SUPPORT_OPTO_ASSEMBLY)
3521 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3522
3523 int pc_digits = 3; // #chars required for pc
3524 int sb_chars = 3; // #chars for "start bundle" indicator
3525 int tab_size = 8;
3526 if (pcs != nullptr) {
3527 int max_pc = 0;
3528 for (uint i = 0; i < pc_limit; i++) {
3529 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3530 }
3531 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3532 }
3533 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3534
3535 bool cut_short = false;
3536 st->print_cr("#");
3537 st->print("# "); C->tf()->dump_on(st); st->cr();
3538 st->print_cr("#");
3539
3540 // For all blocks
3541 int pc = 0x0; // Program counter
3542 char starts_bundle = ' ';
3543 C->regalloc()->dump_frame();
3544
3545 Node *n = nullptr;
3546 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3547 if (VMThread::should_terminate()) {
3548 cut_short = true;
3549 break;
3550 }
3551 Block* block = C->cfg()->get_block(i);
3552 if (block->is_connector() && !Verbose) {
3553 continue;
3554 }
3555 n = block->head();
3556 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3557 pc = pcs[n->_idx];
3558 st->print("%*.*x", pc_digits, pc_digits, pc);
3559 }
3560 st->fill_to(prefix_len);
3561 block->dump_head(C->cfg(), st);
3562 if (block->is_connector()) {
3563 st->fill_to(prefix_len);
3564 st->print_cr("# Empty connector block");
3565 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3566 st->fill_to(prefix_len);
3567 st->print_cr("# Block is sole successor of call");
3568 }
3569
3570 // For all instructions
3571 Node *delay = nullptr;
3572 for (uint j = 0; j < block->number_of_nodes(); j++) {
3573 if (VMThread::should_terminate()) {
3574 cut_short = true;
3575 break;
3576 }
3577 n = block->get_node(j);
3578 if (valid_bundle_info(n)) {
3579 Bundle* bundle = node_bundling(n);
3580 if (bundle->used_in_unconditional_delay()) {
3581 delay = n;
3582 continue;
3583 }
3584 if (bundle->starts_bundle()) {
3585 starts_bundle = '+';
3586 }
3587 }
3588
3589 if (WizardMode) {
3590 n->dump();
3591 }
3592
3593 if( !n->is_Region() && // Dont print in the Assembly
3594 !n->is_Phi() && // a few noisely useless nodes
3595 !n->is_Proj() &&
3596 !n->is_MachTemp() &&
3597 !n->is_SafePointScalarObject() &&
3598 !n->is_Catch() && // Would be nice to print exception table targets
3599 !n->is_MergeMem() && // Not very interesting
3600 !n->is_top() && // Debug info table constants
3601 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3602 ) {
3603 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3604 pc = pcs[n->_idx];
3605 st->print("%*.*x", pc_digits, pc_digits, pc);
3606 } else {
3607 st->fill_to(pc_digits);
3608 }
3609 st->print(" %c ", starts_bundle);
3610 starts_bundle = ' ';
3611 st->fill_to(prefix_len);
3612 n->format(C->regalloc(), st);
3613 st->cr();
3614 }
3615
3616 // If we have an instruction with a delay slot, and have seen a delay,
3617 // then back up and print it
3618 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3619 // Coverity finding - Explicit null dereferenced.
3620 guarantee(delay != nullptr, "no unconditional delay instruction");
3621 if (WizardMode) delay->dump();
3622
3623 if (node_bundling(delay)->starts_bundle())
3624 starts_bundle = '+';
3625 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3626 pc = pcs[n->_idx];
3627 st->print("%*.*x", pc_digits, pc_digits, pc);
3628 } else {
3629 st->fill_to(pc_digits);
3630 }
3631 st->print(" %c ", starts_bundle);
3632 starts_bundle = ' ';
3633 st->fill_to(prefix_len);
3634 delay->format(C->regalloc(), st);
3635 st->cr();
3636 delay = nullptr;
3637 }
3638
3639 // Dump the exception table as well
3640 if( n->is_Catch() && (Verbose || WizardMode) ) {
3641 // Print the exception table for this offset
3642 _handler_table.print_subtable_for(pc);
3643 }
3644 st->bol(); // Make sure we start on a new line
3645 }
3646 st->cr(); // one empty line between blocks
3647 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3648 } // End of per-block dump
3649
3650 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3651 }
3652 #endif
3653
3654 #ifndef PRODUCT
3655 void PhaseOutput::print_statistics() {
3656 Scheduling::print_statistics();
3657 }
3658 #endif