1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/assembler.inline.hpp"
26 #include "code/aotCodeCache.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compilerDirectives.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "gc/shared/barrierSet.hpp"
35 #include "gc/shared/c2/barrierSetC2.hpp"
36 #include "memory/allocation.hpp"
37 #include "opto/ad.hpp"
38 #include "opto/block.hpp"
39 #include "opto/c2_MacroAssembler.hpp"
40 #include "opto/c2compiler.hpp"
41 #include "opto/callnode.hpp"
42 #include "opto/cfgnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/machnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/optoreg.hpp"
47 #include "opto/output.hpp"
48 #include "opto/regalloc.hpp"
49 #include "opto/type.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "utilities/macros.hpp"
52 #include "utilities/powerOfTwo.hpp"
53 #include "utilities/xmlstream.hpp"
54
55 #ifndef PRODUCT
56 #define DEBUG_ARG(x) , x
57 #else
58 #define DEBUG_ARG(x)
59 #endif
60
61 //------------------------------Scheduling----------------------------------
62 // This class contains all the information necessary to implement instruction
63 // scheduling and bundling.
64 class Scheduling {
65
66 private:
67 // Arena to use
68 Arena *_arena;
69
70 // Control-Flow Graph info
71 PhaseCFG *_cfg;
72
73 // Register Allocation info
74 PhaseRegAlloc *_regalloc;
75
76 // Number of nodes in the method
77 uint _node_bundling_limit;
78
79 // List of scheduled nodes. Generated in reverse order
80 Node_List _scheduled;
81
82 // List of nodes currently available for choosing for scheduling
83 Node_List _available;
84
85 // For each instruction beginning a bundle, the number of following
86 // nodes to be bundled with it.
87 Bundle *_node_bundling_base;
88
89 // Mapping from register to Node
90 Node_List _reg_node;
91
92 // Free list for pinch nodes.
93 Node_List _pinch_free_list;
94
95 // Number of uses of this node within the containing basic block.
96 short *_uses;
97
98 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
99 // front, branch+proj at end. Also skips Catch/CProj (same as
100 // branch-at-end), plus just-prior exception-throwing call.
101 uint _bb_start, _bb_end;
102
103 // Latency from the end of the basic block as scheduled
104 unsigned short *_current_latency;
105
106 // Remember the next node
107 Node *_next_node;
108
109 // Length of the current bundle, in instructions
110 uint _bundle_instr_count;
111
112 // Current Cycle number, for computing latencies and bundling
113 uint _bundle_cycle_number;
114
115 // Bundle information
116 Pipeline_Use_Element _bundle_use_elements[resource_count];
117 Pipeline_Use _bundle_use;
118
119 // Dump the available list
120 void dump_available() const;
121
122 public:
123 Scheduling(Arena *arena, Compile &compile);
124
125 // Step ahead "i" cycles
126 void step(uint i);
127
128 // Step ahead 1 cycle, and clear the bundle state (for example,
129 // at a branch target)
130 void step_and_clear();
131
132 Bundle* node_bundling(const Node *n) {
133 assert(valid_bundle_info(n), "oob");
134 return (&_node_bundling_base[n->_idx]);
135 }
136
137 bool valid_bundle_info(const Node *n) const {
138 return (_node_bundling_limit > n->_idx);
139 }
140
141 bool starts_bundle(const Node *n) const {
142 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
143 }
144
145 // Do the scheduling
146 void DoScheduling();
147
148 // Compute the register antidependencies within a basic block
149 void ComputeRegisterAntidependencies(Block *bb);
150 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
151 void verify_good_schedule( Block *b, const char *msg );
152 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
153 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
154
155 // Add a node to the current bundle
156 void AddNodeToBundle(Node *n, const Block *bb);
157
158 // Return an integer less than, equal to, or greater than zero
159 // if the stack offset of the first argument is respectively
160 // less than, equal to, or greater than the second.
161 int compare_two_spill_nodes(Node* first, Node* second);
162
163 // Add a node to the list of available nodes
164 void AddNodeToAvailableList(Node *n);
165
166 // Compute the local use count for the nodes in a block, and compute
167 // the list of instructions with no uses in the block as available
168 void ComputeUseCount(const Block *bb);
169
170 // Choose an instruction from the available list to add to the bundle
171 Node * ChooseNodeToBundle();
172
173 // See if this Node fits into the currently accumulating bundle
174 bool NodeFitsInBundle(Node *n);
175
176 // Decrement the use count for a node
177 void DecrementUseCounts(Node *n, const Block *bb);
178
179 // Garbage collect pinch nodes for reuse by other blocks.
180 void garbage_collect_pinch_nodes();
181 // Clean up a pinch node for reuse (helper for above).
182 void cleanup_pinch( Node *pinch );
183
184 // Information for statistics gathering
185 #ifndef PRODUCT
186 private:
187 // Gather information on size of nops relative to total
188 static uint _total_nop_size, _total_method_size;
189 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
190
191 public:
192 static void print_statistics();
193
194 static void increment_instructions_per_bundle(uint i) {
195 _total_instructions_per_bundle[i]++;
196 }
197
198 static void increment_nop_size(uint s) {
199 _total_nop_size += s;
200 }
201
202 static void increment_method_size(uint s) {
203 _total_method_size += s;
204 }
205 #endif
206
207 };
208
209 PhaseOutput::PhaseOutput()
210 : Phase(Phase::Output),
211 _code_buffer("Compile::Fill_buffer"),
212 _first_block_size(0),
213 _handler_table(),
214 _inc_table(),
215 _stub_list(),
216 _oop_map_set(nullptr),
217 _scratch_buffer_blob(nullptr),
218 _scratch_locs_memory(nullptr),
219 _scratch_const_size(-1),
220 _in_scratch_emit_size(false),
221 _frame_slots(0),
222 _code_offsets(),
223 _node_bundling_limit(0),
224 _node_bundling_base(nullptr),
225 _orig_pc_slot(0),
226 _orig_pc_slot_offset_in_bytes(0),
227 _buf_sizes(),
228 _block(nullptr),
229 _index(0) {
230 C->set_output(this);
231 if (C->stub_name() == nullptr) {
232 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
233 }
234 }
235
236 PhaseOutput::~PhaseOutput() {
237 C->set_output(nullptr);
238 if (_scratch_buffer_blob != nullptr) {
239 BufferBlob::free(_scratch_buffer_blob);
240 }
241 }
242
243 void PhaseOutput::perform_mach_node_analysis() {
244 // Late barrier analysis must be done after schedule and bundle
245 // Otherwise liveness based spilling will fail
246 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
247 bs->late_barrier_analysis();
248
249 pd_perform_mach_node_analysis();
250
251 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
252 }
253
254 // Convert Nodes to instruction bits and pass off to the VM
255 void PhaseOutput::Output() {
256 // RootNode goes
257 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
258
259 // The number of new nodes (mostly MachNop) is proportional to
260 // the number of java calls and inner loops which are aligned.
261 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
262 C->inner_loops()*(OptoLoopAlignment-1)),
263 "out of nodes before code generation" ) ) {
264 return;
265 }
266 // Make sure I can find the Start Node
267 Block *entry = C->cfg()->get_block(1);
268 Block *broot = C->cfg()->get_root_block();
269
270 const StartNode *start = entry->head()->as_Start();
271
272 // Replace StartNode with prolog
273 MachPrologNode *prolog = new MachPrologNode();
274 entry->map_node(prolog, 0);
275 C->cfg()->map_node_to_block(prolog, entry);
276 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
277
278 // Virtual methods need an unverified entry point
279
280 if( C->is_osr_compilation() ) {
281 if( PoisonOSREntry ) {
282 // TODO: Should use a ShouldNotReachHereNode...
283 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
284 }
285 } else {
286 if( C->method() && !C->method()->flags().is_static() ) {
287 // Insert unvalidated entry point
288 C->cfg()->insert( broot, 0, new MachUEPNode() );
289 }
290
291 }
292
293 // Break before main entry point
294 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
295 (OptoBreakpoint && C->is_method_compilation()) ||
296 (OptoBreakpointOSR && C->is_osr_compilation()) ||
297 (OptoBreakpointC2R && !C->method()) ) {
298 // checking for C->method() means that OptoBreakpoint does not apply to
299 // runtime stubs or frame converters
300 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
301 }
302
303 // Insert epilogs before every return
304 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
305 Block* block = C->cfg()->get_block(i);
306 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
307 Node* m = block->end();
308 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
309 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
310 block->add_inst(epilog);
311 C->cfg()->map_node_to_block(epilog, block);
312 }
313 }
314 }
315
316 // Keeper of sizing aspects
317 _buf_sizes = BufferSizingData();
318
319 // Initialize code buffer
320 estimate_buffer_size(_buf_sizes._const);
321 if (C->failing()) return;
322
323 // Pre-compute the length of blocks and replace
324 // long branches with short if machine supports it.
325 // Must be done before ScheduleAndBundle due to SPARC delay slots
326 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
327 blk_starts[0] = 0;
328 shorten_branches(blk_starts);
329
330 ScheduleAndBundle();
331 if (C->failing()) {
332 return;
333 }
334
335 perform_mach_node_analysis();
336
337 // Complete sizing of codebuffer
338 CodeBuffer* cb = init_buffer();
339 if (cb == nullptr || C->failing()) {
340 return;
341 }
342
343 BuildOopMaps();
344
345 if (C->failing()) {
346 return;
347 }
348
349 C2_MacroAssembler masm(cb);
350 fill_buffer(&masm, blk_starts);
351 }
352
353 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
354 // Determine if we need to generate a stack overflow check.
355 // Do it if the method is not a stub function and
356 // has java calls or has frame size > vm_page_size/8.
357 // The debug VM checks that deoptimization doesn't trigger an
358 // unexpected stack overflow (compiled method stack banging should
359 // guarantee it doesn't happen) so we always need the stack bang in
360 // a debug VM.
361 return (C->stub_function() == nullptr &&
362 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
363 DEBUG_ONLY(|| true)));
364 }
365
366 bool PhaseOutput::need_register_stack_bang() const {
367 // Determine if we need to generate a register stack overflow check.
368 // This is only used on architectures which have split register
369 // and memory stacks.
370 // Bang if the method is not a stub function and has java calls
371 return (C->stub_function() == nullptr && C->has_java_calls());
372 }
373
374
375 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
376 // of a loop. When aligning a loop we need to provide enough instructions
377 // in cpu's fetch buffer to feed decoders. The loop alignment could be
378 // avoided if we have enough instructions in fetch buffer at the head of a loop.
379 // By default, the size is set to 999999 by Block's constructor so that
380 // a loop will be aligned if the size is not reset here.
381 //
382 // Note: Mach instructions could contain several HW instructions
383 // so the size is estimated only.
384 //
385 void PhaseOutput::compute_loop_first_inst_sizes() {
386 // The next condition is used to gate the loop alignment optimization.
387 // Don't aligned a loop if there are enough instructions at the head of a loop
388 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
389 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
390 // equal to 11 bytes which is the largest address NOP instruction.
391 if (MaxLoopPad < OptoLoopAlignment - 1) {
392 uint last_block = C->cfg()->number_of_blocks() - 1;
393 for (uint i = 1; i <= last_block; i++) {
394 Block* block = C->cfg()->get_block(i);
395 // Check the first loop's block which requires an alignment.
396 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
397 uint sum_size = 0;
398 uint inst_cnt = NumberOfLoopInstrToAlign;
399 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
400
401 // Check subsequent fallthrough blocks if the loop's first
402 // block(s) does not have enough instructions.
403 Block *nb = block;
404 while(inst_cnt > 0 &&
405 i < last_block &&
406 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
407 !nb->has_successor(block)) {
408 i++;
409 nb = C->cfg()->get_block(i);
410 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
411 } // while( inst_cnt > 0 && i < last_block )
412
413 block->set_first_inst_size(sum_size);
414 } // f( b->head()->is_Loop() )
415 } // for( i <= last_block )
416 } // if( MaxLoopPad < OptoLoopAlignment-1 )
417 }
418
419 // The architecture description provides short branch variants for some long
420 // branch instructions. Replace eligible long branches with short branches.
421 void PhaseOutput::shorten_branches(uint* blk_starts) {
422
423 Compile::TracePhase tp(_t_shortenBranches);
424
425 // Compute size of each block, method size, and relocation information size
426 uint nblocks = C->cfg()->number_of_blocks();
427
428 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
429 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
430 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
431
432 // Collect worst case block paddings
433 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
434 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
435
436 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
437 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
438
439 bool has_short_branch_candidate = false;
440
441 // Initialize the sizes to 0
442 int code_size = 0; // Size in bytes of generated code
443 int stub_size = 0; // Size in bytes of all stub entries
444 // Size in bytes of all relocation entries, including those in local stubs.
445 // Start with 2-bytes of reloc info for the unvalidated entry point
446 int reloc_size = 1; // Number of relocation entries
447
448 // Make three passes. The first computes pessimistic blk_starts,
449 // relative jmp_offset and reloc_size information. The second performs
450 // short branch substitution using the pessimistic sizing. The
451 // third inserts nops where needed.
452
453 // Step one, perform a pessimistic sizing pass.
454 uint last_call_adr = max_juint;
455 uint last_avoid_back_to_back_adr = max_juint;
456 uint nop_size = (new MachNopNode())->size(C->regalloc());
457 for (uint i = 0; i < nblocks; i++) { // For all blocks
458 Block* block = C->cfg()->get_block(i);
459 _block = block;
460
461 // During short branch replacement, we store the relative (to blk_starts)
462 // offset of jump in jmp_offset, rather than the absolute offset of jump.
463 // This is so that we do not need to recompute sizes of all nodes when
464 // we compute correct blk_starts in our next sizing pass.
465 jmp_offset[i] = 0;
466 jmp_size[i] = 0;
467 jmp_nidx[i] = -1;
468 DEBUG_ONLY( jmp_target[i] = 0; )
469 DEBUG_ONLY( jmp_rule[i] = 0; )
470
471 // Sum all instruction sizes to compute block size
472 uint last_inst = block->number_of_nodes();
473 uint blk_size = 0;
474 for (uint j = 0; j < last_inst; j++) {
475 _index = j;
476 Node* nj = block->get_node(_index);
477 // Handle machine instruction nodes
478 if (nj->is_Mach()) {
479 MachNode* mach = nj->as_Mach();
480 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
481 reloc_size += mach->reloc();
482 if (mach->is_MachCall()) {
483 // add size information for trampoline stub
484 // class CallStubImpl is platform-specific and defined in the *.ad files.
485 stub_size += CallStubImpl::size_call_trampoline();
486 reloc_size += CallStubImpl::reloc_call_trampoline();
487
488 MachCallNode *mcall = mach->as_MachCall();
489 // This destination address is NOT PC-relative
490
491 mcall->method_set((intptr_t)mcall->entry_point());
492
493 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
494 stub_size += CompiledDirectCall::to_interp_stub_size();
495 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
496 }
497 } else if (mach->is_MachSafePoint()) {
498 // If call/safepoint are adjacent, account for possible
499 // nop to disambiguate the two safepoints.
500 // ScheduleAndBundle() can rearrange nodes in a block,
501 // check for all offsets inside this block.
502 if (last_call_adr >= blk_starts[i]) {
503 blk_size += nop_size;
504 }
505 }
506 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
507 // Nop is inserted between "avoid back to back" instructions.
508 // ScheduleAndBundle() can rearrange nodes in a block,
509 // check for all offsets inside this block.
510 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
511 blk_size += nop_size;
512 }
513 }
514 if (mach->may_be_short_branch()) {
515 if (!nj->is_MachBranch()) {
516 #ifndef PRODUCT
517 nj->dump(3);
518 #endif
519 Unimplemented();
520 }
521 assert(jmp_nidx[i] == -1, "block should have only one branch");
522 jmp_offset[i] = blk_size;
523 jmp_size[i] = nj->size(C->regalloc());
524 jmp_nidx[i] = j;
525 has_short_branch_candidate = true;
526 }
527 }
528 blk_size += nj->size(C->regalloc());
529 // Remember end of call offset
530 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
531 last_call_adr = blk_starts[i]+blk_size;
532 }
533 // Remember end of avoid_back_to_back offset
534 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
535 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
536 }
537 }
538
539 // When the next block starts a loop, we may insert pad NOP
540 // instructions. Since we cannot know our future alignment,
541 // assume the worst.
542 if (i < nblocks - 1) {
543 Block* nb = C->cfg()->get_block(i + 1);
544 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
545 if (max_loop_pad > 0) {
546 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
547 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
548 // If either is the last instruction in this block, bump by
549 // max_loop_pad in lock-step with blk_size, so sizing
550 // calculations in subsequent blocks still can conservatively
551 // detect that it may the last instruction in this block.
552 if (last_call_adr == blk_starts[i]+blk_size) {
553 last_call_adr += max_loop_pad;
554 }
555 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
556 last_avoid_back_to_back_adr += max_loop_pad;
557 }
558 blk_size += max_loop_pad;
559 block_worst_case_pad[i + 1] = max_loop_pad;
560 }
561 }
562
563 // Save block size; update total method size
564 blk_starts[i+1] = blk_starts[i]+blk_size;
565 }
566
567 // Step two, replace eligible long jumps.
568 bool progress = true;
569 uint last_may_be_short_branch_adr = max_juint;
570 while (has_short_branch_candidate && progress) {
571 progress = false;
572 has_short_branch_candidate = false;
573 int adjust_block_start = 0;
574 for (uint i = 0; i < nblocks; i++) {
575 Block* block = C->cfg()->get_block(i);
576 int idx = jmp_nidx[i];
577 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
578 if (mach != nullptr && mach->may_be_short_branch()) {
579 #ifdef ASSERT
580 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
581 int j;
582 // Find the branch; ignore trailing NOPs.
583 for (j = block->number_of_nodes()-1; j>=0; j--) {
584 Node* n = block->get_node(j);
585 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
586 break;
587 }
588 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
589 #endif
590 int br_size = jmp_size[i];
591 int br_offs = blk_starts[i] + jmp_offset[i];
592
593 // This requires the TRUE branch target be in succs[0]
594 uint bnum = block->non_connector_successor(0)->_pre_order;
595 int offset = blk_starts[bnum] - br_offs;
596 if (bnum > i) { // adjust following block's offset
597 offset -= adjust_block_start;
598 }
599
600 // This block can be a loop header, account for the padding
601 // in the previous block.
602 int block_padding = block_worst_case_pad[i];
603 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
604 // In the following code a nop could be inserted before
605 // the branch which will increase the backward distance.
606 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
607 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
608
609 if (needs_padding && offset <= 0)
610 offset -= nop_size;
611
612 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
613 // We've got a winner. Replace this branch.
614 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
615
616 // Update the jmp_size.
617 int new_size = replacement->size(C->regalloc());
618 int diff = br_size - new_size;
619 assert(diff >= (int)nop_size, "short_branch size should be smaller");
620 // Conservatively take into account padding between
621 // avoid_back_to_back branches. Previous branch could be
622 // converted into avoid_back_to_back branch during next
623 // rounds.
624 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
625 jmp_offset[i] += nop_size;
626 diff -= nop_size;
627 }
628 adjust_block_start += diff;
629 block->map_node(replacement, idx);
630 mach->subsume_by(replacement, C);
631 mach = replacement;
632 progress = true;
633
634 jmp_size[i] = new_size;
635 DEBUG_ONLY( jmp_target[i] = bnum; );
636 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
637 } else {
638 // The jump distance is not short, try again during next iteration.
639 has_short_branch_candidate = true;
640 }
641 } // (mach->may_be_short_branch())
642 if (mach != nullptr && (mach->may_be_short_branch() ||
643 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
644 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
645 }
646 blk_starts[i+1] -= adjust_block_start;
647 }
648 }
649
650 #ifdef ASSERT
651 for (uint i = 0; i < nblocks; i++) { // For all blocks
652 if (jmp_target[i] != 0) {
653 int br_size = jmp_size[i];
654 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
655 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
656 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]);
657 }
658 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
659 }
660 }
661 #endif
662
663 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
664 // after ScheduleAndBundle().
665
666 // ------------------
667 // Compute size for code buffer
668 code_size = blk_starts[nblocks];
669
670 // Relocation records
671 reloc_size += 1; // Relo entry for exception handler
672
673 // Adjust reloc_size to number of record of relocation info
674 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
675 // a relocation index.
676 // The CodeBuffer will expand the locs array if this estimate is too low.
677 reloc_size *= 10 / sizeof(relocInfo);
678
679 _buf_sizes._reloc = reloc_size;
680 _buf_sizes._code = code_size;
681 _buf_sizes._stub = stub_size;
682 }
683
684 //------------------------------FillLocArray-----------------------------------
685 // Create a bit of debug info and append it to the array. The mapping is from
686 // Java local or expression stack to constant, register or stack-slot. For
687 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
688 // entry has been taken care of and caller should skip it).
689 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
690 // This should never have accepted Bad before
691 assert(OptoReg::is_valid(regnum), "location must be valid");
692 return (OptoReg::is_reg(regnum))
693 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
694 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
695 }
696
697
698 ObjectValue*
699 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
700 for (int i = 0; i < objs->length(); i++) {
701 assert(objs->at(i)->is_object(), "corrupt object cache");
702 ObjectValue* sv = objs->at(i)->as_ObjectValue();
703 if (sv->id() == id) {
704 return sv;
705 }
706 }
707 // Otherwise..
708 return nullptr;
709 }
710
711 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
712 ObjectValue* sv ) {
713 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
714 objs->append(sv);
715 }
716
717
718 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
719 GrowableArray<ScopeValue*> *array,
720 GrowableArray<ScopeValue*> *objs ) {
721 assert( local, "use _top instead of null" );
722 if (array->length() != idx) {
723 assert(array->length() == idx + 1, "Unexpected array count");
724 // Old functionality:
725 // return
726 // New functionality:
727 // Assert if the local is not top. In product mode let the new node
728 // override the old entry.
729 assert(local == C->top(), "LocArray collision");
730 if (local == C->top()) {
731 return;
732 }
733 array->pop();
734 }
735 const Type *t = local->bottom_type();
736
737 // Is it a safepoint scalar object node?
738 if (local->is_SafePointScalarObject()) {
739 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
740
741 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
742 if (sv == nullptr) {
743 ciKlass* cik = t->is_oopptr()->exact_klass();
744 assert(cik->is_instance_klass() ||
745 cik->is_array_klass(), "Not supported allocation.");
746 sv = new ObjectValue(spobj->_idx,
747 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
748 set_sv_for_object_node(objs, sv);
749
750 uint first_ind = spobj->first_index(sfpt->jvms());
751 for (uint i = 0; i < spobj->n_fields(); i++) {
752 Node* fld_node = sfpt->in(first_ind+i);
753 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
754 }
755 }
756 array->append(sv);
757 return;
758 } else if (local->is_SafePointScalarMerge()) {
759 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
760 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
761
762 if (mv == nullptr) {
763 GrowableArray<ScopeValue*> deps;
764
765 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
766 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
767 assert(deps.length() == 1, "missing value");
768
769 int selector_idx = smerge->selector_idx(sfpt->jvms());
770 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
771 assert(deps.length() == 2, "missing value");
772
773 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
774 set_sv_for_object_node(objs, mv);
775
776 for (uint i = 1; i < smerge->req(); i++) {
777 Node* obj_node = smerge->in(i);
778 int idx = mv->possible_objects()->length();
779 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
780
781 // By default ObjectValues that are in 'possible_objects' are not root objects.
782 // They will be marked as root later if they are directly referenced in a JVMS.
783 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
784 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
785 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
786 }
787 }
788 array->append(mv);
789 return;
790 }
791
792 // Grab the register number for the local
793 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
794 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
795 // Record the double as two float registers.
796 // The register mask for such a value always specifies two adjacent
797 // float registers, with the lower register number even.
798 // Normally, the allocation of high and low words to these registers
799 // is irrelevant, because nearly all operations on register pairs
800 // (e.g., StoreD) treat them as a single unit.
801 // Here, we assume in addition that the words in these two registers
802 // stored "naturally" (by operations like StoreD and double stores
803 // within the interpreter) such that the lower-numbered register
804 // is written to the lower memory address. This may seem like
805 // a machine dependency, but it is not--it is a requirement on
806 // the author of the <arch>.ad file to ensure that, for every
807 // even/odd double-register pair to which a double may be allocated,
808 // the word in the even single-register is stored to the first
809 // memory word. (Note that register numbers are completely
810 // arbitrary, and are not tied to any machine-level encodings.)
811 #ifdef _LP64
812 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
813 array->append(new ConstantIntValue((jint)0));
814 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
815 } else if ( t->base() == Type::Long ) {
816 array->append(new ConstantIntValue((jint)0));
817 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
818 } else if ( t->base() == Type::RawPtr ) {
819 // jsr/ret return address which must be restored into the full
820 // width 64-bit stack slot.
821 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
822 }
823 #else //_LP64
824 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
825 // Repack the double/long as two jints.
826 // The convention the interpreter uses is that the second local
827 // holds the first raw word of the native double representation.
828 // This is actually reasonable, since locals and stack arrays
829 // grow downwards in all implementations.
830 // (If, on some machine, the interpreter's Java locals or stack
831 // were to grow upwards, the embedded doubles would be word-swapped.)
832 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
833 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
834 }
835 #endif //_LP64
836 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
837 OptoReg::is_reg(regnum) ) {
838 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
839 ? Location::float_in_dbl : Location::normal ));
840 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
841 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
842 ? Location::int_in_long : Location::normal ));
843 } else if( t->base() == Type::NarrowOop ) {
844 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
845 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
846 t->base() == Type::VectorD || t->base() == Type::VectorX ||
847 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
848 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
849 } else if (C->regalloc()->is_oop(local)) {
850 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
851 t->base() == Type::AryPtr,
852 "Unexpected type: %s", t->msg());
853 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
854 } else {
855 assert(t->base() == Type::Int || t->base() == Type::Half ||
856 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
857 "Unexpected type: %s", t->msg());
858 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
859 }
860 return;
861 }
862
863 // No register. It must be constant data.
864 switch (t->base()) {
865 case Type::Half: // Second half of a double
866 ShouldNotReachHere(); // Caller should skip 2nd halves
867 break;
868 case Type::AnyPtr:
869 array->append(new ConstantOopWriteValue(nullptr));
870 break;
871 case Type::AryPtr:
872 case Type::InstPtr: // fall through
873 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
874 break;
875 case Type::NarrowOop:
876 if (t == TypeNarrowOop::NULL_PTR) {
877 array->append(new ConstantOopWriteValue(nullptr));
878 } else {
879 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
880 }
881 break;
882 case Type::Int:
883 array->append(new ConstantIntValue(t->is_int()->get_con()));
884 break;
885 case Type::RawPtr:
886 // A return address (T_ADDRESS).
887 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
888 #ifdef _LP64
889 // Must be restored to the full-width 64-bit stack slot.
890 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
891 #else
892 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
893 #endif
894 break;
895 case Type::FloatCon: {
896 float f = t->is_float_constant()->getf();
897 array->append(new ConstantIntValue(jint_cast(f)));
898 break;
899 }
900 case Type::DoubleCon: {
901 jdouble d = t->is_double_constant()->getd();
902 #ifdef _LP64
903 array->append(new ConstantIntValue((jint)0));
904 array->append(new ConstantDoubleValue(d));
905 #else
906 // Repack the double as two jints.
907 // The convention the interpreter uses is that the second local
908 // holds the first raw word of the native double representation.
909 // This is actually reasonable, since locals and stack arrays
910 // grow downwards in all implementations.
911 // (If, on some machine, the interpreter's Java locals or stack
912 // were to grow upwards, the embedded doubles would be word-swapped.)
913 jlong_accessor acc;
914 acc.long_value = jlong_cast(d);
915 array->append(new ConstantIntValue(acc.words[1]));
916 array->append(new ConstantIntValue(acc.words[0]));
917 #endif
918 break;
919 }
920 case Type::Long: {
921 jlong d = t->is_long()->get_con();
922 #ifdef _LP64
923 array->append(new ConstantIntValue((jint)0));
924 array->append(new ConstantLongValue(d));
925 #else
926 // Repack the long as two jints.
927 // The convention the interpreter uses is that the second local
928 // holds the first raw word of the native double representation.
929 // This is actually reasonable, since locals and stack arrays
930 // grow downwards in all implementations.
931 // (If, on some machine, the interpreter's Java locals or stack
932 // were to grow upwards, the embedded doubles would be word-swapped.)
933 jlong_accessor acc;
934 acc.long_value = d;
935 array->append(new ConstantIntValue(acc.words[1]));
936 array->append(new ConstantIntValue(acc.words[0]));
937 #endif
938 break;
939 }
940 case Type::Top: // Add an illegal value here
941 array->append(new LocationValue(Location()));
942 break;
943 default:
944 ShouldNotReachHere();
945 break;
946 }
947 }
948
949 // Determine if this node starts a bundle
950 bool PhaseOutput::starts_bundle(const Node *n) const {
951 return (_node_bundling_limit > n->_idx &&
952 _node_bundling_base[n->_idx].starts_bundle());
953 }
954
955 // Determine if there is a monitor that has 'ov' as its owner.
956 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
957 for (int k = 0; k < monarray->length(); k++) {
958 MonitorValue* mv = monarray->at(k);
959 if (mv->owner() == ov) {
960 return true;
961 }
962 }
963
964 return false;
965 }
966
967 // Determine if there is a scalar replaced object description represented by 'ov'.
968 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
969 GrowableArray<ScopeValue*>* objs,
970 ObjectValue* ov) const {
971 for (int i = 0; i < jvms->scl_size(); i++) {
972 Node* n = sfn->scalarized_obj(jvms, i);
973 // Other kinds of nodes that we may encounter here, for instance constants
974 // representing values of fields of objects scalarized, aren't relevant for
975 // us, since they don't map to ObjectValue.
976 if (!n->is_SafePointScalarObject()) {
977 continue;
978 }
979
980 ObjectValue* other = sv_for_node_id(objs, n->_idx);
981 if (ov == other) {
982 return true;
983 }
984 }
985 return false;
986 }
987
988 //--------------------------Process_OopMap_Node--------------------------------
989 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
990 // Handle special safepoint nodes for synchronization
991 MachSafePointNode *sfn = mach->as_MachSafePoint();
992 MachCallNode *mcall;
993
994 int safepoint_pc_offset = current_offset;
995 bool is_method_handle_invoke = false;
996 bool return_oop = false;
997 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
998 bool arg_escape = false;
999
1000 // Add the safepoint in the DebugInfoRecorder
1001 if( !mach->is_MachCall() ) {
1002 mcall = nullptr;
1003 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1004 } else {
1005 mcall = mach->as_MachCall();
1006
1007 // Is the call a MethodHandle call?
1008 if (mcall->is_MachCallJava()) {
1009 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1010 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1011 is_method_handle_invoke = true;
1012 }
1013 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1014 }
1015
1016 // Check if a call returns an object.
1017 if (mcall->returns_pointer()) {
1018 return_oop = true;
1019 }
1020 safepoint_pc_offset += mcall->ret_addr_offset();
1021 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1022 }
1023
1024 // Loop over the JVMState list to add scope information
1025 // Do not skip safepoints with a null method, they need monitor info
1026 JVMState* youngest_jvms = sfn->jvms();
1027 int max_depth = youngest_jvms->depth();
1028
1029 // Allocate the object pool for scalar-replaced objects -- the map from
1030 // small-integer keys (which can be recorded in the local and ostack
1031 // arrays) to descriptions of the object state.
1032 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1033
1034 // Visit scopes from oldest to youngest.
1035 for (int depth = 1; depth <= max_depth; depth++) {
1036 JVMState* jvms = youngest_jvms->of_depth(depth);
1037 int idx;
1038 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1039 // Safepoints that do not have method() set only provide oop-map and monitor info
1040 // to support GC; these do not support deoptimization.
1041 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1042 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1043 int num_mon = jvms->nof_monitors();
1044 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1045 "JVMS local count must match that of the method");
1046
1047 // Add Local and Expression Stack Information
1048
1049 // Insert locals into the locarray
1050 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1051 for( idx = 0; idx < num_locs; idx++ ) {
1052 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1053 }
1054
1055 // Insert expression stack entries into the exparray
1056 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1057 for( idx = 0; idx < num_exps; idx++ ) {
1058 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1059 }
1060
1061 // Add in mappings of the monitors
1062 assert( !method ||
1063 !method->is_synchronized() ||
1064 method->is_native() ||
1065 num_mon > 0,
1066 "monitors must always exist for synchronized methods");
1067
1068 // Build the growable array of ScopeValues for exp stack
1069 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1070
1071 // Loop over monitors and insert into array
1072 for (idx = 0; idx < num_mon; idx++) {
1073 // Grab the node that defines this monitor
1074 Node* box_node = sfn->monitor_box(jvms, idx);
1075 Node* obj_node = sfn->monitor_obj(jvms, idx);
1076
1077 // Create ScopeValue for object
1078 ScopeValue *scval = nullptr;
1079
1080 if (obj_node->is_SafePointScalarObject()) {
1081 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1082 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1083 if (scval == nullptr) {
1084 const Type *t = spobj->bottom_type();
1085 ciKlass* cik = t->is_oopptr()->exact_klass();
1086 assert(cik->is_instance_klass() ||
1087 cik->is_array_klass(), "Not supported allocation.");
1088 ObjectValue* sv = new ObjectValue(spobj->_idx,
1089 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1090 PhaseOutput::set_sv_for_object_node(objs, sv);
1091
1092 uint first_ind = spobj->first_index(youngest_jvms);
1093 for (uint i = 0; i < spobj->n_fields(); i++) {
1094 Node* fld_node = sfn->in(first_ind+i);
1095 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1096 }
1097 scval = sv;
1098 }
1099 } else if (obj_node->is_SafePointScalarMerge()) {
1100 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1101 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1102
1103 if (mv == nullptr) {
1104 GrowableArray<ScopeValue*> deps;
1105
1106 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1107 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1108 assert(deps.length() == 1, "missing value");
1109
1110 int selector_idx = smerge->selector_idx(youngest_jvms);
1111 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1112 assert(deps.length() == 2, "missing value");
1113
1114 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1115 set_sv_for_object_node(objs, mv);
1116
1117 for (uint i = 1; i < smerge->req(); i++) {
1118 Node* obj_node = smerge->in(i);
1119 int idx = mv->possible_objects()->length();
1120 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1121
1122 // By default ObjectValues that are in 'possible_objects' are not root objects.
1123 // They will be marked as root later if they are directly referenced in a JVMS.
1124 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1125 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1126 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1127 }
1128 }
1129 scval = mv;
1130 } else if (!obj_node->is_Con()) {
1131 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1132 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1133 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1134 } else {
1135 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1136 }
1137 } else {
1138 const TypePtr *tp = obj_node->get_ptr_type();
1139 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1140 }
1141
1142 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1143 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1144 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1145 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1146 }
1147
1148 // Mark ObjectValue nodes as root nodes if they are directly
1149 // referenced in the JVMS.
1150 for (int i = 0; i < objs->length(); i++) {
1151 ScopeValue* sv = objs->at(i);
1152 if (sv->is_object_merge()) {
1153 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1154
1155 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1156 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1157 if (ov->is_root()) {
1158 // Already flagged as 'root' by something else. We shouldn't change it
1159 // to non-root in a younger JVMS because it may need to be alive in
1160 // a younger JVMS.
1161 } else {
1162 bool is_root = locarray->contains(ov) ||
1163 exparray->contains(ov) ||
1164 contains_as_owner(monarray, ov) ||
1165 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1166 ov->set_root(is_root);
1167 }
1168 }
1169 }
1170 }
1171
1172 // We dump the object pool first, since deoptimization reads it in first.
1173 C->debug_info()->dump_object_pool(objs);
1174
1175 // Build first class objects to pass to scope
1176 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1177 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1178 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1179
1180 // Make method available for all Safepoints
1181 ciMethod* scope_method = method ? method : C->method();
1182 // Describe the scope here
1183 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1184 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1185 // Now we can describe the scope.
1186 methodHandle null_mh;
1187 bool rethrow_exception = false;
1188 C->debug_info()->describe_scope(
1189 safepoint_pc_offset,
1190 null_mh,
1191 scope_method,
1192 jvms->bci(),
1193 jvms->should_reexecute(),
1194 rethrow_exception,
1195 is_method_handle_invoke,
1196 return_oop,
1197 has_ea_local_in_scope,
1198 arg_escape,
1199 locvals,
1200 expvals,
1201 monvals
1202 );
1203 } // End jvms loop
1204
1205 // Mark the end of the scope set.
1206 C->debug_info()->end_safepoint(safepoint_pc_offset);
1207 }
1208
1209
1210
1211 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1212 class NonSafepointEmitter {
1213 Compile* C;
1214 JVMState* _pending_jvms;
1215 int _pending_offset;
1216
1217 void emit_non_safepoint();
1218
1219 public:
1220 NonSafepointEmitter(Compile* compile) {
1221 this->C = compile;
1222 _pending_jvms = nullptr;
1223 _pending_offset = 0;
1224 }
1225
1226 void observe_instruction(Node* n, int pc_offset) {
1227 if (!C->debug_info()->recording_non_safepoints()) return;
1228
1229 Node_Notes* nn = C->node_notes_at(n->_idx);
1230 if (nn == nullptr || nn->jvms() == nullptr) return;
1231 if (_pending_jvms != nullptr &&
1232 _pending_jvms->same_calls_as(nn->jvms())) {
1233 // Repeated JVMS? Stretch it up here.
1234 _pending_offset = pc_offset;
1235 } else {
1236 if (_pending_jvms != nullptr &&
1237 _pending_offset < pc_offset) {
1238 emit_non_safepoint();
1239 }
1240 _pending_jvms = nullptr;
1241 if (pc_offset > C->debug_info()->last_pc_offset()) {
1242 // This is the only way _pending_jvms can become non-null:
1243 _pending_jvms = nn->jvms();
1244 _pending_offset = pc_offset;
1245 }
1246 }
1247 }
1248
1249 // Stay out of the way of real safepoints:
1250 void observe_safepoint(JVMState* jvms, int pc_offset) {
1251 if (_pending_jvms != nullptr &&
1252 !_pending_jvms->same_calls_as(jvms) &&
1253 _pending_offset < pc_offset) {
1254 emit_non_safepoint();
1255 }
1256 _pending_jvms = nullptr;
1257 }
1258
1259 void flush_at_end() {
1260 if (_pending_jvms != nullptr) {
1261 emit_non_safepoint();
1262 }
1263 _pending_jvms = nullptr;
1264 }
1265 };
1266
1267 void NonSafepointEmitter::emit_non_safepoint() {
1268 JVMState* youngest_jvms = _pending_jvms;
1269 int pc_offset = _pending_offset;
1270
1271 // Clear it now:
1272 _pending_jvms = nullptr;
1273
1274 DebugInformationRecorder* debug_info = C->debug_info();
1275 assert(debug_info->recording_non_safepoints(), "sanity");
1276
1277 debug_info->add_non_safepoint(pc_offset);
1278 int max_depth = youngest_jvms->depth();
1279
1280 // Visit scopes from oldest to youngest.
1281 for (int depth = 1; depth <= max_depth; depth++) {
1282 JVMState* jvms = youngest_jvms->of_depth(depth);
1283 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1284 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1285 methodHandle null_mh;
1286 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1287 }
1288
1289 // Mark the end of the scope set.
1290 debug_info->end_non_safepoint(pc_offset);
1291 }
1292
1293 //------------------------------init_buffer------------------------------------
1294 void PhaseOutput::estimate_buffer_size(int& const_req) {
1295
1296 // Set the initially allocated size
1297 const_req = initial_const_capacity;
1298
1299 // The extra spacing after the code is necessary on some platforms.
1300 // Sometimes we need to patch in a jump after the last instruction,
1301 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1302
1303 // Compute the byte offset where we can store the deopt pc.
1304 if (C->fixed_slots() != 0) {
1305 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1306 }
1307
1308 // Compute prolog code size
1309 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1310 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1311
1312 if (C->has_mach_constant_base_node()) {
1313 uint add_size = 0;
1314 // Fill the constant table.
1315 // Note: This must happen before shorten_branches.
1316 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1317 Block* b = C->cfg()->get_block(i);
1318
1319 for (uint j = 0; j < b->number_of_nodes(); j++) {
1320 Node* n = b->get_node(j);
1321
1322 // If the node is a MachConstantNode evaluate the constant
1323 // value section.
1324 if (n->is_MachConstant()) {
1325 MachConstantNode* machcon = n->as_MachConstant();
1326 machcon->eval_constant(C);
1327 } else if (n->is_Mach()) {
1328 // On Power there are more nodes that issue constants.
1329 add_size += (n->as_Mach()->ins_num_consts() * 8);
1330 }
1331 }
1332 }
1333
1334 // Calculate the offsets of the constants and the size of the
1335 // constant table (including the padding to the next section).
1336 constant_table().calculate_offsets_and_size();
1337 const_req = constant_table().alignment() + constant_table().size() + add_size;
1338 }
1339
1340 // Initialize the space for the BufferBlob used to find and verify
1341 // instruction size in MachNode::emit_size()
1342 init_scratch_buffer_blob(const_req);
1343 }
1344
1345 CodeBuffer* PhaseOutput::init_buffer() {
1346 int stub_req = _buf_sizes._stub;
1347 int code_req = _buf_sizes._code;
1348 int const_req = _buf_sizes._const;
1349
1350 int pad_req = NativeCall::byte_size();
1351
1352 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1353 stub_req += bs->estimate_stub_size();
1354
1355 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1356 // class HandlerImpl is platform-specific and defined in the *.ad files.
1357 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1358 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1359 stub_req += MAX_stubs_size; // ensure per-stub margin
1360 code_req += MAX_inst_size; // ensure per-instruction margin
1361
1362 if (StressCodeBuffers)
1363 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1364
1365 int total_req =
1366 const_req +
1367 code_req +
1368 pad_req +
1369 stub_req +
1370 exception_handler_req +
1371 deopt_handler_req; // deopt handler
1372
1373 if (C->has_method_handle_invokes())
1374 total_req += deopt_handler_req; // deopt MH handler
1375
1376 CodeBuffer* cb = code_buffer();
1377 cb->set_const_section_alignment(constant_table().alignment());
1378 cb->initialize(total_req, _buf_sizes._reloc);
1379
1380 // Have we run out of code space?
1381 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1382 C->record_failure("CodeCache is full");
1383 return nullptr;
1384 }
1385 // Configure the code buffer.
1386 cb->initialize_consts_size(const_req);
1387 cb->initialize_stubs_size(stub_req);
1388 cb->initialize_oop_recorder(C->env()->oop_recorder());
1389
1390 return cb;
1391 }
1392
1393 //------------------------------fill_buffer------------------------------------
1394 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1395 // blk_starts[] contains offsets calculated during short branches processing,
1396 // offsets should not be increased during following steps.
1397
1398 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1399 // of a loop. It is used to determine the padding for loop alignment.
1400 Compile::TracePhase tp(_t_fillBuffer);
1401
1402 compute_loop_first_inst_sizes();
1403
1404 // Create oopmap set.
1405 _oop_map_set = new OopMapSet();
1406
1407 // !!!!! This preserves old handling of oopmaps for now
1408 C->debug_info()->set_oopmaps(_oop_map_set);
1409
1410 uint nblocks = C->cfg()->number_of_blocks();
1411 // Count and start of implicit null check instructions
1412 uint inct_cnt = 0;
1413 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1414
1415 // Count and start of calls
1416 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1417
1418 uint return_offset = 0;
1419 int nop_size = (new MachNopNode())->size(C->regalloc());
1420
1421 int previous_offset = 0;
1422 int current_offset = 0;
1423 int last_call_offset = -1;
1424 int last_avoid_back_to_back_offset = -1;
1425 #ifdef ASSERT
1426 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1427 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1428 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1429 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1430 #endif
1431
1432 // Create an array of unused labels, one for each basic block, if printing is enabled
1433 #if defined(SUPPORT_OPTO_ASSEMBLY)
1434 int* node_offsets = nullptr;
1435 uint node_offset_limit = C->unique();
1436
1437 if (C->print_assembly()) {
1438 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1439 }
1440 if (node_offsets != nullptr) {
1441 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1442 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1443 }
1444 #endif
1445
1446 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1447
1448 // Emit the constant table.
1449 if (C->has_mach_constant_base_node()) {
1450 if (!constant_table().emit(masm)) {
1451 C->record_failure("consts section overflow");
1452 return;
1453 }
1454 }
1455
1456 // Create an array of labels, one for each basic block
1457 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1458 for (uint i = 0; i <= nblocks; i++) {
1459 blk_labels[i].init();
1460 }
1461
1462 // Now fill in the code buffer
1463 for (uint i = 0; i < nblocks; i++) {
1464 Block* block = C->cfg()->get_block(i);
1465 _block = block;
1466 Node* head = block->head();
1467
1468 // If this block needs to start aligned (i.e, can be reached other
1469 // than by falling-thru from the previous block), then force the
1470 // start of a new bundle.
1471 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1472 masm->code()->flush_bundle(true);
1473 }
1474
1475 #ifdef ASSERT
1476 if (!block->is_connector()) {
1477 stringStream st;
1478 block->dump_head(C->cfg(), &st);
1479 masm->block_comment(st.freeze());
1480 }
1481 jmp_target[i] = 0;
1482 jmp_offset[i] = 0;
1483 jmp_size[i] = 0;
1484 jmp_rule[i] = 0;
1485 #endif
1486 int blk_offset = current_offset;
1487
1488 // Define the label at the beginning of the basic block
1489 masm->bind(blk_labels[block->_pre_order]);
1490
1491 uint last_inst = block->number_of_nodes();
1492
1493 // Emit block normally, except for last instruction.
1494 // Emit means "dump code bits into code buffer".
1495 for (uint j = 0; j<last_inst; j++) {
1496 _index = j;
1497
1498 // Get the node
1499 Node* n = block->get_node(j);
1500
1501 // If this starts a new instruction group, then flush the current one
1502 // (but allow split bundles)
1503 if (Pipeline::requires_bundling() && starts_bundle(n))
1504 masm->code()->flush_bundle(false);
1505
1506 // Special handling for SafePoint/Call Nodes
1507 bool is_mcall = false;
1508 if (n->is_Mach()) {
1509 MachNode *mach = n->as_Mach();
1510 is_mcall = n->is_MachCall();
1511 bool is_sfn = n->is_MachSafePoint();
1512
1513 // If this requires all previous instructions be flushed, then do so
1514 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1515 masm->code()->flush_bundle(true);
1516 current_offset = masm->offset();
1517 }
1518
1519 // align the instruction if necessary
1520 int padding = mach->compute_padding(current_offset);
1521 // Make sure safepoint node for polling is distinct from a call's
1522 // return by adding a nop if needed.
1523 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1524 padding = nop_size;
1525 }
1526 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1527 current_offset == last_avoid_back_to_back_offset) {
1528 // Avoid back to back some instructions.
1529 padding = nop_size;
1530 }
1531
1532 if (padding > 0) {
1533 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1534 int nops_cnt = padding / nop_size;
1535 MachNode *nop = new MachNopNode(nops_cnt);
1536 block->insert_node(nop, j++);
1537 last_inst++;
1538 C->cfg()->map_node_to_block(nop, block);
1539 // Ensure enough space.
1540 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1541 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1542 C->record_failure("CodeCache is full");
1543 return;
1544 }
1545 nop->emit(masm, C->regalloc());
1546 masm->code()->flush_bundle(true);
1547 current_offset = masm->offset();
1548 }
1549
1550 bool observe_safepoint = is_sfn;
1551 // Remember the start of the last call in a basic block
1552 if (is_mcall) {
1553 MachCallNode *mcall = mach->as_MachCall();
1554
1555 // This destination address is NOT PC-relative
1556 mcall->method_set((intptr_t)mcall->entry_point());
1557
1558 // Save the return address
1559 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1560
1561 observe_safepoint = mcall->guaranteed_safepoint();
1562 }
1563
1564 // sfn will be valid whenever mcall is valid now because of inheritance
1565 if (observe_safepoint) {
1566 // Handle special safepoint nodes for synchronization
1567 if (!is_mcall) {
1568 MachSafePointNode *sfn = mach->as_MachSafePoint();
1569 // !!!!! Stubs only need an oopmap right now, so bail out
1570 if (sfn->jvms()->method() == nullptr) {
1571 // Write the oopmap directly to the code blob??!!
1572 continue;
1573 }
1574 } // End synchronization
1575
1576 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1577 current_offset);
1578 Process_OopMap_Node(mach, current_offset);
1579 } // End if safepoint
1580
1581 // If this is a null check, then add the start of the previous instruction to the list
1582 else if( mach->is_MachNullCheck() ) {
1583 inct_starts[inct_cnt++] = previous_offset;
1584 }
1585
1586 // If this is a branch, then fill in the label with the target BB's label
1587 else if (mach->is_MachBranch()) {
1588 // This requires the TRUE branch target be in succs[0]
1589 uint block_num = block->non_connector_successor(0)->_pre_order;
1590
1591 // Try to replace long branch,
1592 // it is mostly for back branches since forward branch's
1593 // distance is not updated yet.
1594 if (mach->may_be_short_branch()) {
1595 int br_size = n->size(C->regalloc());
1596 int offset = blk_starts[block_num] - current_offset;
1597 if (block_num >= i) {
1598 // Current and following block's offset are not
1599 // finalized yet, adjust distance by the difference
1600 // between calculated and final offsets of current block.
1601 offset -= (blk_starts[i] - blk_offset);
1602 }
1603 // In the following code a nop could be inserted before
1604 // the branch which will increase the backward distance.
1605 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1606 if (needs_padding && offset <= 0)
1607 offset -= nop_size;
1608
1609 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1610 // We've got a winner. Replace this branch.
1611 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1612
1613 // Update the jmp_size.
1614 int new_size = replacement->size(C->regalloc());
1615 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1616 // Insert padding between avoid_back_to_back branches.
1617 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1618 MachNode *nop = new MachNopNode();
1619 block->insert_node(nop, j++);
1620 C->cfg()->map_node_to_block(nop, block);
1621 last_inst++;
1622 nop->emit(masm, C->regalloc());
1623 masm->code()->flush_bundle(true);
1624 current_offset = masm->offset();
1625 }
1626 #ifdef ASSERT
1627 jmp_target[i] = block_num;
1628 jmp_offset[i] = current_offset - blk_offset;
1629 jmp_size[i] = new_size;
1630 jmp_rule[i] = mach->rule();
1631 #endif
1632 block->map_node(replacement, j);
1633 mach->subsume_by(replacement, C);
1634 n = replacement;
1635 mach = replacement;
1636 }
1637 }
1638 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1639 } else if (mach->ideal_Opcode() == Op_Jump) {
1640 for (uint h = 0; h < block->_num_succs; h++) {
1641 Block* succs_block = block->_succs[h];
1642 for (uint j = 1; j < succs_block->num_preds(); j++) {
1643 Node* jpn = succs_block->pred(j);
1644 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1645 uint block_num = succs_block->non_connector()->_pre_order;
1646 Label *blkLabel = &blk_labels[block_num];
1647 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1648 }
1649 }
1650 }
1651 } else if (!n->is_Proj()) {
1652 // Remember the beginning of the previous instruction, in case
1653 // it's followed by a flag-kill and a null-check. Happens on
1654 // Intel all the time, with add-to-memory kind of opcodes.
1655 previous_offset = current_offset;
1656 }
1657
1658 // Not an else-if!
1659 // If this is a trap based cmp then add its offset to the list.
1660 if (mach->is_TrapBasedCheckNode()) {
1661 inct_starts[inct_cnt++] = current_offset;
1662 }
1663 }
1664
1665 // Verify that there is sufficient space remaining
1666 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1667 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1668 C->record_failure("CodeCache is full");
1669 return;
1670 }
1671
1672 // Save the offset for the listing
1673 #if defined(SUPPORT_OPTO_ASSEMBLY)
1674 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1675 node_offsets[n->_idx] = masm->offset();
1676 }
1677 #endif
1678 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1679
1680 // "Normal" instruction case
1681 DEBUG_ONLY(uint instr_offset = masm->offset());
1682 n->emit(masm, C->regalloc());
1683 current_offset = masm->offset();
1684
1685 // Above we only verified that there is enough space in the instruction section.
1686 // However, the instruction may emit stubs that cause code buffer expansion.
1687 // Bail out here if expansion failed due to a lack of code cache space.
1688 if (C->failing()) {
1689 return;
1690 }
1691
1692 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1693 "ret_addr_offset() not within emitted code");
1694
1695 #ifdef ASSERT
1696 uint n_size = n->size(C->regalloc());
1697 if (n_size < (current_offset-instr_offset)) {
1698 MachNode* mach = n->as_Mach();
1699 n->dump();
1700 mach->dump_format(C->regalloc(), tty);
1701 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1702 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1703 tty->print_cr(" ------------------- ");
1704 BufferBlob* blob = this->scratch_buffer_blob();
1705 address blob_begin = blob->content_begin();
1706 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1707 assert(false, "wrong size of mach node");
1708 }
1709 #endif
1710 non_safepoints.observe_instruction(n, current_offset);
1711
1712 // mcall is last "call" that can be a safepoint
1713 // record it so we can see if a poll will directly follow it
1714 // in which case we'll need a pad to make the PcDesc sites unique
1715 // see 5010568. This can be slightly inaccurate but conservative
1716 // in the case that return address is not actually at current_offset.
1717 // This is a small price to pay.
1718
1719 if (is_mcall) {
1720 last_call_offset = current_offset;
1721 }
1722
1723 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1724 // Avoid back to back some instructions.
1725 last_avoid_back_to_back_offset = current_offset;
1726 }
1727
1728 } // End for all instructions in block
1729
1730 // If the next block is the top of a loop, pad this block out to align
1731 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1732 if (i < nblocks-1) {
1733 Block *nb = C->cfg()->get_block(i + 1);
1734 int padding = nb->alignment_padding(current_offset);
1735 if( padding > 0 ) {
1736 MachNode *nop = new MachNopNode(padding / nop_size);
1737 block->insert_node(nop, block->number_of_nodes());
1738 C->cfg()->map_node_to_block(nop, block);
1739 nop->emit(masm, C->regalloc());
1740 current_offset = masm->offset();
1741 }
1742 }
1743 // Verify that the distance for generated before forward
1744 // short branches is still valid.
1745 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1746
1747 // Save new block start offset
1748 blk_starts[i] = blk_offset;
1749 } // End of for all blocks
1750 blk_starts[nblocks] = current_offset;
1751
1752 non_safepoints.flush_at_end();
1753
1754 // Offset too large?
1755 if (C->failing()) return;
1756
1757 // Define a pseudo-label at the end of the code
1758 masm->bind( blk_labels[nblocks] );
1759
1760 // Compute the size of the first block
1761 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1762
1763 #ifdef ASSERT
1764 for (uint i = 0; i < nblocks; i++) { // For all blocks
1765 if (jmp_target[i] != 0) {
1766 int br_size = jmp_size[i];
1767 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1768 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1769 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]);
1770 assert(false, "Displacement too large for short jmp");
1771 }
1772 }
1773 }
1774 #endif
1775
1776 if (!masm->code()->finalize_stubs()) {
1777 C->record_failure("CodeCache is full");
1778 return;
1779 }
1780
1781 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1782 bs->emit_stubs(*masm->code());
1783 if (C->failing()) return;
1784
1785 // Fill in stubs.
1786 assert(masm->inst_mark() == nullptr, "should be.");
1787 _stub_list.emit(*masm);
1788 if (C->failing()) return;
1789
1790 #ifndef PRODUCT
1791 // Information on the size of the method, without the extraneous code
1792 Scheduling::increment_method_size(masm->offset());
1793 #endif
1794
1795 // ------------------
1796 // Fill in exception table entries.
1797 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1798
1799 // Only java methods have exception handlers and deopt handlers
1800 // class HandlerImpl is platform-specific and defined in the *.ad files.
1801 if (C->method()) {
1802 // Emit the exception handler code.
1803 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1804 if (C->failing()) {
1805 return; // CodeBuffer::expand failed
1806 }
1807 // Emit the deopt handler code.
1808 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1809
1810 // Emit the MethodHandle deopt handler code (if required).
1811 if (C->has_method_handle_invokes() && !C->failing()) {
1812 // We can use the same code as for the normal deopt handler, we
1813 // just need a different entry point address.
1814 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1815 }
1816 }
1817
1818 // One last check for failed CodeBuffer::expand:
1819 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1820 C->record_failure("CodeCache is full");
1821 return;
1822 }
1823
1824 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1825 if (C->print_assembly()) {
1826 tty->cr();
1827 tty->print_cr("============================= C2-compiled nmethod ==============================");
1828 }
1829 #endif
1830
1831 #if defined(SUPPORT_OPTO_ASSEMBLY)
1832 // Dump the assembly code, including basic-block numbers
1833 if (C->print_assembly()) {
1834 ttyLocker ttyl; // keep the following output all in one block
1835 if (!VMThread::should_terminate()) { // test this under the tty lock
1836 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1837 // We call them first and write to a stringStream, then we retake the lock to
1838 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1839 ResourceMark rm;
1840 stringStream method_metadata_str;
1841 if (C->method() != nullptr) {
1842 C->method()->print_metadata(&method_metadata_str);
1843 }
1844 stringStream dump_asm_str;
1845 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1846
1847 NoSafepointVerifier nsv;
1848 ttyLocker ttyl2;
1849 // This output goes directly to the tty, not the compiler log.
1850 // To enable tools to match it up with the compilation activity,
1851 // be sure to tag this tty output with the compile ID.
1852 if (xtty != nullptr) {
1853 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1854 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1855 }
1856 if (C->method() != nullptr) {
1857 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1858 tty->print_raw(method_metadata_str.freeze());
1859 } else if (C->stub_name() != nullptr) {
1860 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1861 }
1862 tty->cr();
1863 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1864 tty->print_raw(dump_asm_str.freeze());
1865 tty->print_cr("--------------------------------------------------------------------------------");
1866 if (xtty != nullptr) {
1867 xtty->tail("opto_assembly");
1868 }
1869 }
1870 }
1871 #endif
1872 }
1873
1874 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1875 _inc_table.set_size(cnt);
1876
1877 uint inct_cnt = 0;
1878 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1879 Block* block = C->cfg()->get_block(i);
1880 Node *n = nullptr;
1881 int j;
1882
1883 // Find the branch; ignore trailing NOPs.
1884 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1885 n = block->get_node(j);
1886 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1887 break;
1888 }
1889 }
1890
1891 // If we didn't find anything, continue
1892 if (j < 0) {
1893 continue;
1894 }
1895
1896 // Compute ExceptionHandlerTable subtable entry and add it
1897 // (skip empty blocks)
1898 if (n->is_Catch()) {
1899
1900 // Get the offset of the return from the call
1901 uint call_return = call_returns[block->_pre_order];
1902 #ifdef ASSERT
1903 assert( call_return > 0, "no call seen for this basic block" );
1904 while (block->get_node(--j)->is_MachProj()) ;
1905 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1906 #endif
1907 // last instruction is a CatchNode, find it's CatchProjNodes
1908 int nof_succs = block->_num_succs;
1909 // allocate space
1910 GrowableArray<intptr_t> handler_bcis(nof_succs);
1911 GrowableArray<intptr_t> handler_pcos(nof_succs);
1912 // iterate through all successors
1913 for (int j = 0; j < nof_succs; j++) {
1914 Block* s = block->_succs[j];
1915 bool found_p = false;
1916 for (uint k = 1; k < s->num_preds(); k++) {
1917 Node* pk = s->pred(k);
1918 if (pk->is_CatchProj() && pk->in(0) == n) {
1919 const CatchProjNode* p = pk->as_CatchProj();
1920 found_p = true;
1921 // add the corresponding handler bci & pco information
1922 if (p->_con != CatchProjNode::fall_through_index) {
1923 // p leads to an exception handler (and is not fall through)
1924 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1925 // no duplicates, please
1926 if (!handler_bcis.contains(p->handler_bci())) {
1927 uint block_num = s->non_connector()->_pre_order;
1928 handler_bcis.append(p->handler_bci());
1929 handler_pcos.append(blk_labels[block_num].loc_pos());
1930 }
1931 }
1932 }
1933 }
1934 assert(found_p, "no matching predecessor found");
1935 // Note: Due to empty block removal, one block may have
1936 // several CatchProj inputs, from the same Catch.
1937 }
1938
1939 // Set the offset of the return from the call
1940 assert(handler_bcis.find(-1) != -1, "must have default handler");
1941 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
1942 continue;
1943 }
1944
1945 // Handle implicit null exception table updates
1946 if (n->is_MachNullCheck()) {
1947 MachNode* access = n->in(1)->as_Mach();
1948 assert(access->barrier_data() == 0 ||
1949 access->is_late_expanded_null_check_candidate(),
1950 "Implicit null checks on memory accesses with barriers are only supported on nodes explicitly marked as null-check candidates");
1951 uint block_num = block->non_connector_successor(0)->_pre_order;
1952 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1953 continue;
1954 }
1955 // Handle implicit exception table updates: trap instructions.
1956 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1957 uint block_num = block->non_connector_successor(0)->_pre_order;
1958 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1959 continue;
1960 }
1961 } // End of for all blocks fill in exception table entries
1962 }
1963
1964 // Static Variables
1965 #ifndef PRODUCT
1966 uint Scheduling::_total_nop_size = 0;
1967 uint Scheduling::_total_method_size = 0;
1968 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1969 #endif
1970
1971 // Initializer for class Scheduling
1972
1973 Scheduling::Scheduling(Arena *arena, Compile &compile)
1974 : _arena(arena),
1975 _cfg(compile.cfg()),
1976 _regalloc(compile.regalloc()),
1977 _scheduled(arena),
1978 _available(arena),
1979 _reg_node(arena),
1980 _pinch_free_list(arena),
1981 _next_node(nullptr),
1982 _bundle_instr_count(0),
1983 _bundle_cycle_number(0),
1984 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
1985 {
1986 // Save the count
1987 _node_bundling_limit = compile.unique();
1988 uint node_max = _regalloc->node_regs_max_index();
1989
1990 compile.output()->set_node_bundling_limit(_node_bundling_limit);
1991
1992 // This one is persistent within the Compile class
1993 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1994
1995 // Allocate space for fixed-size arrays
1996 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1997 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1998
1999 // Clear the arrays
2000 for (uint i = 0; i < node_max; i++) {
2001 ::new (&_node_bundling_base[i]) Bundle();
2002 }
2003 memset(_uses, 0, node_max * sizeof(short));
2004 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2005
2006 // Clear the bundling information
2007 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2008
2009 // Get the last node
2010 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2011
2012 _next_node = block->get_node(block->number_of_nodes() - 1);
2013 }
2014
2015 // Step ahead "i" cycles
2016 void Scheduling::step(uint i) {
2017
2018 Bundle *bundle = node_bundling(_next_node);
2019 bundle->set_starts_bundle();
2020
2021 // Update the bundle record, but leave the flags information alone
2022 if (_bundle_instr_count > 0) {
2023 bundle->set_instr_count(_bundle_instr_count);
2024 bundle->set_resources_used(_bundle_use.resourcesUsed());
2025 }
2026
2027 // Update the state information
2028 _bundle_instr_count = 0;
2029 _bundle_cycle_number += i;
2030 _bundle_use.step(i);
2031 }
2032
2033 void Scheduling::step_and_clear() {
2034 Bundle *bundle = node_bundling(_next_node);
2035 bundle->set_starts_bundle();
2036
2037 // Update the bundle record
2038 if (_bundle_instr_count > 0) {
2039 bundle->set_instr_count(_bundle_instr_count);
2040 bundle->set_resources_used(_bundle_use.resourcesUsed());
2041
2042 _bundle_cycle_number += 1;
2043 }
2044
2045 // Clear the bundling information
2046 _bundle_instr_count = 0;
2047 _bundle_use.reset();
2048
2049 memcpy(_bundle_use_elements,
2050 Pipeline_Use::elaborated_elements,
2051 sizeof(Pipeline_Use::elaborated_elements));
2052 }
2053
2054 // Perform instruction scheduling and bundling over the sequence of
2055 // instructions in backwards order.
2056 void PhaseOutput::ScheduleAndBundle() {
2057
2058 // Don't optimize this if it isn't a method
2059 if (!C->method())
2060 return;
2061
2062 // Don't optimize this if scheduling is disabled
2063 if (!C->do_scheduling())
2064 return;
2065
2066 // Scheduling code works only with pairs (8 bytes) maximum.
2067 // And when the scalable vector register is used, we may spill/unspill
2068 // the whole reg regardless of the max vector size.
2069 if (C->max_vector_size() > 8 ||
2070 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2071 return;
2072 }
2073
2074 Compile::TracePhase tp(_t_instrSched);
2075
2076 // Create a data structure for all the scheduling information
2077 Scheduling scheduling(Thread::current()->resource_area(), *C);
2078
2079 // Walk backwards over each basic block, computing the needed alignment
2080 // Walk over all the basic blocks
2081 scheduling.DoScheduling();
2082
2083 #ifndef PRODUCT
2084 if (C->trace_opto_output()) {
2085 // Buffer and print all at once
2086 ResourceMark rm;
2087 stringStream ss;
2088 ss.print("\n---- After ScheduleAndBundle ----\n");
2089 print_scheduling(&ss);
2090 tty->print("%s", ss.as_string());
2091 }
2092 #endif
2093 }
2094
2095 #ifndef PRODUCT
2096 // Separated out so that it can be called directly from debugger
2097 void PhaseOutput::print_scheduling() {
2098 print_scheduling(tty);
2099 }
2100
2101 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2102 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2103 output_stream->print("\nBB#%03d:\n", i);
2104 Block* block = C->cfg()->get_block(i);
2105 for (uint j = 0; j < block->number_of_nodes(); j++) {
2106 Node* n = block->get_node(j);
2107 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2108 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2109 n->dump("\n", false, output_stream);
2110 }
2111 }
2112 }
2113 #endif
2114
2115 // See if this node fits into the present instruction bundle
2116 bool Scheduling::NodeFitsInBundle(Node *n) {
2117 uint n_idx = n->_idx;
2118
2119 // If the node cannot be scheduled this cycle, skip it
2120 if (_current_latency[n_idx] > _bundle_cycle_number) {
2121 #ifndef PRODUCT
2122 if (_cfg->C->trace_opto_output())
2123 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2124 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2125 #endif
2126 return (false);
2127 }
2128
2129 const Pipeline *node_pipeline = n->pipeline();
2130
2131 uint instruction_count = node_pipeline->instructionCount();
2132 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2133 instruction_count = 0;
2134
2135 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2136 #ifndef PRODUCT
2137 if (_cfg->C->trace_opto_output())
2138 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2139 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2140 #endif
2141 return (false);
2142 }
2143
2144 // Don't allow non-machine nodes to be handled this way
2145 if (!n->is_Mach() && instruction_count == 0)
2146 return (false);
2147
2148 // See if there is any overlap
2149 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2150
2151 if (delay > 0) {
2152 #ifndef PRODUCT
2153 if (_cfg->C->trace_opto_output())
2154 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2155 #endif
2156 return false;
2157 }
2158
2159 #ifndef PRODUCT
2160 if (_cfg->C->trace_opto_output())
2161 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2162 #endif
2163
2164 return true;
2165 }
2166
2167 Node * Scheduling::ChooseNodeToBundle() {
2168 uint siz = _available.size();
2169
2170 if (siz == 0) {
2171
2172 #ifndef PRODUCT
2173 if (_cfg->C->trace_opto_output())
2174 tty->print("# ChooseNodeToBundle: null\n");
2175 #endif
2176 return (nullptr);
2177 }
2178
2179 // Fast path, if only 1 instruction in the bundle
2180 if (siz == 1) {
2181 #ifndef PRODUCT
2182 if (_cfg->C->trace_opto_output()) {
2183 tty->print("# ChooseNodeToBundle (only 1): ");
2184 _available[0]->dump();
2185 }
2186 #endif
2187 return (_available[0]);
2188 }
2189
2190 // Don't bother, if the bundle is already full
2191 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2192 for ( uint i = 0; i < siz; i++ ) {
2193 Node *n = _available[i];
2194
2195 // Skip projections, we'll handle them another way
2196 if (n->is_Proj())
2197 continue;
2198
2199 // This presupposed that instructions are inserted into the
2200 // available list in a legality order; i.e. instructions that
2201 // must be inserted first are at the head of the list
2202 if (NodeFitsInBundle(n)) {
2203 #ifndef PRODUCT
2204 if (_cfg->C->trace_opto_output()) {
2205 tty->print("# ChooseNodeToBundle: ");
2206 n->dump();
2207 }
2208 #endif
2209 return (n);
2210 }
2211 }
2212 }
2213
2214 // Nothing fits in this bundle, choose the highest priority
2215 #ifndef PRODUCT
2216 if (_cfg->C->trace_opto_output()) {
2217 tty->print("# ChooseNodeToBundle: ");
2218 _available[0]->dump();
2219 }
2220 #endif
2221
2222 return _available[0];
2223 }
2224
2225 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2226 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2227
2228 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2229 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2230 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2231 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2232
2233 // Comparison between stack -> reg and stack -> reg
2234 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2235 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2236 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2237 }
2238
2239 // Comparison between reg -> stack and reg -> stack
2240 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2241 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2242 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2243 }
2244
2245 return 0; // Not comparable
2246 }
2247
2248 void Scheduling::AddNodeToAvailableList(Node *n) {
2249 assert( !n->is_Proj(), "projections never directly made available" );
2250 #ifndef PRODUCT
2251 if (_cfg->C->trace_opto_output()) {
2252 tty->print("# AddNodeToAvailableList: ");
2253 n->dump();
2254 }
2255 #endif
2256
2257 int latency = _current_latency[n->_idx];
2258
2259 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2260 // for stack spilling or unspilling have the same latency, we sort
2261 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2262 // have more opportunities to do ld/st merging
2263 uint i;
2264 for (i = 0; i < _available.size(); i++) {
2265 if (_current_latency[_available[i]->_idx] > latency) {
2266 break;
2267 } else if (_current_latency[_available[i]->_idx] == latency &&
2268 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2269 compare_two_spill_nodes(n, _available[i]) > 0) {
2270 break;
2271 }
2272 }
2273
2274 // Special Check for compares following branches
2275 if( n->is_Mach() && _scheduled.size() > 0 ) {
2276 int op = n->as_Mach()->ideal_Opcode();
2277 Node *last = _scheduled[0];
2278 if( last->is_MachIf() && last->in(1) == n &&
2279 ( op == Op_CmpI ||
2280 op == Op_CmpU ||
2281 op == Op_CmpUL ||
2282 op == Op_CmpP ||
2283 op == Op_CmpF ||
2284 op == Op_CmpD ||
2285 op == Op_CmpL ) ) {
2286
2287 // Recalculate position, moving to front of same latency
2288 for ( i=0 ; i < _available.size(); i++ )
2289 if (_current_latency[_available[i]->_idx] >= latency)
2290 break;
2291 }
2292 }
2293
2294 // Insert the node in the available list
2295 _available.insert(i, n);
2296
2297 #ifndef PRODUCT
2298 if (_cfg->C->trace_opto_output())
2299 dump_available();
2300 #endif
2301 }
2302
2303 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2304 for ( uint i=0; i < n->len(); i++ ) {
2305 Node *def = n->in(i);
2306 if (!def) continue;
2307 if( def->is_Proj() ) // If this is a machine projection, then
2308 def = def->in(0); // propagate usage thru to the base instruction
2309
2310 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2311 continue;
2312 }
2313
2314 // Compute the latency
2315 uint l = _bundle_cycle_number + n->latency(i);
2316 if (_current_latency[def->_idx] < l)
2317 _current_latency[def->_idx] = l;
2318
2319 // If this does not have uses then schedule it
2320 if ((--_uses[def->_idx]) == 0)
2321 AddNodeToAvailableList(def);
2322 }
2323 }
2324
2325 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2326 #ifndef PRODUCT
2327 if (_cfg->C->trace_opto_output()) {
2328 tty->print("# AddNodeToBundle: ");
2329 n->dump();
2330 }
2331 #endif
2332
2333 // Remove this from the available list
2334 uint i;
2335 for (i = 0; i < _available.size(); i++)
2336 if (_available[i] == n)
2337 break;
2338 assert(i < _available.size(), "entry in _available list not found");
2339 _available.remove(i);
2340
2341 // See if this fits in the current bundle
2342 const Pipeline *node_pipeline = n->pipeline();
2343 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2344
2345
2346 // Get the number of instructions
2347 uint instruction_count = node_pipeline->instructionCount();
2348 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2349 instruction_count = 0;
2350
2351 // Compute the latency information
2352 uint delay = 0;
2353
2354 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2355 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2356 if (relative_latency < 0)
2357 relative_latency = 0;
2358
2359 delay = _bundle_use.full_latency(relative_latency, node_usage);
2360
2361 // Does not fit in this bundle, start a new one
2362 if (delay > 0) {
2363 step(delay);
2364
2365 #ifndef PRODUCT
2366 if (_cfg->C->trace_opto_output())
2367 tty->print("# *** STEP(%d) ***\n", delay);
2368 #endif
2369 }
2370 }
2371
2372 if (delay == 0) {
2373 if (node_pipeline->hasMultipleBundles()) {
2374 #ifndef PRODUCT
2375 if (_cfg->C->trace_opto_output())
2376 tty->print("# *** STEP(multiple instructions) ***\n");
2377 #endif
2378 step(1);
2379 }
2380
2381 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2382 #ifndef PRODUCT
2383 if (_cfg->C->trace_opto_output())
2384 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2385 instruction_count + _bundle_instr_count,
2386 Pipeline::_max_instrs_per_cycle);
2387 #endif
2388 step(1);
2389 }
2390 }
2391
2392 // Set the node's latency
2393 _current_latency[n->_idx] = _bundle_cycle_number;
2394
2395 // Now merge the functional unit information
2396 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2397 _bundle_use.add_usage(node_usage);
2398
2399 // Increment the number of instructions in this bundle
2400 _bundle_instr_count += instruction_count;
2401
2402 // Remember this node for later
2403 if (n->is_Mach())
2404 _next_node = n;
2405
2406 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2407 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2408 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2409 // into the block. All other scheduled nodes get put in the schedule here.
2410 int op = n->Opcode();
2411 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2412 (op != Op_Node && // Not an unused antidepedence node and
2413 // not an unallocated boxlock
2414 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2415
2416 // Push any trailing projections
2417 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2418 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2419 Node *foi = n->fast_out(i);
2420 if( foi->is_Proj() )
2421 _scheduled.push(foi);
2422 }
2423 }
2424
2425 // Put the instruction in the schedule list
2426 _scheduled.push(n);
2427 }
2428
2429 #ifndef PRODUCT
2430 if (_cfg->C->trace_opto_output())
2431 dump_available();
2432 #endif
2433
2434 // Walk all the definitions, decrementing use counts, and
2435 // if a definition has a 0 use count, place it in the available list.
2436 DecrementUseCounts(n,bb);
2437 }
2438
2439 // This method sets the use count within a basic block. We will ignore all
2440 // uses outside the current basic block. As we are doing a backwards walk,
2441 // any node we reach that has a use count of 0 may be scheduled. This also
2442 // avoids the problem of cyclic references from phi nodes, as long as phi
2443 // nodes are at the front of the basic block. This method also initializes
2444 // the available list to the set of instructions that have no uses within this
2445 // basic block.
2446 void Scheduling::ComputeUseCount(const Block *bb) {
2447 #ifndef PRODUCT
2448 if (_cfg->C->trace_opto_output())
2449 tty->print("# -> ComputeUseCount\n");
2450 #endif
2451
2452 // Clear the list of available and scheduled instructions, just in case
2453 _available.clear();
2454 _scheduled.clear();
2455
2456 #ifdef ASSERT
2457 for( uint i=0; i < bb->number_of_nodes(); i++ )
2458 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2459 #endif
2460
2461 // Force the _uses count to never go to zero for unscheduable pieces
2462 // of the block
2463 for( uint k = 0; k < _bb_start; k++ )
2464 _uses[bb->get_node(k)->_idx] = 1;
2465 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2466 _uses[bb->get_node(l)->_idx] = 1;
2467
2468 // Iterate backwards over the instructions in the block. Don't count the
2469 // branch projections at end or the block header instructions.
2470 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2471 Node *n = bb->get_node(j);
2472 if( n->is_Proj() ) continue; // Projections handled another way
2473
2474 // Account for all uses
2475 for ( uint k = 0; k < n->len(); k++ ) {
2476 Node *inp = n->in(k);
2477 if (!inp) continue;
2478 assert(inp != n, "no cycles allowed" );
2479 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2480 if (inp->is_Proj()) { // Skip through Proj's
2481 inp = inp->in(0);
2482 }
2483 ++_uses[inp->_idx]; // Count 1 block-local use
2484 }
2485 }
2486
2487 // If this instruction has a 0 use count, then it is available
2488 if (!_uses[n->_idx]) {
2489 _current_latency[n->_idx] = _bundle_cycle_number;
2490 AddNodeToAvailableList(n);
2491 }
2492
2493 #ifndef PRODUCT
2494 if (_cfg->C->trace_opto_output()) {
2495 tty->print("# uses: %3d: ", _uses[n->_idx]);
2496 n->dump();
2497 }
2498 #endif
2499 }
2500
2501 #ifndef PRODUCT
2502 if (_cfg->C->trace_opto_output())
2503 tty->print("# <- ComputeUseCount\n");
2504 #endif
2505 }
2506
2507 // This routine performs scheduling on each basic block in reverse order,
2508 // using instruction latencies and taking into account function unit
2509 // availability.
2510 void Scheduling::DoScheduling() {
2511 #ifndef PRODUCT
2512 if (_cfg->C->trace_opto_output())
2513 tty->print("# -> DoScheduling\n");
2514 #endif
2515
2516 Block *succ_bb = nullptr;
2517 Block *bb;
2518 Compile* C = Compile::current();
2519
2520 // Walk over all the basic blocks in reverse order
2521 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2522 bb = _cfg->get_block(i);
2523
2524 #ifndef PRODUCT
2525 if (_cfg->C->trace_opto_output()) {
2526 tty->print("# Schedule BB#%03d (initial)\n", i);
2527 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2528 bb->get_node(j)->dump();
2529 }
2530 }
2531 #endif
2532
2533 // On the head node, skip processing
2534 if (bb == _cfg->get_root_block()) {
2535 continue;
2536 }
2537
2538 // Skip empty, connector blocks
2539 if (bb->is_connector())
2540 continue;
2541
2542 // If the following block is not the sole successor of
2543 // this one, then reset the pipeline information
2544 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2545 #ifndef PRODUCT
2546 if (_cfg->C->trace_opto_output()) {
2547 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2548 _next_node->_idx, _bundle_instr_count);
2549 }
2550 #endif
2551 step_and_clear();
2552 }
2553
2554 // Leave untouched the starting instruction, any Phis, a CreateEx node
2555 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2556 _bb_end = bb->number_of_nodes()-1;
2557 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2558 Node *n = bb->get_node(_bb_start);
2559 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2560 // Also, MachIdealNodes do not get scheduled
2561 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2562 MachNode *mach = n->as_Mach();
2563 int iop = mach->ideal_Opcode();
2564 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2565 if( iop == Op_Con ) continue; // Do not schedule Top
2566 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2567 mach->pipeline() == MachNode::pipeline_class() &&
2568 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2569 continue;
2570 break; // Funny loop structure to be sure...
2571 }
2572 // Compute last "interesting" instruction in block - last instruction we
2573 // might schedule. _bb_end points just after last schedulable inst.
2574 Node *last = bb->get_node(_bb_end);
2575 // Ignore trailing NOPs.
2576 while (_bb_end > 0 && last->is_Mach() &&
2577 last->as_Mach()->ideal_Opcode() == Op_Con) {
2578 last = bb->get_node(--_bb_end);
2579 }
2580 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2581 if( last->is_Catch() ||
2582 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2583 // There might be a prior call. Skip it.
2584 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2585 } else if( last->is_MachNullCheck() ) {
2586 // Backup so the last null-checked memory instruction is
2587 // outside the schedulable range. Skip over the nullcheck,
2588 // projection, and the memory nodes.
2589 Node *mem = last->in(1);
2590 do {
2591 _bb_end--;
2592 } while (mem != bb->get_node(_bb_end));
2593 } else {
2594 // Set _bb_end to point after last schedulable inst.
2595 _bb_end++;
2596 }
2597
2598 assert( _bb_start <= _bb_end, "inverted block ends" );
2599
2600 // Compute the register antidependencies for the basic block
2601 ComputeRegisterAntidependencies(bb);
2602 if (C->failing()) return; // too many D-U pinch points
2603
2604 // Compute the usage within the block, and set the list of all nodes
2605 // in the block that have no uses within the block.
2606 ComputeUseCount(bb);
2607
2608 // Schedule the remaining instructions in the block
2609 while ( _available.size() > 0 ) {
2610 Node *n = ChooseNodeToBundle();
2611 guarantee(n != nullptr, "no nodes available");
2612 AddNodeToBundle(n,bb);
2613 }
2614
2615 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2616 #ifdef ASSERT
2617 for( uint l = _bb_start; l < _bb_end; l++ ) {
2618 Node *n = bb->get_node(l);
2619 uint m;
2620 for( m = 0; m < _bb_end-_bb_start; m++ )
2621 if( _scheduled[m] == n )
2622 break;
2623 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2624 }
2625 #endif
2626
2627 // Now copy the instructions (in reverse order) back to the block
2628 for ( uint k = _bb_start; k < _bb_end; k++ )
2629 bb->map_node(_scheduled[_bb_end-k-1], k);
2630
2631 #ifndef PRODUCT
2632 if (_cfg->C->trace_opto_output()) {
2633 tty->print("# Schedule BB#%03d (final)\n", i);
2634 uint current = 0;
2635 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2636 Node *n = bb->get_node(j);
2637 if( valid_bundle_info(n) ) {
2638 Bundle *bundle = node_bundling(n);
2639 if (bundle->instr_count() > 0) {
2640 tty->print("*** Bundle: ");
2641 bundle->dump();
2642 }
2643 n->dump();
2644 }
2645 }
2646 }
2647 #endif
2648 #ifdef ASSERT
2649 verify_good_schedule(bb,"after block local scheduling");
2650 #endif
2651 }
2652
2653 #ifndef PRODUCT
2654 if (_cfg->C->trace_opto_output())
2655 tty->print("# <- DoScheduling\n");
2656 #endif
2657
2658 // Record final node-bundling array location
2659 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2660
2661 } // end DoScheduling
2662
2663 // Verify that no live-range used in the block is killed in the block by a
2664 // wrong DEF. This doesn't verify live-ranges that span blocks.
2665
2666 // Check for edge existence. Used to avoid adding redundant precedence edges.
2667 static bool edge_from_to( Node *from, Node *to ) {
2668 for( uint i=0; i<from->len(); i++ )
2669 if( from->in(i) == to )
2670 return true;
2671 return false;
2672 }
2673
2674 #ifdef ASSERT
2675 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2676 // Check for bad kills
2677 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2678 Node *prior_use = _reg_node[def];
2679 if( prior_use && !edge_from_to(prior_use,n) ) {
2680 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2681 n->dump();
2682 tty->print_cr("...");
2683 prior_use->dump();
2684 assert(edge_from_to(prior_use,n), "%s", msg);
2685 }
2686 _reg_node.map(def,nullptr); // Kill live USEs
2687 }
2688 }
2689
2690 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2691
2692 // Zap to something reasonable for the verify code
2693 _reg_node.clear();
2694
2695 // Walk over the block backwards. Check to make sure each DEF doesn't
2696 // kill a live value (other than the one it's supposed to). Add each
2697 // USE to the live set.
2698 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2699 Node *n = b->get_node(i);
2700 int n_op = n->Opcode();
2701 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2702 // Fat-proj kills a slew of registers
2703 RegMaskIterator rmi(n->out_RegMask());
2704 while (rmi.has_next()) {
2705 OptoReg::Name kill = rmi.next();
2706 verify_do_def(n, kill, msg);
2707 }
2708 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2709 // Get DEF'd registers the normal way
2710 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2711 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2712 }
2713
2714 // Now make all USEs live
2715 for( uint i=1; i<n->req(); i++ ) {
2716 Node *def = n->in(i);
2717 assert(def != nullptr, "input edge required");
2718 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2719 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2720 if( OptoReg::is_valid(reg_lo) ) {
2721 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2722 _reg_node.map(reg_lo,n);
2723 }
2724 if( OptoReg::is_valid(reg_hi) ) {
2725 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2726 _reg_node.map(reg_hi,n);
2727 }
2728 }
2729
2730 }
2731
2732 // Zap to something reasonable for the Antidependence code
2733 _reg_node.clear();
2734 }
2735 #endif
2736
2737 // Conditionally add precedence edges. Avoid putting edges on Projs.
2738 static void add_prec_edge_from_to( Node *from, Node *to ) {
2739 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2740 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2741 from = from->in(0);
2742 }
2743 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2744 !edge_from_to( from, to ) ) // Avoid duplicate edge
2745 from->add_prec(to);
2746 }
2747
2748 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2749 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2750 return;
2751
2752 if (OptoReg::is_reg(def_reg)) {
2753 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2754 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2755 // This is one of the high slots of a vector register.
2756 // ScheduleAndBundle already checked there are no live wide
2757 // vectors in this method so it can be safely ignored.
2758 return;
2759 }
2760 }
2761
2762 Node *pinch = _reg_node[def_reg]; // Get pinch point
2763 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2764 is_def ) { // Check for a true def (not a kill)
2765 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2766 return;
2767 }
2768
2769 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2770 DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); )
2771
2772 // After some number of kills there _may_ be a later def
2773 Node *later_def = nullptr;
2774
2775 Compile* C = Compile::current();
2776
2777 // Finding a kill requires a real pinch-point.
2778 // Check for not already having a pinch-point.
2779 // Pinch points are Op_Node's.
2780 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2781 later_def = pinch; // Must be def/kill as optimistic pinch-point
2782 if ( _pinch_free_list.size() > 0) {
2783 pinch = _pinch_free_list.pop();
2784 } else {
2785 pinch = new Node(1); // Pinch point to-be
2786 }
2787 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2788 DEBUG_ONLY( pinch->dump(); );
2789 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2790 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2791 return;
2792 }
2793 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2794 _reg_node.map(def_reg,pinch); // Record pinch-point
2795 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2796 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2797 pinch->init_req(0, C->top()); // set not null for the next call
2798 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2799 later_def = nullptr; // and no later def
2800 }
2801 pinch->set_req(0,later_def); // Hook later def so we can find it
2802 } else { // Else have valid pinch point
2803 if( pinch->in(0) ) // If there is a later-def
2804 later_def = pinch->in(0); // Get it
2805 }
2806
2807 // Add output-dependence edge from later def to kill
2808 if( later_def ) // If there is some original def
2809 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2810
2811 // See if current kill is also a use, and so is forced to be the pinch-point.
2812 if( pinch->Opcode() == Op_Node ) {
2813 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2814 for( uint i=1; i<uses->req(); i++ ) {
2815 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2816 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2817 // Yes, found a use/kill pinch-point
2818 pinch->set_req(0,nullptr); //
2819 pinch->replace_by(kill); // Move anti-dep edges up
2820 pinch = kill;
2821 _reg_node.map(def_reg,pinch);
2822 return;
2823 }
2824 }
2825 }
2826
2827 // Add edge from kill to pinch-point
2828 add_prec_edge_from_to(kill,pinch);
2829 }
2830
2831 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2832 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2833 return;
2834 Node *pinch = _reg_node[use_reg]; // Get pinch point
2835 // Check for no later def_reg/kill in block
2836 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
2837 // Use has to be block-local as well
2838 _cfg->get_block_for_node(use) == b) {
2839 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2840 pinch->req() == 1 ) { // pinch not yet in block?
2841 pinch->del_req(0); // yank pointer to later-def, also set flag
2842 // Insert the pinch-point in the block just after the last use
2843 b->insert_node(pinch, b->find_node(use) + 1);
2844 _bb_end++; // Increase size scheduled region in block
2845 }
2846
2847 add_prec_edge_from_to(pinch,use);
2848 }
2849 }
2850
2851 // We insert antidependences between the reads and following write of
2852 // allocated registers to prevent illegal code motion. Hopefully, the
2853 // number of added references should be fairly small, especially as we
2854 // are only adding references within the current basic block.
2855 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2856
2857 #ifdef ASSERT
2858 verify_good_schedule(b,"before block local scheduling");
2859 #endif
2860
2861 // A valid schedule, for each register independently, is an endless cycle
2862 // of: a def, then some uses (connected to the def by true dependencies),
2863 // then some kills (defs with no uses), finally the cycle repeats with a new
2864 // def. The uses are allowed to float relative to each other, as are the
2865 // kills. No use is allowed to slide past a kill (or def). This requires
2866 // antidependencies between all uses of a single def and all kills that
2867 // follow, up to the next def. More edges are redundant, because later defs
2868 // & kills are already serialized with true or antidependencies. To keep
2869 // the edge count down, we add a 'pinch point' node if there's more than
2870 // one use or more than one kill/def.
2871
2872 // We add dependencies in one bottom-up pass.
2873
2874 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2875
2876 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2877 // register. If not, we record the DEF/KILL in _reg_node, the
2878 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2879 // "pinch point", a new Node that's in the graph but not in the block.
2880 // We put edges from the prior and current DEF/KILLs to the pinch point.
2881 // We put the pinch point in _reg_node. If there's already a pinch point
2882 // we merely add an edge from the current DEF/KILL to the pinch point.
2883
2884 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2885 // put an edge from the pinch point to the USE.
2886
2887 // To be expedient, the _reg_node array is pre-allocated for the whole
2888 // compilation. _reg_node is lazily initialized; it either contains a null,
2889 // or a valid def/kill/pinch-point, or a leftover node from some prior
2890 // block. Leftover node from some prior block is treated like a null (no
2891 // prior def, so no anti-dependence needed). Valid def is distinguished by
2892 // it being in the current block.
2893 bool fat_proj_seen = false;
2894 uint last_safept = _bb_end-1;
2895 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
2896 Node* last_safept_node = end_node;
2897 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2898 Node *n = b->get_node(i);
2899 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2900 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2901 // Fat-proj kills a slew of registers
2902 // This can add edges to 'n' and obscure whether or not it was a def,
2903 // hence the is_def flag.
2904 fat_proj_seen = true;
2905 RegMaskIterator rmi(n->out_RegMask());
2906 while (rmi.has_next()) {
2907 OptoReg::Name kill = rmi.next();
2908 anti_do_def(b, n, kill, is_def);
2909 }
2910 } else {
2911 // Get DEF'd registers the normal way
2912 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2913 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2914 }
2915
2916 // Kill projections on a branch should appear to occur on the
2917 // branch, not afterwards, so grab the masks from the projections
2918 // and process them.
2919 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
2920 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2921 Node* use = n->fast_out(i);
2922 if (use->is_Proj()) {
2923 RegMaskIterator rmi(use->out_RegMask());
2924 while (rmi.has_next()) {
2925 OptoReg::Name kill = rmi.next();
2926 anti_do_def(b, n, kill, false);
2927 }
2928 }
2929 }
2930 }
2931
2932 // Check each register used by this instruction for a following DEF/KILL
2933 // that must occur afterward and requires an anti-dependence edge.
2934 for( uint j=0; j<n->req(); j++ ) {
2935 Node *def = n->in(j);
2936 if( def ) {
2937 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2938 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2939 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2940 }
2941 }
2942 // Do not allow defs of new derived values to float above GC
2943 // points unless the base is definitely available at the GC point.
2944
2945 Node *m = b->get_node(i);
2946
2947 // Add precedence edge from following safepoint to use of derived pointer
2948 if( last_safept_node != end_node &&
2949 m != last_safept_node) {
2950 for (uint k = 1; k < m->req(); k++) {
2951 const Type *t = m->in(k)->bottom_type();
2952 if( t->isa_oop_ptr() &&
2953 t->is_ptr()->offset() != 0 ) {
2954 last_safept_node->add_prec( m );
2955 break;
2956 }
2957 }
2958 }
2959
2960 if( n->jvms() ) { // Precedence edge from derived to safept
2961 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2962 if( b->get_node(last_safept) != last_safept_node ) {
2963 last_safept = b->find_node(last_safept_node);
2964 }
2965 for( uint j=last_safept; j > i; j-- ) {
2966 Node *mach = b->get_node(j);
2967 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2968 mach->add_prec( n );
2969 }
2970 last_safept = i;
2971 last_safept_node = m;
2972 }
2973 }
2974
2975 if (fat_proj_seen) {
2976 // Garbage collect pinch nodes that were not consumed.
2977 // They are usually created by a fat kill MachProj for a call.
2978 garbage_collect_pinch_nodes();
2979 }
2980 }
2981
2982 // Garbage collect pinch nodes for reuse by other blocks.
2983 //
2984 // The block scheduler's insertion of anti-dependence
2985 // edges creates many pinch nodes when the block contains
2986 // 2 or more Calls. A pinch node is used to prevent a
2987 // combinatorial explosion of edges. If a set of kills for a
2988 // register is anti-dependent on a set of uses (or defs), rather
2989 // than adding an edge in the graph between each pair of kill
2990 // and use (or def), a pinch is inserted between them:
2991 //
2992 // use1 use2 use3
2993 // \ | /
2994 // \ | /
2995 // pinch
2996 // / | \
2997 // / | \
2998 // kill1 kill2 kill3
2999 //
3000 // One pinch node is created per register killed when
3001 // the second call is encountered during a backwards pass
3002 // over the block. Most of these pinch nodes are never
3003 // wired into the graph because the register is never
3004 // used or def'ed in the block.
3005 //
3006 void Scheduling::garbage_collect_pinch_nodes() {
3007 #ifndef PRODUCT
3008 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3009 #endif
3010 int trace_cnt = 0;
3011 for (uint k = 0; k < _reg_node.max(); k++) {
3012 Node* pinch = _reg_node[k];
3013 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3014 // no predecence input edges
3015 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3016 cleanup_pinch(pinch);
3017 _pinch_free_list.push(pinch);
3018 _reg_node.map(k, nullptr);
3019 #ifndef PRODUCT
3020 if (_cfg->C->trace_opto_output()) {
3021 trace_cnt++;
3022 if (trace_cnt > 40) {
3023 tty->print("\n");
3024 trace_cnt = 0;
3025 }
3026 tty->print(" %d", pinch->_idx);
3027 }
3028 #endif
3029 }
3030 }
3031 #ifndef PRODUCT
3032 if (_cfg->C->trace_opto_output()) tty->print("\n");
3033 #endif
3034 }
3035
3036 // Clean up a pinch node for reuse.
3037 void Scheduling::cleanup_pinch( Node *pinch ) {
3038 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3039
3040 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3041 Node* use = pinch->last_out(i);
3042 uint uses_found = 0;
3043 for (uint j = use->req(); j < use->len(); j++) {
3044 if (use->in(j) == pinch) {
3045 use->rm_prec(j);
3046 uses_found++;
3047 }
3048 }
3049 assert(uses_found > 0, "must be a precedence edge");
3050 i -= uses_found; // we deleted 1 or more copies of this edge
3051 }
3052 // May have a later_def entry
3053 pinch->set_req(0, nullptr);
3054 }
3055
3056 #ifndef PRODUCT
3057
3058 void Scheduling::dump_available() const {
3059 tty->print("#Availist ");
3060 for (uint i = 0; i < _available.size(); i++)
3061 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3062 tty->cr();
3063 }
3064
3065 // Print Scheduling Statistics
3066 void Scheduling::print_statistics() {
3067 // Print the size added by nops for bundling
3068 tty->print("Nops added %d bytes to total of %d bytes",
3069 _total_nop_size, _total_method_size);
3070 if (_total_method_size > 0)
3071 tty->print(", for %.2f%%",
3072 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3073 tty->print("\n");
3074
3075 uint total_instructions = 0, total_bundles = 0;
3076
3077 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3078 uint bundle_count = _total_instructions_per_bundle[i];
3079 total_instructions += bundle_count * i;
3080 total_bundles += bundle_count;
3081 }
3082
3083 if (total_bundles > 0)
3084 tty->print("Average ILP (excluding nops) is %.2f\n",
3085 ((double)total_instructions) / ((double)total_bundles));
3086 }
3087 #endif
3088
3089 //-----------------------init_scratch_buffer_blob------------------------------
3090 // Construct a temporary BufferBlob and cache it for this compile.
3091 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3092 // If there is already a scratch buffer blob allocated and the
3093 // constant section is big enough, use it. Otherwise free the
3094 // current and allocate a new one.
3095 BufferBlob* blob = scratch_buffer_blob();
3096 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3097 // Use the current blob.
3098 } else {
3099 if (blob != nullptr) {
3100 BufferBlob::free(blob);
3101 }
3102
3103 ResourceMark rm;
3104 _scratch_const_size = const_size;
3105 int size = C2Compiler::initial_code_buffer_size(const_size);
3106 blob = BufferBlob::create("Compile::scratch_buffer", size);
3107 // Record the buffer blob for next time.
3108 set_scratch_buffer_blob(blob);
3109 // Have we run out of code space?
3110 if (scratch_buffer_blob() == nullptr) {
3111 // Let CompilerBroker disable further compilations.
3112 C->record_failure("Not enough space for scratch buffer in CodeCache");
3113 return;
3114 }
3115 }
3116
3117 // Initialize the relocation buffers
3118 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3119 set_scratch_locs_memory(locs_buf);
3120 }
3121
3122
3123 //-----------------------scratch_emit_size-------------------------------------
3124 // Helper function that computes size by emitting code
3125 uint PhaseOutput::scratch_emit_size(const Node* n) {
3126 // Start scratch_emit_size section.
3127 set_in_scratch_emit_size(true);
3128
3129 // Emit into a trash buffer and count bytes emitted.
3130 // This is a pretty expensive way to compute a size,
3131 // but it works well enough if seldom used.
3132 // All common fixed-size instructions are given a size
3133 // method by the AD file.
3134 // Note that the scratch buffer blob and locs memory are
3135 // allocated at the beginning of the compile task, and
3136 // may be shared by several calls to scratch_emit_size.
3137 // The allocation of the scratch buffer blob is particularly
3138 // expensive, since it has to grab the code cache lock.
3139 BufferBlob* blob = this->scratch_buffer_blob();
3140 assert(blob != nullptr, "Initialize BufferBlob at start");
3141 assert(blob->size() > MAX_inst_size, "sanity");
3142 relocInfo* locs_buf = scratch_locs_memory();
3143 address blob_begin = blob->content_begin();
3144 address blob_end = (address)locs_buf;
3145 assert(blob->contains(blob_end), "sanity");
3146 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3147 buf.initialize_consts_size(_scratch_const_size);
3148 buf.initialize_stubs_size(MAX_stubs_size);
3149 assert(locs_buf != nullptr, "sanity");
3150 int lsize = MAX_locs_size / 3;
3151 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3152 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3153 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3154 // Mark as scratch buffer.
3155 buf.consts()->set_scratch_emit();
3156 buf.insts()->set_scratch_emit();
3157 buf.stubs()->set_scratch_emit();
3158
3159 // Do the emission.
3160
3161 Label fakeL; // Fake label for branch instructions.
3162 Label* saveL = nullptr;
3163 uint save_bnum = 0;
3164 bool is_branch = n->is_MachBranch();
3165 C2_MacroAssembler masm(&buf);
3166 masm.bind(fakeL);
3167 if (is_branch) {
3168 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3169 n->as_MachBranch()->label_set(&fakeL, 0);
3170 }
3171 n->emit(&masm, C->regalloc());
3172
3173 // Emitting into the scratch buffer should not fail
3174 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3175
3176 if (is_branch) // Restore label.
3177 n->as_MachBranch()->label_set(saveL, save_bnum);
3178
3179 // End scratch_emit_size section.
3180 set_in_scratch_emit_size(false);
3181
3182 return buf.insts_size();
3183 }
3184
3185 void PhaseOutput::install() {
3186 if (!C->should_install_code()) {
3187 return;
3188 } else if (C->stub_function() != nullptr) {
3189 install_stub(C->stub_name());
3190 } else {
3191 install_code(C->method(),
3192 C->entry_bci(),
3193 CompileBroker::compiler2(),
3194 C->has_unsafe_access(),
3195 SharedRuntime::is_wide_vector(C->max_vector_size()));
3196 }
3197 }
3198
3199 void PhaseOutput::install_code(ciMethod* target,
3200 int entry_bci,
3201 AbstractCompiler* compiler,
3202 bool has_unsafe_access,
3203 bool has_wide_vectors) {
3204 // Check if we want to skip execution of all compiled code.
3205 {
3206 #ifndef PRODUCT
3207 if (OptoNoExecute) {
3208 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3209 return;
3210 }
3211 #endif
3212 Compile::TracePhase tp(_t_registerMethod);
3213
3214 if (C->is_osr_compilation()) {
3215 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3216 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3217 } else {
3218 if (!target->is_static()) {
3219 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3220 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3221 // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3222 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3223 }
3224 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3225 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3226 }
3227
3228 C->env()->register_method(target,
3229 entry_bci,
3230 &_code_offsets,
3231 _orig_pc_slot_offset_in_bytes,
3232 code_buffer(),
3233 frame_size_in_words(),
3234 oop_map_set(),
3235 &_handler_table,
3236 inc_table(),
3237 compiler,
3238 has_unsafe_access,
3239 SharedRuntime::is_wide_vector(C->max_vector_size()),
3240 C->has_monitors(),
3241 C->has_scoped_access(),
3242 0);
3243
3244 if (C->log() != nullptr) { // Print code cache state into compiler log
3245 C->log()->code_cache_state();
3246 }
3247 }
3248 }
3249 void PhaseOutput::install_stub(const char* stub_name) {
3250 // Entry point will be accessed using stub_entry_point();
3251 if (code_buffer() == nullptr) {
3252 Matcher::soft_match_failure();
3253 } else {
3254 if (PrintAssembly && (WizardMode || Verbose))
3255 tty->print_cr("### Stub::%s", stub_name);
3256
3257 if (!C->failing()) {
3258 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3259
3260 // Make the NMethod
3261 // For now we mark the frame as never safe for profile stackwalking
3262 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3263 code_buffer(),
3264 CodeOffsets::frame_never_safe,
3265 // _code_offsets.value(CodeOffsets::Frame_Complete),
3266 frame_size_in_words(),
3267 oop_map_set(),
3268 false,
3269 false);
3270
3271 if (rs == nullptr) {
3272 C->record_failure("CodeCache is full");
3273 } else {
3274 assert(rs->is_runtime_stub(), "sanity check");
3275 C->set_stub_entry_point(rs->entry_point());
3276 BlobId blob_id = StubInfo::blob(C->stub_id());
3277 AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, blob_id);
3278 }
3279 }
3280 }
3281 }
3282
3283 // Support for bundling info
3284 Bundle* PhaseOutput::node_bundling(const Node *n) {
3285 assert(valid_bundle_info(n), "oob");
3286 return &_node_bundling_base[n->_idx];
3287 }
3288
3289 bool PhaseOutput::valid_bundle_info(const Node *n) {
3290 return (_node_bundling_limit > n->_idx);
3291 }
3292
3293 //------------------------------frame_size_in_words-----------------------------
3294 // frame_slots in units of words
3295 int PhaseOutput::frame_size_in_words() const {
3296 // shift is 0 in LP32 and 1 in LP64
3297 const int shift = (LogBytesPerWord - LogBytesPerInt);
3298 int words = _frame_slots >> shift;
3299 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3300 return words;
3301 }
3302
3303 // To bang the stack of this compiled method we use the stack size
3304 // that the interpreter would need in case of a deoptimization. This
3305 // removes the need to bang the stack in the deoptimization blob which
3306 // in turn simplifies stack overflow handling.
3307 int PhaseOutput::bang_size_in_bytes() const {
3308 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3309 }
3310
3311 //------------------------------dump_asm---------------------------------------
3312 // Dump formatted assembly
3313 #if defined(SUPPORT_OPTO_ASSEMBLY)
3314 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3315
3316 int pc_digits = 3; // #chars required for pc
3317 int sb_chars = 3; // #chars for "start bundle" indicator
3318 int tab_size = 8;
3319 if (pcs != nullptr) {
3320 int max_pc = 0;
3321 for (uint i = 0; i < pc_limit; i++) {
3322 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3323 }
3324 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3325 }
3326 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3327
3328 bool cut_short = false;
3329 st->print_cr("#");
3330 st->print("# "); C->tf()->dump_on(st); st->cr();
3331 st->print_cr("#");
3332
3333 // For all blocks
3334 int pc = 0x0; // Program counter
3335 char starts_bundle = ' ';
3336 C->regalloc()->dump_frame();
3337
3338 Node *n = nullptr;
3339 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3340 if (VMThread::should_terminate()) {
3341 cut_short = true;
3342 break;
3343 }
3344 Block* block = C->cfg()->get_block(i);
3345 if (block->is_connector() && !Verbose) {
3346 continue;
3347 }
3348 n = block->head();
3349 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3350 pc = pcs[n->_idx];
3351 st->print("%*.*x", pc_digits, pc_digits, pc);
3352 }
3353 st->fill_to(prefix_len);
3354 block->dump_head(C->cfg(), st);
3355 if (block->is_connector()) {
3356 st->fill_to(prefix_len);
3357 st->print_cr("# Empty connector block");
3358 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3359 st->fill_to(prefix_len);
3360 st->print_cr("# Block is sole successor of call");
3361 }
3362
3363 // For all instructions
3364 for (uint j = 0; j < block->number_of_nodes(); j++) {
3365 if (VMThread::should_terminate()) {
3366 cut_short = true;
3367 break;
3368 }
3369 n = block->get_node(j);
3370 if (valid_bundle_info(n)) {
3371 Bundle* bundle = node_bundling(n);
3372 if (bundle->starts_bundle()) {
3373 starts_bundle = '+';
3374 }
3375 }
3376
3377 if (WizardMode) {
3378 n->dump();
3379 }
3380
3381 if( !n->is_Region() && // Dont print in the Assembly
3382 !n->is_Phi() && // a few noisely useless nodes
3383 !n->is_Proj() &&
3384 !n->is_MachTemp() &&
3385 !n->is_SafePointScalarObject() &&
3386 !n->is_Catch() && // Would be nice to print exception table targets
3387 !n->is_MergeMem() && // Not very interesting
3388 !n->is_top() && // Debug info table constants
3389 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3390 ) {
3391 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3392 pc = pcs[n->_idx];
3393 st->print("%*.*x", pc_digits, pc_digits, pc);
3394 } else {
3395 st->fill_to(pc_digits);
3396 }
3397 st->print(" %c ", starts_bundle);
3398 starts_bundle = ' ';
3399 st->fill_to(prefix_len);
3400 n->format(C->regalloc(), st);
3401 st->cr();
3402 }
3403
3404 // Dump the exception table as well
3405 if( n->is_Catch() && (Verbose || WizardMode) ) {
3406 // Print the exception table for this offset
3407 _handler_table.print_subtable_for(pc);
3408 }
3409 st->bol(); // Make sure we start on a new line
3410 }
3411 st->cr(); // one empty line between blocks
3412 } // End of per-block dump
3413
3414 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3415 }
3416 #endif
3417
3418 #ifndef PRODUCT
3419 void PhaseOutput::print_statistics() {
3420 Scheduling::print_statistics();
3421 }
3422 #endif