1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/debugInfo.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/compilerDirectives.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "gc/shared/barrierSet.hpp"
36 #include "gc/shared/c2/barrierSetC2.hpp"
37 #include "memory/allocation.inline.hpp"
38 #include "memory/allocation.hpp"
39 #include "opto/ad.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/c2_MacroAssembler.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/cfgnode.hpp"
45 #include "opto/locknode.hpp"
46 #include "opto/machnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/optoreg.hpp"
49 #include "opto/output.hpp"
50 #include "opto/regalloc.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/type.hpp"
54 #include "runtime/handles.inline.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #include "utilities/xmlstream.hpp"
59
60 #ifndef PRODUCT
61 #define DEBUG_ARG(x) , x
62 #else
63 #define DEBUG_ARG(x)
64 #endif
65
66 //------------------------------Scheduling----------------------------------
67 // This class contains all the information necessary to implement instruction
68 // scheduling and bundling.
69 class Scheduling {
70
71 private:
72 // Arena to use
73 Arena *_arena;
74
75 // Control-Flow Graph info
76 PhaseCFG *_cfg;
77
78 // Register Allocation info
79 PhaseRegAlloc *_regalloc;
80
81 // Number of nodes in the method
82 uint _node_bundling_limit;
83
84 // List of scheduled nodes. Generated in reverse order
85 Node_List _scheduled;
86
87 // List of nodes currently available for choosing for scheduling
88 Node_List _available;
89
90 // For each instruction beginning a bundle, the number of following
91 // nodes to be bundled with it.
92 Bundle *_node_bundling_base;
93
94 // Mapping from register to Node
95 Node_List _reg_node;
96
97 // Free list for pinch nodes.
98 Node_List _pinch_free_list;
99
100 // Number of uses of this node within the containing basic block.
101 short *_uses;
102
103 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
104 // front, branch+proj at end. Also skips Catch/CProj (same as
105 // branch-at-end), plus just-prior exception-throwing call.
106 uint _bb_start, _bb_end;
107
108 // Latency from the end of the basic block as scheduled
109 unsigned short *_current_latency;
110
111 // Remember the next node
112 Node *_next_node;
113
114 // Use this for an unconditional branch delay slot
115 Node *_unconditional_delay_slot;
116
117 // Pointer to a Nop
118 MachNopNode *_nop;
119
120 // Length of the current bundle, in instructions
121 uint _bundle_instr_count;
122
123 // Current Cycle number, for computing latencies and bundling
124 uint _bundle_cycle_number;
125
126 // Bundle information
127 Pipeline_Use_Element _bundle_use_elements[resource_count];
128 Pipeline_Use _bundle_use;
129
130 // Dump the available list
131 void dump_available() const;
132
133 public:
134 Scheduling(Arena *arena, Compile &compile);
135
136 // Destructor
137 NOT_PRODUCT( ~Scheduling(); )
138
139 // Step ahead "i" cycles
140 void step(uint i);
141
142 // Step ahead 1 cycle, and clear the bundle state (for example,
143 // at a branch target)
144 void step_and_clear();
145
146 Bundle* node_bundling(const Node *n) {
147 assert(valid_bundle_info(n), "oob");
148 return (&_node_bundling_base[n->_idx]);
149 }
150
151 bool valid_bundle_info(const Node *n) const {
152 return (_node_bundling_limit > n->_idx);
153 }
154
155 bool starts_bundle(const Node *n) const {
156 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
157 }
158
159 // Do the scheduling
160 void DoScheduling();
161
162 // Compute the register antidependencies within a basic block
163 void ComputeRegisterAntidependencies(Block *bb);
164 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
165 void verify_good_schedule( Block *b, const char *msg );
166 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
167 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
168
169 // Add a node to the current bundle
170 void AddNodeToBundle(Node *n, const Block *bb);
171
172 // Return an integer less than, equal to, or greater than zero
173 // if the stack offset of the first argument is respectively
174 // less than, equal to, or greater than the second.
175 int compare_two_spill_nodes(Node* first, Node* second);
176
177 // Add a node to the list of available nodes
178 void AddNodeToAvailableList(Node *n);
179
180 // Compute the local use count for the nodes in a block, and compute
181 // the list of instructions with no uses in the block as available
182 void ComputeUseCount(const Block *bb);
183
184 // Choose an instruction from the available list to add to the bundle
185 Node * ChooseNodeToBundle();
186
187 // See if this Node fits into the currently accumulating bundle
188 bool NodeFitsInBundle(Node *n);
189
190 // Decrement the use count for a node
191 void DecrementUseCounts(Node *n, const Block *bb);
192
193 // Garbage collect pinch nodes for reuse by other blocks.
194 void garbage_collect_pinch_nodes();
195 // Clean up a pinch node for reuse (helper for above).
196 void cleanup_pinch( Node *pinch );
197
198 // Information for statistics gathering
199 #ifndef PRODUCT
200 private:
201 // Gather information on size of nops relative to total
202 uint _branches, _unconditional_delays;
203
204 static uint _total_nop_size, _total_method_size;
205 static uint _total_branches, _total_unconditional_delays;
206 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
207
208 public:
209 static void print_statistics();
210
211 static void increment_instructions_per_bundle(uint i) {
212 _total_instructions_per_bundle[i]++;
213 }
214
215 static void increment_nop_size(uint s) {
216 _total_nop_size += s;
217 }
218
219 static void increment_method_size(uint s) {
220 _total_method_size += s;
221 }
222 #endif
223
224 };
225
226 PhaseOutput::PhaseOutput()
227 : Phase(Phase::Output),
228 _code_buffer("Compile::Fill_buffer"),
229 _first_block_size(0),
230 _handler_table(),
231 _inc_table(),
232 _stub_list(),
233 _oop_map_set(nullptr),
234 _scratch_buffer_blob(nullptr),
235 _scratch_locs_memory(nullptr),
236 _scratch_const_size(-1),
237 _in_scratch_emit_size(false),
238 _frame_slots(0),
239 _code_offsets(),
240 _node_bundling_limit(0),
241 _node_bundling_base(nullptr),
242 _orig_pc_slot(0),
243 _orig_pc_slot_offset_in_bytes(0),
244 _buf_sizes(),
245 _block(nullptr),
246 _index(0) {
247 C->set_output(this);
248 if (C->stub_name() == nullptr) {
249 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
250 }
251 }
252
253 PhaseOutput::~PhaseOutput() {
254 C->set_output(nullptr);
255 if (_scratch_buffer_blob != nullptr) {
256 BufferBlob::free(_scratch_buffer_blob);
257 }
258 }
259
260 void PhaseOutput::perform_mach_node_analysis() {
261 // Late barrier analysis must be done after schedule and bundle
262 // Otherwise liveness based spilling will fail
263 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
264 bs->late_barrier_analysis();
265
266 pd_perform_mach_node_analysis();
267
268 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
269 }
270
271 // Convert Nodes to instruction bits and pass off to the VM
272 void PhaseOutput::Output() {
273 // RootNode goes
274 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
275
276 // The number of new nodes (mostly MachNop) is proportional to
277 // the number of java calls and inner loops which are aligned.
278 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
279 C->inner_loops()*(OptoLoopAlignment-1)),
280 "out of nodes before code generation" ) ) {
281 return;
282 }
283 // Make sure I can find the Start Node
284 Block *entry = C->cfg()->get_block(1);
285 Block *broot = C->cfg()->get_root_block();
286
287 const StartNode *start = entry->head()->as_Start();
288
289 // Replace StartNode with prolog
290 MachPrologNode *prolog = new MachPrologNode();
291 entry->map_node(prolog, 0);
292 C->cfg()->map_node_to_block(prolog, entry);
293 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
294
295 // Virtual methods need an unverified entry point
296
297 if( C->is_osr_compilation() ) {
298 if( PoisonOSREntry ) {
299 // TODO: Should use a ShouldNotReachHereNode...
300 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
301 }
302 } else {
303 if( C->method() && !C->method()->flags().is_static() ) {
304 // Insert unvalidated entry point
305 C->cfg()->insert( broot, 0, new MachUEPNode() );
306 }
307
308 }
309
310 // Break before main entry point
311 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
312 (OptoBreakpoint && C->is_method_compilation()) ||
313 (OptoBreakpointOSR && C->is_osr_compilation()) ||
314 (OptoBreakpointC2R && !C->method()) ) {
315 // checking for C->method() means that OptoBreakpoint does not apply to
316 // runtime stubs or frame converters
317 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
318 }
319
320 // Insert epilogs before every return
321 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
322 Block* block = C->cfg()->get_block(i);
323 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
324 Node* m = block->end();
325 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
326 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
327 block->add_inst(epilog);
328 C->cfg()->map_node_to_block(epilog, block);
329 }
330 }
331 }
332
333 // Keeper of sizing aspects
334 _buf_sizes = BufferSizingData();
335
336 // Initialize code buffer
337 estimate_buffer_size(_buf_sizes._const);
338 if (C->failing()) return;
339
340 // Pre-compute the length of blocks and replace
341 // long branches with short if machine supports it.
342 // Must be done before ScheduleAndBundle due to SPARC delay slots
343 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
344 blk_starts[0] = 0;
345 shorten_branches(blk_starts);
346
347 ScheduleAndBundle();
348 if (C->failing()) {
349 return;
350 }
351
352 perform_mach_node_analysis();
353
354 // Complete sizing of codebuffer
355 CodeBuffer* cb = init_buffer();
356 if (cb == nullptr || C->failing()) {
357 return;
358 }
359
360 BuildOopMaps();
361
362 if (C->failing()) {
363 return;
364 }
365
366 C2_MacroAssembler masm(cb);
367 fill_buffer(&masm, blk_starts);
368 }
369
370 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
371 // Determine if we need to generate a stack overflow check.
372 // Do it if the method is not a stub function and
373 // has java calls or has frame size > vm_page_size/8.
374 // The debug VM checks that deoptimization doesn't trigger an
375 // unexpected stack overflow (compiled method stack banging should
376 // guarantee it doesn't happen) so we always need the stack bang in
377 // a debug VM.
378 return (C->stub_function() == nullptr &&
379 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
380 DEBUG_ONLY(|| true)));
381 }
382
383 bool PhaseOutput::need_register_stack_bang() const {
384 // Determine if we need to generate a register stack overflow check.
385 // This is only used on architectures which have split register
386 // and memory stacks (ie. IA64).
387 // Bang if the method is not a stub function and has java calls
388 return (C->stub_function() == nullptr && C->has_java_calls());
389 }
390
391
392 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
393 // of a loop. When aligning a loop we need to provide enough instructions
394 // in cpu's fetch buffer to feed decoders. The loop alignment could be
395 // avoided if we have enough instructions in fetch buffer at the head of a loop.
396 // By default, the size is set to 999999 by Block's constructor so that
397 // a loop will be aligned if the size is not reset here.
398 //
399 // Note: Mach instructions could contain several HW instructions
400 // so the size is estimated only.
401 //
402 void PhaseOutput::compute_loop_first_inst_sizes() {
403 // The next condition is used to gate the loop alignment optimization.
404 // Don't aligned a loop if there are enough instructions at the head of a loop
405 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
406 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
407 // equal to 11 bytes which is the largest address NOP instruction.
408 if (MaxLoopPad < OptoLoopAlignment - 1) {
409 uint last_block = C->cfg()->number_of_blocks() - 1;
410 for (uint i = 1; i <= last_block; i++) {
411 Block* block = C->cfg()->get_block(i);
412 // Check the first loop's block which requires an alignment.
413 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
414 uint sum_size = 0;
415 uint inst_cnt = NumberOfLoopInstrToAlign;
416 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
417
418 // Check subsequent fallthrough blocks if the loop's first
419 // block(s) does not have enough instructions.
420 Block *nb = block;
421 while(inst_cnt > 0 &&
422 i < last_block &&
423 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
424 !nb->has_successor(block)) {
425 i++;
426 nb = C->cfg()->get_block(i);
427 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
428 } // while( inst_cnt > 0 && i < last_block )
429
430 block->set_first_inst_size(sum_size);
431 } // f( b->head()->is_Loop() )
432 } // for( i <= last_block )
433 } // if( MaxLoopPad < OptoLoopAlignment-1 )
434 }
435
436 // The architecture description provides short branch variants for some long
437 // branch instructions. Replace eligible long branches with short branches.
438 void PhaseOutput::shorten_branches(uint* blk_starts) {
439
440 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
441
442 // Compute size of each block, method size, and relocation information size
443 uint nblocks = C->cfg()->number_of_blocks();
444
445 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
446 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
447 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
448
449 // Collect worst case block paddings
450 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
451 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
452
453 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
454 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
455
456 bool has_short_branch_candidate = false;
457
458 // Initialize the sizes to 0
459 int code_size = 0; // Size in bytes of generated code
460 int stub_size = 0; // Size in bytes of all stub entries
461 // Size in bytes of all relocation entries, including those in local stubs.
462 // Start with 2-bytes of reloc info for the unvalidated entry point
463 int reloc_size = 1; // Number of relocation entries
464
465 // Make three passes. The first computes pessimistic blk_starts,
466 // relative jmp_offset and reloc_size information. The second performs
467 // short branch substitution using the pessimistic sizing. The
468 // third inserts nops where needed.
469
470 // Step one, perform a pessimistic sizing pass.
471 uint last_call_adr = max_juint;
472 uint last_avoid_back_to_back_adr = max_juint;
473 uint nop_size = (new MachNopNode())->size(C->regalloc());
474 for (uint i = 0; i < nblocks; i++) { // For all blocks
475 Block* block = C->cfg()->get_block(i);
476 _block = block;
477
478 // During short branch replacement, we store the relative (to blk_starts)
479 // offset of jump in jmp_offset, rather than the absolute offset of jump.
480 // This is so that we do not need to recompute sizes of all nodes when
481 // we compute correct blk_starts in our next sizing pass.
482 jmp_offset[i] = 0;
483 jmp_size[i] = 0;
484 jmp_nidx[i] = -1;
485 DEBUG_ONLY( jmp_target[i] = 0; )
486 DEBUG_ONLY( jmp_rule[i] = 0; )
487
488 // Sum all instruction sizes to compute block size
489 uint last_inst = block->number_of_nodes();
490 uint blk_size = 0;
491 for (uint j = 0; j < last_inst; j++) {
492 _index = j;
493 Node* nj = block->get_node(_index);
494 // Handle machine instruction nodes
495 if (nj->is_Mach()) {
496 MachNode* mach = nj->as_Mach();
497 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
498 reloc_size += mach->reloc();
499 if (mach->is_MachCall()) {
500 // add size information for trampoline stub
501 // class CallStubImpl is platform-specific and defined in the *.ad files.
502 stub_size += CallStubImpl::size_call_trampoline();
503 reloc_size += CallStubImpl::reloc_call_trampoline();
504
505 MachCallNode *mcall = mach->as_MachCall();
506 // This destination address is NOT PC-relative
507
508 mcall->method_set((intptr_t)mcall->entry_point());
509
510 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
511 stub_size += CompiledDirectCall::to_interp_stub_size();
512 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
513 }
514 } else if (mach->is_MachSafePoint()) {
515 // If call/safepoint are adjacent, account for possible
516 // nop to disambiguate the two safepoints.
517 // ScheduleAndBundle() can rearrange nodes in a block,
518 // check for all offsets inside this block.
519 if (last_call_adr >= blk_starts[i]) {
520 blk_size += nop_size;
521 }
522 }
523 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
524 // Nop is inserted between "avoid back to back" instructions.
525 // ScheduleAndBundle() can rearrange nodes in a block,
526 // check for all offsets inside this block.
527 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
528 blk_size += nop_size;
529 }
530 }
531 if (mach->may_be_short_branch()) {
532 if (!nj->is_MachBranch()) {
533 #ifndef PRODUCT
534 nj->dump(3);
535 #endif
536 Unimplemented();
537 }
538 assert(jmp_nidx[i] == -1, "block should have only one branch");
539 jmp_offset[i] = blk_size;
540 jmp_size[i] = nj->size(C->regalloc());
541 jmp_nidx[i] = j;
542 has_short_branch_candidate = true;
543 }
544 }
545 blk_size += nj->size(C->regalloc());
546 // Remember end of call offset
547 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
548 last_call_adr = blk_starts[i]+blk_size;
549 }
550 // Remember end of avoid_back_to_back offset
551 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
552 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
553 }
554 }
555
556 // When the next block starts a loop, we may insert pad NOP
557 // instructions. Since we cannot know our future alignment,
558 // assume the worst.
559 if (i < nblocks - 1) {
560 Block* nb = C->cfg()->get_block(i + 1);
561 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
562 if (max_loop_pad > 0) {
563 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
564 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
565 // If either is the last instruction in this block, bump by
566 // max_loop_pad in lock-step with blk_size, so sizing
567 // calculations in subsequent blocks still can conservatively
568 // detect that it may the last instruction in this block.
569 if (last_call_adr == blk_starts[i]+blk_size) {
570 last_call_adr += max_loop_pad;
571 }
572 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
573 last_avoid_back_to_back_adr += max_loop_pad;
574 }
575 blk_size += max_loop_pad;
576 block_worst_case_pad[i + 1] = max_loop_pad;
577 }
578 }
579
580 // Save block size; update total method size
581 blk_starts[i+1] = blk_starts[i]+blk_size;
582 }
583
584 // Step two, replace eligible long jumps.
585 bool progress = true;
586 uint last_may_be_short_branch_adr = max_juint;
587 while (has_short_branch_candidate && progress) {
588 progress = false;
589 has_short_branch_candidate = false;
590 int adjust_block_start = 0;
591 for (uint i = 0; i < nblocks; i++) {
592 Block* block = C->cfg()->get_block(i);
593 int idx = jmp_nidx[i];
594 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
595 if (mach != nullptr && mach->may_be_short_branch()) {
596 #ifdef ASSERT
597 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
598 int j;
599 // Find the branch; ignore trailing NOPs.
600 for (j = block->number_of_nodes()-1; j>=0; j--) {
601 Node* n = block->get_node(j);
602 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
603 break;
604 }
605 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
606 #endif
607 int br_size = jmp_size[i];
608 int br_offs = blk_starts[i] + jmp_offset[i];
609
610 // This requires the TRUE branch target be in succs[0]
611 uint bnum = block->non_connector_successor(0)->_pre_order;
612 int offset = blk_starts[bnum] - br_offs;
613 if (bnum > i) { // adjust following block's offset
614 offset -= adjust_block_start;
615 }
616
617 // This block can be a loop header, account for the padding
618 // in the previous block.
619 int block_padding = block_worst_case_pad[i];
620 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
621 // In the following code a nop could be inserted before
622 // the branch which will increase the backward distance.
623 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
624 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
625
626 if (needs_padding && offset <= 0)
627 offset -= nop_size;
628
629 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
630 // We've got a winner. Replace this branch.
631 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
632
633 // Update the jmp_size.
634 int new_size = replacement->size(C->regalloc());
635 int diff = br_size - new_size;
636 assert(diff >= (int)nop_size, "short_branch size should be smaller");
637 // Conservatively take into account padding between
638 // avoid_back_to_back branches. Previous branch could be
639 // converted into avoid_back_to_back branch during next
640 // rounds.
641 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
642 jmp_offset[i] += nop_size;
643 diff -= nop_size;
644 }
645 adjust_block_start += diff;
646 block->map_node(replacement, idx);
647 mach->subsume_by(replacement, C);
648 mach = replacement;
649 progress = true;
650
651 jmp_size[i] = new_size;
652 DEBUG_ONLY( jmp_target[i] = bnum; );
653 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
654 } else {
655 // The jump distance is not short, try again during next iteration.
656 has_short_branch_candidate = true;
657 }
658 } // (mach->may_be_short_branch())
659 if (mach != nullptr && (mach->may_be_short_branch() ||
660 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
661 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
662 }
663 blk_starts[i+1] -= adjust_block_start;
664 }
665 }
666
667 #ifdef ASSERT
668 for (uint i = 0; i < nblocks; i++) { // For all blocks
669 if (jmp_target[i] != 0) {
670 int br_size = jmp_size[i];
671 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
672 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
673 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
674 }
675 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
676 }
677 }
678 #endif
679
680 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
681 // after ScheduleAndBundle().
682
683 // ------------------
684 // Compute size for code buffer
685 code_size = blk_starts[nblocks];
686
687 // Relocation records
688 reloc_size += 1; // Relo entry for exception handler
689
690 // Adjust reloc_size to number of record of relocation info
691 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
692 // a relocation index.
693 // The CodeBuffer will expand the locs array if this estimate is too low.
694 reloc_size *= 10 / sizeof(relocInfo);
695
696 _buf_sizes._reloc = reloc_size;
697 _buf_sizes._code = code_size;
698 _buf_sizes._stub = stub_size;
699 }
700
701 //------------------------------FillLocArray-----------------------------------
702 // Create a bit of debug info and append it to the array. The mapping is from
703 // Java local or expression stack to constant, register or stack-slot. For
704 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
705 // entry has been taken care of and caller should skip it).
706 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
707 // This should never have accepted Bad before
708 assert(OptoReg::is_valid(regnum), "location must be valid");
709 return (OptoReg::is_reg(regnum))
710 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
711 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
712 }
713
714
715 ObjectValue*
716 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
717 for (int i = 0; i < objs->length(); i++) {
718 assert(objs->at(i)->is_object(), "corrupt object cache");
719 ObjectValue* sv = objs->at(i)->as_ObjectValue();
720 if (sv->id() == id) {
721 return sv;
722 }
723 }
724 // Otherwise..
725 return nullptr;
726 }
727
728 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
729 ObjectValue* sv ) {
730 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
731 objs->append(sv);
732 }
733
734
735 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
736 GrowableArray<ScopeValue*> *array,
737 GrowableArray<ScopeValue*> *objs ) {
738 assert( local, "use _top instead of null" );
739 if (array->length() != idx) {
740 assert(array->length() == idx + 1, "Unexpected array count");
741 // Old functionality:
742 // return
743 // New functionality:
744 // Assert if the local is not top. In product mode let the new node
745 // override the old entry.
746 assert(local == C->top(), "LocArray collision");
747 if (local == C->top()) {
748 return;
749 }
750 array->pop();
751 }
752 const Type *t = local->bottom_type();
753
754 // Is it a safepoint scalar object node?
755 if (local->is_SafePointScalarObject()) {
756 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
757
758 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
759 if (sv == nullptr) {
760 ciKlass* cik = t->is_oopptr()->exact_klass();
761 assert(cik->is_instance_klass() ||
762 cik->is_array_klass(), "Not supported allocation.");
763 sv = new ObjectValue(spobj->_idx,
764 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
765 set_sv_for_object_node(objs, sv);
766
767 uint first_ind = spobj->first_index(sfpt->jvms());
768 for (uint i = 0; i < spobj->n_fields(); i++) {
769 Node* fld_node = sfpt->in(first_ind+i);
770 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
771 }
772 }
773 array->append(sv);
774 return;
775 } else if (local->is_SafePointScalarMerge()) {
776 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
777 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
778
779 if (mv == nullptr) {
780 GrowableArray<ScopeValue*> deps;
781
782 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
783 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
784 assert(deps.length() == 1, "missing value");
785
786 int selector_idx = smerge->selector_idx(sfpt->jvms());
787 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
788 assert(deps.length() == 2, "missing value");
789
790 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
791 set_sv_for_object_node(objs, mv);
792
793 for (uint i = 1; i < smerge->req(); i++) {
794 Node* obj_node = smerge->in(i);
795 int idx = mv->possible_objects()->length();
796 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
797
798 // By default ObjectValues that are in 'possible_objects' are not root objects.
799 // They will be marked as root later if they are directly referenced in a JVMS.
800 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
801 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
802 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
803 }
804 }
805 array->append(mv);
806 return;
807 }
808
809 // Grab the register number for the local
810 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
811 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
812 // Record the double as two float registers.
813 // The register mask for such a value always specifies two adjacent
814 // float registers, with the lower register number even.
815 // Normally, the allocation of high and low words to these registers
816 // is irrelevant, because nearly all operations on register pairs
817 // (e.g., StoreD) treat them as a single unit.
818 // Here, we assume in addition that the words in these two registers
819 // stored "naturally" (by operations like StoreD and double stores
820 // within the interpreter) such that the lower-numbered register
821 // is written to the lower memory address. This may seem like
822 // a machine dependency, but it is not--it is a requirement on
823 // the author of the <arch>.ad file to ensure that, for every
824 // even/odd double-register pair to which a double may be allocated,
825 // the word in the even single-register is stored to the first
826 // memory word. (Note that register numbers are completely
827 // arbitrary, and are not tied to any machine-level encodings.)
828 #ifdef _LP64
829 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
830 array->append(new ConstantIntValue((jint)0));
831 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
832 } else if ( t->base() == Type::Long ) {
833 array->append(new ConstantIntValue((jint)0));
834 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
835 } else if ( t->base() == Type::RawPtr ) {
836 // jsr/ret return address which must be restored into the full
837 // width 64-bit stack slot.
838 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
839 }
840 #else //_LP64
841 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
842 // Repack the double/long as two jints.
843 // The convention the interpreter uses is that the second local
844 // holds the first raw word of the native double representation.
845 // This is actually reasonable, since locals and stack arrays
846 // grow downwards in all implementations.
847 // (If, on some machine, the interpreter's Java locals or stack
848 // were to grow upwards, the embedded doubles would be word-swapped.)
849 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
850 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
851 }
852 #endif //_LP64
853 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
854 OptoReg::is_reg(regnum) ) {
855 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
856 ? Location::float_in_dbl : Location::normal ));
857 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
858 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
859 ? Location::int_in_long : Location::normal ));
860 } else if( t->base() == Type::NarrowOop ) {
861 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
862 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
863 t->base() == Type::VectorD || t->base() == Type::VectorX ||
864 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
865 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
866 } else if (C->regalloc()->is_oop(local)) {
867 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
868 t->base() == Type::AryPtr,
869 "Unexpected type: %s", t->msg());
870 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
871 } else {
872 assert(t->base() == Type::Int || t->base() == Type::Half ||
873 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
874 "Unexpected type: %s", t->msg());
875 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
876 }
877 return;
878 }
879
880 // No register. It must be constant data.
881 switch (t->base()) {
882 case Type::Half: // Second half of a double
883 ShouldNotReachHere(); // Caller should skip 2nd halves
884 break;
885 case Type::AnyPtr:
886 array->append(new ConstantOopWriteValue(nullptr));
887 break;
888 case Type::AryPtr:
889 case Type::InstPtr: // fall through
890 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
891 break;
892 case Type::NarrowOop:
893 if (t == TypeNarrowOop::NULL_PTR) {
894 array->append(new ConstantOopWriteValue(nullptr));
895 } else {
896 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
897 }
898 break;
899 case Type::Int:
900 array->append(new ConstantIntValue(t->is_int()->get_con()));
901 break;
902 case Type::RawPtr:
903 // A return address (T_ADDRESS).
904 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
905 #ifdef _LP64
906 // Must be restored to the full-width 64-bit stack slot.
907 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
908 #else
909 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
910 #endif
911 break;
912 case Type::FloatCon: {
913 float f = t->is_float_constant()->getf();
914 array->append(new ConstantIntValue(jint_cast(f)));
915 break;
916 }
917 case Type::DoubleCon: {
918 jdouble d = t->is_double_constant()->getd();
919 #ifdef _LP64
920 array->append(new ConstantIntValue((jint)0));
921 array->append(new ConstantDoubleValue(d));
922 #else
923 // Repack the double as two jints.
924 // The convention the interpreter uses is that the second local
925 // holds the first raw word of the native double representation.
926 // This is actually reasonable, since locals and stack arrays
927 // grow downwards in all implementations.
928 // (If, on some machine, the interpreter's Java locals or stack
929 // were to grow upwards, the embedded doubles would be word-swapped.)
930 jlong_accessor acc;
931 acc.long_value = jlong_cast(d);
932 array->append(new ConstantIntValue(acc.words[1]));
933 array->append(new ConstantIntValue(acc.words[0]));
934 #endif
935 break;
936 }
937 case Type::Long: {
938 jlong d = t->is_long()->get_con();
939 #ifdef _LP64
940 array->append(new ConstantIntValue((jint)0));
941 array->append(new ConstantLongValue(d));
942 #else
943 // Repack the long as two jints.
944 // The convention the interpreter uses is that the second local
945 // holds the first raw word of the native double representation.
946 // This is actually reasonable, since locals and stack arrays
947 // grow downwards in all implementations.
948 // (If, on some machine, the interpreter's Java locals or stack
949 // were to grow upwards, the embedded doubles would be word-swapped.)
950 jlong_accessor acc;
951 acc.long_value = d;
952 array->append(new ConstantIntValue(acc.words[1]));
953 array->append(new ConstantIntValue(acc.words[0]));
954 #endif
955 break;
956 }
957 case Type::Top: // Add an illegal value here
958 array->append(new LocationValue(Location()));
959 break;
960 default:
961 ShouldNotReachHere();
962 break;
963 }
964 }
965
966 // Determine if this node starts a bundle
967 bool PhaseOutput::starts_bundle(const Node *n) const {
968 return (_node_bundling_limit > n->_idx &&
969 _node_bundling_base[n->_idx].starts_bundle());
970 }
971
972 // Determine if there is a monitor that has 'ov' as its owner.
973 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
974 for (int k = 0; k < monarray->length(); k++) {
975 MonitorValue* mv = monarray->at(k);
976 if (mv->owner() == ov) {
977 return true;
978 }
979 }
980
981 return false;
982 }
983
984 // Determine if there is a scalar replaced object description represented by 'ov'.
985 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
986 GrowableArray<ScopeValue*>* objs,
987 ObjectValue* ov) const {
988 for (int i = 0; i < jvms->scl_size(); i++) {
989 Node* n = sfn->scalarized_obj(jvms, i);
990 // Other kinds of nodes that we may encounter here, for instance constants
991 // representing values of fields of objects scalarized, aren't relevant for
992 // us, since they don't map to ObjectValue.
993 if (!n->is_SafePointScalarObject()) {
994 continue;
995 }
996
997 ObjectValue* other = sv_for_node_id(objs, n->_idx);
998 if (ov == other) {
999 return true;
1000 }
1001 }
1002 return false;
1003 }
1004
1005 //--------------------------Process_OopMap_Node--------------------------------
1006 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1007 // Handle special safepoint nodes for synchronization
1008 MachSafePointNode *sfn = mach->as_MachSafePoint();
1009 MachCallNode *mcall;
1010
1011 int safepoint_pc_offset = current_offset;
1012 bool is_method_handle_invoke = false;
1013 bool return_oop = false;
1014 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1015 bool arg_escape = false;
1016
1017 // Add the safepoint in the DebugInfoRecorder
1018 if( !mach->is_MachCall() ) {
1019 mcall = nullptr;
1020 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1021 } else {
1022 mcall = mach->as_MachCall();
1023
1024 // Is the call a MethodHandle call?
1025 if (mcall->is_MachCallJava()) {
1026 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1027 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1028 is_method_handle_invoke = true;
1029 }
1030 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1031 }
1032
1033 // Check if a call returns an object.
1034 if (mcall->returns_pointer()) {
1035 return_oop = true;
1036 }
1037 safepoint_pc_offset += mcall->ret_addr_offset();
1038 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1039 }
1040
1041 // Loop over the JVMState list to add scope information
1042 // Do not skip safepoints with a null method, they need monitor info
1043 JVMState* youngest_jvms = sfn->jvms();
1044 int max_depth = youngest_jvms->depth();
1045
1046 // Allocate the object pool for scalar-replaced objects -- the map from
1047 // small-integer keys (which can be recorded in the local and ostack
1048 // arrays) to descriptions of the object state.
1049 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1050
1051 // Visit scopes from oldest to youngest.
1052 for (int depth = 1; depth <= max_depth; depth++) {
1053 JVMState* jvms = youngest_jvms->of_depth(depth);
1054 int idx;
1055 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1056 // Safepoints that do not have method() set only provide oop-map and monitor info
1057 // to support GC; these do not support deoptimization.
1058 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1059 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1060 int num_mon = jvms->nof_monitors();
1061 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1062 "JVMS local count must match that of the method");
1063
1064 // Add Local and Expression Stack Information
1065
1066 // Insert locals into the locarray
1067 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1068 for( idx = 0; idx < num_locs; idx++ ) {
1069 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1070 }
1071
1072 // Insert expression stack entries into the exparray
1073 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1074 for( idx = 0; idx < num_exps; idx++ ) {
1075 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1076 }
1077
1078 // Add in mappings of the monitors
1079 assert( !method ||
1080 !method->is_synchronized() ||
1081 method->is_native() ||
1082 num_mon > 0 ||
1083 !GenerateSynchronizationCode,
1084 "monitors must always exist for synchronized methods");
1085
1086 // Build the growable array of ScopeValues for exp stack
1087 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1088
1089 // Loop over monitors and insert into array
1090 for (idx = 0; idx < num_mon; idx++) {
1091 // Grab the node that defines this monitor
1092 Node* box_node = sfn->monitor_box(jvms, idx);
1093 Node* obj_node = sfn->monitor_obj(jvms, idx);
1094
1095 // Create ScopeValue for object
1096 ScopeValue *scval = nullptr;
1097
1098 if (obj_node->is_SafePointScalarObject()) {
1099 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1100 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1101 if (scval == nullptr) {
1102 const Type *t = spobj->bottom_type();
1103 ciKlass* cik = t->is_oopptr()->exact_klass();
1104 assert(cik->is_instance_klass() ||
1105 cik->is_array_klass(), "Not supported allocation.");
1106 ObjectValue* sv = new ObjectValue(spobj->_idx,
1107 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1108 PhaseOutput::set_sv_for_object_node(objs, sv);
1109
1110 uint first_ind = spobj->first_index(youngest_jvms);
1111 for (uint i = 0; i < spobj->n_fields(); i++) {
1112 Node* fld_node = sfn->in(first_ind+i);
1113 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1114 }
1115 scval = sv;
1116 }
1117 } else if (obj_node->is_SafePointScalarMerge()) {
1118 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1119 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1120
1121 if (mv == nullptr) {
1122 GrowableArray<ScopeValue*> deps;
1123
1124 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1125 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1126 assert(deps.length() == 1, "missing value");
1127
1128 int selector_idx = smerge->selector_idx(youngest_jvms);
1129 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1130 assert(deps.length() == 2, "missing value");
1131
1132 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1133 set_sv_for_object_node(objs, mv);
1134
1135 for (uint i = 1; i < smerge->req(); i++) {
1136 Node* obj_node = smerge->in(i);
1137 int idx = mv->possible_objects()->length();
1138 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1139
1140 // By default ObjectValues that are in 'possible_objects' are not root objects.
1141 // They will be marked as root later if they are directly referenced in a JVMS.
1142 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1143 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1144 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1145 }
1146 }
1147 scval = mv;
1148 } else if (!obj_node->is_Con()) {
1149 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1150 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1151 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1152 } else {
1153 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1154 }
1155 } else {
1156 const TypePtr *tp = obj_node->get_ptr_type();
1157 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1158 }
1159
1160 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1161 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1162 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1163 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1164 }
1165
1166 // Mark ObjectValue nodes as root nodes if they are directly
1167 // referenced in the JVMS.
1168 for (int i = 0; i < objs->length(); i++) {
1169 ScopeValue* sv = objs->at(i);
1170 if (sv->is_object_merge()) {
1171 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1172
1173 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1174 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1175 if (ov->is_root()) {
1176 // Already flagged as 'root' by something else. We shouldn't change it
1177 // to non-root in a younger JVMS because it may need to be alive in
1178 // a younger JVMS.
1179 } else {
1180 bool is_root = locarray->contains(ov) ||
1181 exparray->contains(ov) ||
1182 contains_as_owner(monarray, ov) ||
1183 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1184 ov->set_root(is_root);
1185 }
1186 }
1187 }
1188 }
1189
1190 // We dump the object pool first, since deoptimization reads it in first.
1191 C->debug_info()->dump_object_pool(objs);
1192
1193 // Build first class objects to pass to scope
1194 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1195 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1196 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1197
1198 // Make method available for all Safepoints
1199 ciMethod* scope_method = method ? method : C->method();
1200 // Describe the scope here
1201 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1202 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1203 // Now we can describe the scope.
1204 methodHandle null_mh;
1205 bool rethrow_exception = false;
1206 C->debug_info()->describe_scope(
1207 safepoint_pc_offset,
1208 null_mh,
1209 scope_method,
1210 jvms->bci(),
1211 jvms->should_reexecute(),
1212 rethrow_exception,
1213 is_method_handle_invoke,
1214 return_oop,
1215 has_ea_local_in_scope,
1216 arg_escape,
1217 locvals,
1218 expvals,
1219 monvals
1220 );
1221 } // End jvms loop
1222
1223 // Mark the end of the scope set.
1224 C->debug_info()->end_safepoint(safepoint_pc_offset);
1225 }
1226
1227
1228
1229 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1230 class NonSafepointEmitter {
1231 Compile* C;
1232 JVMState* _pending_jvms;
1233 int _pending_offset;
1234
1235 void emit_non_safepoint();
1236
1237 public:
1238 NonSafepointEmitter(Compile* compile) {
1239 this->C = compile;
1240 _pending_jvms = nullptr;
1241 _pending_offset = 0;
1242 }
1243
1244 void observe_instruction(Node* n, int pc_offset) {
1245 if (!C->debug_info()->recording_non_safepoints()) return;
1246
1247 Node_Notes* nn = C->node_notes_at(n->_idx);
1248 if (nn == nullptr || nn->jvms() == nullptr) return;
1249 if (_pending_jvms != nullptr &&
1250 _pending_jvms->same_calls_as(nn->jvms())) {
1251 // Repeated JVMS? Stretch it up here.
1252 _pending_offset = pc_offset;
1253 } else {
1254 if (_pending_jvms != nullptr &&
1255 _pending_offset < pc_offset) {
1256 emit_non_safepoint();
1257 }
1258 _pending_jvms = nullptr;
1259 if (pc_offset > C->debug_info()->last_pc_offset()) {
1260 // This is the only way _pending_jvms can become non-null:
1261 _pending_jvms = nn->jvms();
1262 _pending_offset = pc_offset;
1263 }
1264 }
1265 }
1266
1267 // Stay out of the way of real safepoints:
1268 void observe_safepoint(JVMState* jvms, int pc_offset) {
1269 if (_pending_jvms != nullptr &&
1270 !_pending_jvms->same_calls_as(jvms) &&
1271 _pending_offset < pc_offset) {
1272 emit_non_safepoint();
1273 }
1274 _pending_jvms = nullptr;
1275 }
1276
1277 void flush_at_end() {
1278 if (_pending_jvms != nullptr) {
1279 emit_non_safepoint();
1280 }
1281 _pending_jvms = nullptr;
1282 }
1283 };
1284
1285 void NonSafepointEmitter::emit_non_safepoint() {
1286 JVMState* youngest_jvms = _pending_jvms;
1287 int pc_offset = _pending_offset;
1288
1289 // Clear it now:
1290 _pending_jvms = nullptr;
1291
1292 DebugInformationRecorder* debug_info = C->debug_info();
1293 assert(debug_info->recording_non_safepoints(), "sanity");
1294
1295 debug_info->add_non_safepoint(pc_offset);
1296 int max_depth = youngest_jvms->depth();
1297
1298 // Visit scopes from oldest to youngest.
1299 for (int depth = 1; depth <= max_depth; depth++) {
1300 JVMState* jvms = youngest_jvms->of_depth(depth);
1301 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1302 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1303 methodHandle null_mh;
1304 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1305 }
1306
1307 // Mark the end of the scope set.
1308 debug_info->end_non_safepoint(pc_offset);
1309 }
1310
1311 //------------------------------init_buffer------------------------------------
1312 void PhaseOutput::estimate_buffer_size(int& const_req) {
1313
1314 // Set the initially allocated size
1315 const_req = initial_const_capacity;
1316
1317 // The extra spacing after the code is necessary on some platforms.
1318 // Sometimes we need to patch in a jump after the last instruction,
1319 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1320
1321 // Compute the byte offset where we can store the deopt pc.
1322 if (C->fixed_slots() != 0) {
1323 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1324 }
1325
1326 // Compute prolog code size
1327 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1328 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1329
1330 if (C->has_mach_constant_base_node()) {
1331 uint add_size = 0;
1332 // Fill the constant table.
1333 // Note: This must happen before shorten_branches.
1334 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1335 Block* b = C->cfg()->get_block(i);
1336
1337 for (uint j = 0; j < b->number_of_nodes(); j++) {
1338 Node* n = b->get_node(j);
1339
1340 // If the node is a MachConstantNode evaluate the constant
1341 // value section.
1342 if (n->is_MachConstant()) {
1343 MachConstantNode* machcon = n->as_MachConstant();
1344 machcon->eval_constant(C);
1345 } else if (n->is_Mach()) {
1346 // On Power there are more nodes that issue constants.
1347 add_size += (n->as_Mach()->ins_num_consts() * 8);
1348 }
1349 }
1350 }
1351
1352 // Calculate the offsets of the constants and the size of the
1353 // constant table (including the padding to the next section).
1354 constant_table().calculate_offsets_and_size();
1355 const_req = constant_table().size() + add_size;
1356 }
1357
1358 // Initialize the space for the BufferBlob used to find and verify
1359 // instruction size in MachNode::emit_size()
1360 init_scratch_buffer_blob(const_req);
1361 }
1362
1363 CodeBuffer* PhaseOutput::init_buffer() {
1364 int stub_req = _buf_sizes._stub;
1365 int code_req = _buf_sizes._code;
1366 int const_req = _buf_sizes._const;
1367
1368 int pad_req = NativeCall::byte_size();
1369
1370 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1371 stub_req += bs->estimate_stub_size();
1372
1373 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1374 // class HandlerImpl is platform-specific and defined in the *.ad files.
1375 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1376 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1377 stub_req += MAX_stubs_size; // ensure per-stub margin
1378 code_req += MAX_inst_size; // ensure per-instruction margin
1379
1380 if (StressCodeBuffers)
1381 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1382
1383 int total_req =
1384 const_req +
1385 code_req +
1386 pad_req +
1387 stub_req +
1388 exception_handler_req +
1389 deopt_handler_req; // deopt handler
1390
1391 if (C->has_method_handle_invokes())
1392 total_req += deopt_handler_req; // deopt MH handler
1393
1394 CodeBuffer* cb = code_buffer();
1395 cb->initialize(total_req, _buf_sizes._reloc);
1396
1397 // Have we run out of code space?
1398 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1399 C->record_failure("CodeCache is full");
1400 return nullptr;
1401 }
1402 // Configure the code buffer.
1403 cb->initialize_consts_size(const_req);
1404 cb->initialize_stubs_size(stub_req);
1405 cb->initialize_oop_recorder(C->env()->oop_recorder());
1406
1407 // fill in the nop array for bundling computations
1408 MachNode *_nop_list[Bundle::_nop_count];
1409 Bundle::initialize_nops(_nop_list);
1410
1411 return cb;
1412 }
1413
1414 //------------------------------fill_buffer------------------------------------
1415 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1416 // blk_starts[] contains offsets calculated during short branches processing,
1417 // offsets should not be increased during following steps.
1418
1419 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1420 // of a loop. It is used to determine the padding for loop alignment.
1421 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1422
1423 compute_loop_first_inst_sizes();
1424
1425 // Create oopmap set.
1426 _oop_map_set = new OopMapSet();
1427
1428 // !!!!! This preserves old handling of oopmaps for now
1429 C->debug_info()->set_oopmaps(_oop_map_set);
1430
1431 uint nblocks = C->cfg()->number_of_blocks();
1432 // Count and start of implicit null check instructions
1433 uint inct_cnt = 0;
1434 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1435
1436 // Count and start of calls
1437 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1438
1439 uint return_offset = 0;
1440 int nop_size = (new MachNopNode())->size(C->regalloc());
1441
1442 int previous_offset = 0;
1443 int current_offset = 0;
1444 int last_call_offset = -1;
1445 int last_avoid_back_to_back_offset = -1;
1446 #ifdef ASSERT
1447 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1448 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1449 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1450 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1451 #endif
1452
1453 // Create an array of unused labels, one for each basic block, if printing is enabled
1454 #if defined(SUPPORT_OPTO_ASSEMBLY)
1455 int* node_offsets = nullptr;
1456 uint node_offset_limit = C->unique();
1457
1458 if (C->print_assembly()) {
1459 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1460 }
1461 if (node_offsets != nullptr) {
1462 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1463 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1464 }
1465 #endif
1466
1467 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1468
1469 // Emit the constant table.
1470 if (C->has_mach_constant_base_node()) {
1471 if (!constant_table().emit(masm)) {
1472 C->record_failure("consts section overflow");
1473 return;
1474 }
1475 }
1476
1477 // Create an array of labels, one for each basic block
1478 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1479 for (uint i = 0; i <= nblocks; i++) {
1480 blk_labels[i].init();
1481 }
1482
1483 // Now fill in the code buffer
1484 Node* delay_slot = nullptr;
1485 for (uint i = 0; i < nblocks; i++) {
1486 Block* block = C->cfg()->get_block(i);
1487 _block = block;
1488 Node* head = block->head();
1489
1490 // If this block needs to start aligned (i.e, can be reached other
1491 // than by falling-thru from the previous block), then force the
1492 // start of a new bundle.
1493 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1494 masm->code()->flush_bundle(true);
1495 }
1496
1497 #ifdef ASSERT
1498 if (!block->is_connector()) {
1499 stringStream st;
1500 block->dump_head(C->cfg(), &st);
1501 masm->block_comment(st.freeze());
1502 }
1503 jmp_target[i] = 0;
1504 jmp_offset[i] = 0;
1505 jmp_size[i] = 0;
1506 jmp_rule[i] = 0;
1507 #endif
1508 int blk_offset = current_offset;
1509
1510 // Define the label at the beginning of the basic block
1511 masm->bind(blk_labels[block->_pre_order]);
1512
1513 uint last_inst = block->number_of_nodes();
1514
1515 // Emit block normally, except for last instruction.
1516 // Emit means "dump code bits into code buffer".
1517 for (uint j = 0; j<last_inst; j++) {
1518 _index = j;
1519
1520 // Get the node
1521 Node* n = block->get_node(j);
1522
1523 // See if delay slots are supported
1524 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1525 assert(delay_slot == nullptr, "no use of delay slot node");
1526 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1527
1528 delay_slot = n;
1529 continue;
1530 }
1531
1532 // If this starts a new instruction group, then flush the current one
1533 // (but allow split bundles)
1534 if (Pipeline::requires_bundling() && starts_bundle(n))
1535 masm->code()->flush_bundle(false);
1536
1537 // Special handling for SafePoint/Call Nodes
1538 bool is_mcall = false;
1539 if (n->is_Mach()) {
1540 MachNode *mach = n->as_Mach();
1541 is_mcall = n->is_MachCall();
1542 bool is_sfn = n->is_MachSafePoint();
1543
1544 // If this requires all previous instructions be flushed, then do so
1545 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1546 masm->code()->flush_bundle(true);
1547 current_offset = masm->offset();
1548 }
1549
1550 // A padding may be needed again since a previous instruction
1551 // could be moved to delay slot.
1552
1553 // align the instruction if necessary
1554 int padding = mach->compute_padding(current_offset);
1555 // Make sure safepoint node for polling is distinct from a call's
1556 // return by adding a nop if needed.
1557 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1558 padding = nop_size;
1559 }
1560 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1561 current_offset == last_avoid_back_to_back_offset) {
1562 // Avoid back to back some instructions.
1563 padding = nop_size;
1564 }
1565
1566 if (padding > 0) {
1567 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1568 int nops_cnt = padding / nop_size;
1569 MachNode *nop = new MachNopNode(nops_cnt);
1570 block->insert_node(nop, j++);
1571 last_inst++;
1572 C->cfg()->map_node_to_block(nop, block);
1573 // Ensure enough space.
1574 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1575 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1576 C->record_failure("CodeCache is full");
1577 return;
1578 }
1579 nop->emit(masm, C->regalloc());
1580 masm->code()->flush_bundle(true);
1581 current_offset = masm->offset();
1582 }
1583
1584 bool observe_safepoint = is_sfn;
1585 // Remember the start of the last call in a basic block
1586 if (is_mcall) {
1587 MachCallNode *mcall = mach->as_MachCall();
1588
1589 // This destination address is NOT PC-relative
1590 mcall->method_set((intptr_t)mcall->entry_point());
1591
1592 // Save the return address
1593 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1594
1595 observe_safepoint = mcall->guaranteed_safepoint();
1596 }
1597
1598 // sfn will be valid whenever mcall is valid now because of inheritance
1599 if (observe_safepoint) {
1600 // Handle special safepoint nodes for synchronization
1601 if (!is_mcall) {
1602 MachSafePointNode *sfn = mach->as_MachSafePoint();
1603 // !!!!! Stubs only need an oopmap right now, so bail out
1604 if (sfn->jvms()->method() == nullptr) {
1605 // Write the oopmap directly to the code blob??!!
1606 continue;
1607 }
1608 } // End synchronization
1609
1610 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1611 current_offset);
1612 Process_OopMap_Node(mach, current_offset);
1613 } // End if safepoint
1614
1615 // If this is a null check, then add the start of the previous instruction to the list
1616 else if( mach->is_MachNullCheck() ) {
1617 inct_starts[inct_cnt++] = previous_offset;
1618 }
1619
1620 // If this is a branch, then fill in the label with the target BB's label
1621 else if (mach->is_MachBranch()) {
1622 // This requires the TRUE branch target be in succs[0]
1623 uint block_num = block->non_connector_successor(0)->_pre_order;
1624
1625 // Try to replace long branch if delay slot is not used,
1626 // it is mostly for back branches since forward branch's
1627 // distance is not updated yet.
1628 bool delay_slot_is_used = valid_bundle_info(n) &&
1629 C->output()->node_bundling(n)->use_unconditional_delay();
1630 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1631 assert(delay_slot == nullptr, "not expecting delay slot node");
1632 int br_size = n->size(C->regalloc());
1633 int offset = blk_starts[block_num] - current_offset;
1634 if (block_num >= i) {
1635 // Current and following block's offset are not
1636 // finalized yet, adjust distance by the difference
1637 // between calculated and final offsets of current block.
1638 offset -= (blk_starts[i] - blk_offset);
1639 }
1640 // In the following code a nop could be inserted before
1641 // the branch which will increase the backward distance.
1642 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1643 if (needs_padding && offset <= 0)
1644 offset -= nop_size;
1645
1646 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1647 // We've got a winner. Replace this branch.
1648 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1649
1650 // Update the jmp_size.
1651 int new_size = replacement->size(C->regalloc());
1652 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1653 // Insert padding between avoid_back_to_back branches.
1654 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1655 MachNode *nop = new MachNopNode();
1656 block->insert_node(nop, j++);
1657 C->cfg()->map_node_to_block(nop, block);
1658 last_inst++;
1659 nop->emit(masm, C->regalloc());
1660 masm->code()->flush_bundle(true);
1661 current_offset = masm->offset();
1662 }
1663 #ifdef ASSERT
1664 jmp_target[i] = block_num;
1665 jmp_offset[i] = current_offset - blk_offset;
1666 jmp_size[i] = new_size;
1667 jmp_rule[i] = mach->rule();
1668 #endif
1669 block->map_node(replacement, j);
1670 mach->subsume_by(replacement, C);
1671 n = replacement;
1672 mach = replacement;
1673 }
1674 }
1675 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1676 } else if (mach->ideal_Opcode() == Op_Jump) {
1677 for (uint h = 0; h < block->_num_succs; h++) {
1678 Block* succs_block = block->_succs[h];
1679 for (uint j = 1; j < succs_block->num_preds(); j++) {
1680 Node* jpn = succs_block->pred(j);
1681 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1682 uint block_num = succs_block->non_connector()->_pre_order;
1683 Label *blkLabel = &blk_labels[block_num];
1684 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1685 }
1686 }
1687 }
1688 } else if (!n->is_Proj()) {
1689 // Remember the beginning of the previous instruction, in case
1690 // it's followed by a flag-kill and a null-check. Happens on
1691 // Intel all the time, with add-to-memory kind of opcodes.
1692 previous_offset = current_offset;
1693 }
1694
1695 // Not an else-if!
1696 // If this is a trap based cmp then add its offset to the list.
1697 if (mach->is_TrapBasedCheckNode()) {
1698 inct_starts[inct_cnt++] = current_offset;
1699 }
1700 }
1701
1702 // Verify that there is sufficient space remaining
1703 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1704 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1705 C->record_failure("CodeCache is full");
1706 return;
1707 }
1708
1709 // Save the offset for the listing
1710 #if defined(SUPPORT_OPTO_ASSEMBLY)
1711 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1712 node_offsets[n->_idx] = masm->offset();
1713 }
1714 #endif
1715 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1716
1717 // "Normal" instruction case
1718 DEBUG_ONLY(uint instr_offset = masm->offset());
1719 n->emit(masm, C->regalloc());
1720 current_offset = masm->offset();
1721
1722 // Above we only verified that there is enough space in the instruction section.
1723 // However, the instruction may emit stubs that cause code buffer expansion.
1724 // Bail out here if expansion failed due to a lack of code cache space.
1725 if (C->failing()) {
1726 return;
1727 }
1728
1729 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1730 "ret_addr_offset() not within emitted code");
1731
1732 #ifdef ASSERT
1733 uint n_size = n->size(C->regalloc());
1734 if (n_size < (current_offset-instr_offset)) {
1735 MachNode* mach = n->as_Mach();
1736 n->dump();
1737 mach->dump_format(C->regalloc(), tty);
1738 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1739 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1740 tty->print_cr(" ------------------- ");
1741 BufferBlob* blob = this->scratch_buffer_blob();
1742 address blob_begin = blob->content_begin();
1743 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1744 assert(false, "wrong size of mach node");
1745 }
1746 #endif
1747 non_safepoints.observe_instruction(n, current_offset);
1748
1749 // mcall is last "call" that can be a safepoint
1750 // record it so we can see if a poll will directly follow it
1751 // in which case we'll need a pad to make the PcDesc sites unique
1752 // see 5010568. This can be slightly inaccurate but conservative
1753 // in the case that return address is not actually at current_offset.
1754 // This is a small price to pay.
1755
1756 if (is_mcall) {
1757 last_call_offset = current_offset;
1758 }
1759
1760 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1761 // Avoid back to back some instructions.
1762 last_avoid_back_to_back_offset = current_offset;
1763 }
1764
1765 // See if this instruction has a delay slot
1766 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1767 guarantee(delay_slot != nullptr, "expecting delay slot node");
1768
1769 // Back up 1 instruction
1770 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1771
1772 // Save the offset for the listing
1773 #if defined(SUPPORT_OPTO_ASSEMBLY)
1774 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1775 node_offsets[delay_slot->_idx] = masm->offset();
1776 }
1777 #endif
1778
1779 // Support a SafePoint in the delay slot
1780 if (delay_slot->is_MachSafePoint()) {
1781 MachNode *mach = delay_slot->as_Mach();
1782 // !!!!! Stubs only need an oopmap right now, so bail out
1783 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1784 // Write the oopmap directly to the code blob??!!
1785 delay_slot = nullptr;
1786 continue;
1787 }
1788
1789 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1790 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1791 adjusted_offset);
1792 // Generate an OopMap entry
1793 Process_OopMap_Node(mach, adjusted_offset);
1794 }
1795
1796 // Insert the delay slot instruction
1797 delay_slot->emit(masm, C->regalloc());
1798
1799 // Don't reuse it
1800 delay_slot = nullptr;
1801 }
1802
1803 } // End for all instructions in block
1804
1805 // If the next block is the top of a loop, pad this block out to align
1806 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1807 if (i < nblocks-1) {
1808 Block *nb = C->cfg()->get_block(i + 1);
1809 int padding = nb->alignment_padding(current_offset);
1810 if( padding > 0 ) {
1811 MachNode *nop = new MachNopNode(padding / nop_size);
1812 block->insert_node(nop, block->number_of_nodes());
1813 C->cfg()->map_node_to_block(nop, block);
1814 nop->emit(masm, C->regalloc());
1815 current_offset = masm->offset();
1816 }
1817 }
1818 // Verify that the distance for generated before forward
1819 // short branches is still valid.
1820 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1821
1822 // Save new block start offset
1823 blk_starts[i] = blk_offset;
1824 } // End of for all blocks
1825 blk_starts[nblocks] = current_offset;
1826
1827 non_safepoints.flush_at_end();
1828
1829 // Offset too large?
1830 if (C->failing()) return;
1831
1832 // Define a pseudo-label at the end of the code
1833 masm->bind( blk_labels[nblocks] );
1834
1835 // Compute the size of the first block
1836 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1837
1838 #ifdef ASSERT
1839 for (uint i = 0; i < nblocks; i++) { // For all blocks
1840 if (jmp_target[i] != 0) {
1841 int br_size = jmp_size[i];
1842 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1843 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1844 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]);
1845 assert(false, "Displacement too large for short jmp");
1846 }
1847 }
1848 }
1849 #endif
1850
1851 if (!masm->code()->finalize_stubs()) {
1852 C->record_failure("CodeCache is full");
1853 return;
1854 }
1855
1856 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1857 bs->emit_stubs(*masm->code());
1858 if (C->failing()) return;
1859
1860 // Fill in stubs.
1861 assert(masm->inst_mark() == nullptr, "should be.");
1862 _stub_list.emit(*masm);
1863 if (C->failing()) return;
1864
1865 #ifndef PRODUCT
1866 // Information on the size of the method, without the extraneous code
1867 Scheduling::increment_method_size(masm->offset());
1868 #endif
1869
1870 // ------------------
1871 // Fill in exception table entries.
1872 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1873
1874 // Only java methods have exception handlers and deopt handlers
1875 // class HandlerImpl is platform-specific and defined in the *.ad files.
1876 if (C->method()) {
1877 // Emit the exception handler code.
1878 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1879 if (C->failing()) {
1880 return; // CodeBuffer::expand failed
1881 }
1882 // Emit the deopt handler code.
1883 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1884
1885 // Emit the MethodHandle deopt handler code (if required).
1886 if (C->has_method_handle_invokes() && !C->failing()) {
1887 // We can use the same code as for the normal deopt handler, we
1888 // just need a different entry point address.
1889 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1890 }
1891 }
1892
1893 // One last check for failed CodeBuffer::expand:
1894 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1895 C->record_failure("CodeCache is full");
1896 return;
1897 }
1898
1899 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1900 if (C->print_assembly()) {
1901 tty->cr();
1902 tty->print_cr("============================= C2-compiled nmethod ==============================");
1903 }
1904 #endif
1905
1906 #if defined(SUPPORT_OPTO_ASSEMBLY)
1907 // Dump the assembly code, including basic-block numbers
1908 if (C->print_assembly()) {
1909 ttyLocker ttyl; // keep the following output all in one block
1910 if (!VMThread::should_terminate()) { // test this under the tty lock
1911 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1912 // We call them first and write to a stringStream, then we retake the lock to
1913 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1914 ResourceMark rm;
1915 stringStream method_metadata_str;
1916 if (C->method() != nullptr) {
1917 C->method()->print_metadata(&method_metadata_str);
1918 }
1919 stringStream dump_asm_str;
1920 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1921
1922 NoSafepointVerifier nsv;
1923 ttyLocker ttyl2;
1924 // This output goes directly to the tty, not the compiler log.
1925 // To enable tools to match it up with the compilation activity,
1926 // be sure to tag this tty output with the compile ID.
1927 if (xtty != nullptr) {
1928 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1929 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1930 }
1931 if (C->method() != nullptr) {
1932 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1933 tty->print_raw(method_metadata_str.freeze());
1934 } else if (C->stub_name() != nullptr) {
1935 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1936 }
1937 tty->cr();
1938 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1939 tty->print_raw(dump_asm_str.freeze());
1940 tty->print_cr("--------------------------------------------------------------------------------");
1941 if (xtty != nullptr) {
1942 xtty->tail("opto_assembly");
1943 }
1944 }
1945 }
1946 #endif
1947 }
1948
1949 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1950 _inc_table.set_size(cnt);
1951
1952 uint inct_cnt = 0;
1953 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1954 Block* block = C->cfg()->get_block(i);
1955 Node *n = nullptr;
1956 int j;
1957
1958 // Find the branch; ignore trailing NOPs.
1959 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1960 n = block->get_node(j);
1961 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1962 break;
1963 }
1964 }
1965
1966 // If we didn't find anything, continue
1967 if (j < 0) {
1968 continue;
1969 }
1970
1971 // Compute ExceptionHandlerTable subtable entry and add it
1972 // (skip empty blocks)
1973 if (n->is_Catch()) {
1974
1975 // Get the offset of the return from the call
1976 uint call_return = call_returns[block->_pre_order];
1977 #ifdef ASSERT
1978 assert( call_return > 0, "no call seen for this basic block" );
1979 while (block->get_node(--j)->is_MachProj()) ;
1980 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1981 #endif
1982 // last instruction is a CatchNode, find it's CatchProjNodes
1983 int nof_succs = block->_num_succs;
1984 // allocate space
1985 GrowableArray<intptr_t> handler_bcis(nof_succs);
1986 GrowableArray<intptr_t> handler_pcos(nof_succs);
1987 // iterate through all successors
1988 for (int j = 0; j < nof_succs; j++) {
1989 Block* s = block->_succs[j];
1990 bool found_p = false;
1991 for (uint k = 1; k < s->num_preds(); k++) {
1992 Node* pk = s->pred(k);
1993 if (pk->is_CatchProj() && pk->in(0) == n) {
1994 const CatchProjNode* p = pk->as_CatchProj();
1995 found_p = true;
1996 // add the corresponding handler bci & pco information
1997 if (p->_con != CatchProjNode::fall_through_index) {
1998 // p leads to an exception handler (and is not fall through)
1999 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2000 // no duplicates, please
2001 if (!handler_bcis.contains(p->handler_bci())) {
2002 uint block_num = s->non_connector()->_pre_order;
2003 handler_bcis.append(p->handler_bci());
2004 handler_pcos.append(blk_labels[block_num].loc_pos());
2005 }
2006 }
2007 }
2008 }
2009 assert(found_p, "no matching predecessor found");
2010 // Note: Due to empty block removal, one block may have
2011 // several CatchProj inputs, from the same Catch.
2012 }
2013
2014 // Set the offset of the return from the call
2015 assert(handler_bcis.find(-1) != -1, "must have default handler");
2016 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2017 continue;
2018 }
2019
2020 // Handle implicit null exception table updates
2021 if (n->is_MachNullCheck()) {
2022 assert(n->in(1)->as_Mach()->barrier_data() == 0,
2023 "Implicit null checks on memory accesses with barriers are not yet supported");
2024 uint block_num = block->non_connector_successor(0)->_pre_order;
2025 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2026 continue;
2027 }
2028 // Handle implicit exception table updates: trap instructions.
2029 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2030 uint block_num = block->non_connector_successor(0)->_pre_order;
2031 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2032 continue;
2033 }
2034 } // End of for all blocks fill in exception table entries
2035 }
2036
2037 // Static Variables
2038 #ifndef PRODUCT
2039 uint Scheduling::_total_nop_size = 0;
2040 uint Scheduling::_total_method_size = 0;
2041 uint Scheduling::_total_branches = 0;
2042 uint Scheduling::_total_unconditional_delays = 0;
2043 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2044 #endif
2045
2046 // Initializer for class Scheduling
2047
2048 Scheduling::Scheduling(Arena *arena, Compile &compile)
2049 : _arena(arena),
2050 _cfg(compile.cfg()),
2051 _regalloc(compile.regalloc()),
2052 _scheduled(arena),
2053 _available(arena),
2054 _reg_node(arena),
2055 _pinch_free_list(arena),
2056 _next_node(nullptr),
2057 _bundle_instr_count(0),
2058 _bundle_cycle_number(0),
2059 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2060 #ifndef PRODUCT
2061 , _branches(0)
2062 , _unconditional_delays(0)
2063 #endif
2064 {
2065 // Create a MachNopNode
2066 _nop = new MachNopNode();
2067
2068 // Now that the nops are in the array, save the count
2069 // (but allow entries for the nops)
2070 _node_bundling_limit = compile.unique();
2071 uint node_max = _regalloc->node_regs_max_index();
2072
2073 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2074
2075 // This one is persistent within the Compile class
2076 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2077
2078 // Allocate space for fixed-size arrays
2079 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2080 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2081
2082 // Clear the arrays
2083 for (uint i = 0; i < node_max; i++) {
2084 ::new (&_node_bundling_base[i]) Bundle();
2085 }
2086 memset(_uses, 0, node_max * sizeof(short));
2087 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2088
2089 // Clear the bundling information
2090 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2091
2092 // Get the last node
2093 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2094
2095 _next_node = block->get_node(block->number_of_nodes() - 1);
2096 }
2097
2098 #ifndef PRODUCT
2099 // Scheduling destructor
2100 Scheduling::~Scheduling() {
2101 _total_branches += _branches;
2102 _total_unconditional_delays += _unconditional_delays;
2103 }
2104 #endif
2105
2106 // Step ahead "i" cycles
2107 void Scheduling::step(uint i) {
2108
2109 Bundle *bundle = node_bundling(_next_node);
2110 bundle->set_starts_bundle();
2111
2112 // Update the bundle record, but leave the flags information alone
2113 if (_bundle_instr_count > 0) {
2114 bundle->set_instr_count(_bundle_instr_count);
2115 bundle->set_resources_used(_bundle_use.resourcesUsed());
2116 }
2117
2118 // Update the state information
2119 _bundle_instr_count = 0;
2120 _bundle_cycle_number += i;
2121 _bundle_use.step(i);
2122 }
2123
2124 void Scheduling::step_and_clear() {
2125 Bundle *bundle = node_bundling(_next_node);
2126 bundle->set_starts_bundle();
2127
2128 // Update the bundle record
2129 if (_bundle_instr_count > 0) {
2130 bundle->set_instr_count(_bundle_instr_count);
2131 bundle->set_resources_used(_bundle_use.resourcesUsed());
2132
2133 _bundle_cycle_number += 1;
2134 }
2135
2136 // Clear the bundling information
2137 _bundle_instr_count = 0;
2138 _bundle_use.reset();
2139
2140 memcpy(_bundle_use_elements,
2141 Pipeline_Use::elaborated_elements,
2142 sizeof(Pipeline_Use::elaborated_elements));
2143 }
2144
2145 // Perform instruction scheduling and bundling over the sequence of
2146 // instructions in backwards order.
2147 void PhaseOutput::ScheduleAndBundle() {
2148
2149 // Don't optimize this if it isn't a method
2150 if (!C->method())
2151 return;
2152
2153 // Don't optimize this if scheduling is disabled
2154 if (!C->do_scheduling())
2155 return;
2156
2157 // Scheduling code works only with pairs (8 bytes) maximum.
2158 // And when the scalable vector register is used, we may spill/unspill
2159 // the whole reg regardless of the max vector size.
2160 if (C->max_vector_size() > 8 ||
2161 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2162 return;
2163 }
2164
2165 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2166
2167 // Create a data structure for all the scheduling information
2168 Scheduling scheduling(Thread::current()->resource_area(), *C);
2169
2170 // Walk backwards over each basic block, computing the needed alignment
2171 // Walk over all the basic blocks
2172 scheduling.DoScheduling();
2173
2174 #ifndef PRODUCT
2175 if (C->trace_opto_output()) {
2176 // Buffer and print all at once
2177 ResourceMark rm;
2178 stringStream ss;
2179 ss.print("\n---- After ScheduleAndBundle ----\n");
2180 print_scheduling(&ss);
2181 tty->print("%s", ss.as_string());
2182 }
2183 #endif
2184 }
2185
2186 #ifndef PRODUCT
2187 // Separated out so that it can be called directly from debugger
2188 void PhaseOutput::print_scheduling() {
2189 print_scheduling(tty);
2190 }
2191
2192 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2193 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2194 output_stream->print("\nBB#%03d:\n", i);
2195 Block* block = C->cfg()->get_block(i);
2196 for (uint j = 0; j < block->number_of_nodes(); j++) {
2197 Node* n = block->get_node(j);
2198 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2199 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2200 n->dump("\n", false, output_stream);
2201 }
2202 }
2203 }
2204 #endif
2205
2206 // See if this node fits into the present instruction bundle
2207 bool Scheduling::NodeFitsInBundle(Node *n) {
2208 uint n_idx = n->_idx;
2209
2210 // If this is the unconditional delay instruction, then it fits
2211 if (n == _unconditional_delay_slot) {
2212 #ifndef PRODUCT
2213 if (_cfg->C->trace_opto_output())
2214 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2215 #endif
2216 return (true);
2217 }
2218
2219 // If the node cannot be scheduled this cycle, skip it
2220 if (_current_latency[n_idx] > _bundle_cycle_number) {
2221 #ifndef PRODUCT
2222 if (_cfg->C->trace_opto_output())
2223 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2224 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2225 #endif
2226 return (false);
2227 }
2228
2229 const Pipeline *node_pipeline = n->pipeline();
2230
2231 uint instruction_count = node_pipeline->instructionCount();
2232 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2233 instruction_count = 0;
2234 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2235 instruction_count++;
2236
2237 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2238 #ifndef PRODUCT
2239 if (_cfg->C->trace_opto_output())
2240 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2241 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2242 #endif
2243 return (false);
2244 }
2245
2246 // Don't allow non-machine nodes to be handled this way
2247 if (!n->is_Mach() && instruction_count == 0)
2248 return (false);
2249
2250 // See if there is any overlap
2251 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2252
2253 if (delay > 0) {
2254 #ifndef PRODUCT
2255 if (_cfg->C->trace_opto_output())
2256 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2257 #endif
2258 return false;
2259 }
2260
2261 #ifndef PRODUCT
2262 if (_cfg->C->trace_opto_output())
2263 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2264 #endif
2265
2266 return true;
2267 }
2268
2269 Node * Scheduling::ChooseNodeToBundle() {
2270 uint siz = _available.size();
2271
2272 if (siz == 0) {
2273
2274 #ifndef PRODUCT
2275 if (_cfg->C->trace_opto_output())
2276 tty->print("# ChooseNodeToBundle: null\n");
2277 #endif
2278 return (nullptr);
2279 }
2280
2281 // Fast path, if only 1 instruction in the bundle
2282 if (siz == 1) {
2283 #ifndef PRODUCT
2284 if (_cfg->C->trace_opto_output()) {
2285 tty->print("# ChooseNodeToBundle (only 1): ");
2286 _available[0]->dump();
2287 }
2288 #endif
2289 return (_available[0]);
2290 }
2291
2292 // Don't bother, if the bundle is already full
2293 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2294 for ( uint i = 0; i < siz; i++ ) {
2295 Node *n = _available[i];
2296
2297 // Skip projections, we'll handle them another way
2298 if (n->is_Proj())
2299 continue;
2300
2301 // This presupposed that instructions are inserted into the
2302 // available list in a legality order; i.e. instructions that
2303 // must be inserted first are at the head of the list
2304 if (NodeFitsInBundle(n)) {
2305 #ifndef PRODUCT
2306 if (_cfg->C->trace_opto_output()) {
2307 tty->print("# ChooseNodeToBundle: ");
2308 n->dump();
2309 }
2310 #endif
2311 return (n);
2312 }
2313 }
2314 }
2315
2316 // Nothing fits in this bundle, choose the highest priority
2317 #ifndef PRODUCT
2318 if (_cfg->C->trace_opto_output()) {
2319 tty->print("# ChooseNodeToBundle: ");
2320 _available[0]->dump();
2321 }
2322 #endif
2323
2324 return _available[0];
2325 }
2326
2327 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2328 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2329
2330 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2331 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2332 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2333 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2334
2335 // Comparison between stack -> reg and stack -> reg
2336 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2337 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2338 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2339 }
2340
2341 // Comparison between reg -> stack and reg -> stack
2342 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2343 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2344 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2345 }
2346
2347 return 0; // Not comparable
2348 }
2349
2350 void Scheduling::AddNodeToAvailableList(Node *n) {
2351 assert( !n->is_Proj(), "projections never directly made available" );
2352 #ifndef PRODUCT
2353 if (_cfg->C->trace_opto_output()) {
2354 tty->print("# AddNodeToAvailableList: ");
2355 n->dump();
2356 }
2357 #endif
2358
2359 int latency = _current_latency[n->_idx];
2360
2361 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2362 // for stack spilling or unspilling have the same latency, we sort
2363 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2364 // have more opportunities to do ld/st merging
2365 uint i;
2366 for (i = 0; i < _available.size(); i++) {
2367 if (_current_latency[_available[i]->_idx] > latency) {
2368 break;
2369 } else if (_current_latency[_available[i]->_idx] == latency &&
2370 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2371 compare_two_spill_nodes(n, _available[i]) > 0) {
2372 break;
2373 }
2374 }
2375
2376 // Special Check for compares following branches
2377 if( n->is_Mach() && _scheduled.size() > 0 ) {
2378 int op = n->as_Mach()->ideal_Opcode();
2379 Node *last = _scheduled[0];
2380 if( last->is_MachIf() && last->in(1) == n &&
2381 ( op == Op_CmpI ||
2382 op == Op_CmpU ||
2383 op == Op_CmpUL ||
2384 op == Op_CmpP ||
2385 op == Op_CmpF ||
2386 op == Op_CmpD ||
2387 op == Op_CmpL ) ) {
2388
2389 // Recalculate position, moving to front of same latency
2390 for ( i=0 ; i < _available.size(); i++ )
2391 if (_current_latency[_available[i]->_idx] >= latency)
2392 break;
2393 }
2394 }
2395
2396 // Insert the node in the available list
2397 _available.insert(i, n);
2398
2399 #ifndef PRODUCT
2400 if (_cfg->C->trace_opto_output())
2401 dump_available();
2402 #endif
2403 }
2404
2405 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2406 for ( uint i=0; i < n->len(); i++ ) {
2407 Node *def = n->in(i);
2408 if (!def) continue;
2409 if( def->is_Proj() ) // If this is a machine projection, then
2410 def = def->in(0); // propagate usage thru to the base instruction
2411
2412 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2413 continue;
2414 }
2415
2416 // Compute the latency
2417 uint l = _bundle_cycle_number + n->latency(i);
2418 if (_current_latency[def->_idx] < l)
2419 _current_latency[def->_idx] = l;
2420
2421 // If this does not have uses then schedule it
2422 if ((--_uses[def->_idx]) == 0)
2423 AddNodeToAvailableList(def);
2424 }
2425 }
2426
2427 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2428 #ifndef PRODUCT
2429 if (_cfg->C->trace_opto_output()) {
2430 tty->print("# AddNodeToBundle: ");
2431 n->dump();
2432 }
2433 #endif
2434
2435 // Remove this from the available list
2436 uint i;
2437 for (i = 0; i < _available.size(); i++)
2438 if (_available[i] == n)
2439 break;
2440 assert(i < _available.size(), "entry in _available list not found");
2441 _available.remove(i);
2442
2443 // See if this fits in the current bundle
2444 const Pipeline *node_pipeline = n->pipeline();
2445 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2446
2447 // Check for instructions to be placed in the delay slot. We
2448 // do this before we actually schedule the current instruction,
2449 // because the delay slot follows the current instruction.
2450 if (Pipeline::_branch_has_delay_slot &&
2451 node_pipeline->hasBranchDelay() &&
2452 !_unconditional_delay_slot) {
2453
2454 uint siz = _available.size();
2455
2456 // Conditional branches can support an instruction that
2457 // is unconditionally executed and not dependent by the
2458 // branch, OR a conditionally executed instruction if
2459 // the branch is taken. In practice, this means that
2460 // the first instruction at the branch target is
2461 // copied to the delay slot, and the branch goes to
2462 // the instruction after that at the branch target
2463 if ( n->is_MachBranch() ) {
2464
2465 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2466 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2467
2468 #ifndef PRODUCT
2469 _branches++;
2470 #endif
2471
2472 // At least 1 instruction is on the available list
2473 // that is not dependent on the branch
2474 for (uint i = 0; i < siz; i++) {
2475 Node *d = _available[i];
2476 const Pipeline *avail_pipeline = d->pipeline();
2477
2478 // Don't allow safepoints in the branch shadow, that will
2479 // cause a number of difficulties
2480 if ( avail_pipeline->instructionCount() == 1 &&
2481 !avail_pipeline->hasMultipleBundles() &&
2482 !avail_pipeline->hasBranchDelay() &&
2483 Pipeline::instr_has_unit_size() &&
2484 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2485 NodeFitsInBundle(d) &&
2486 !node_bundling(d)->used_in_delay()) {
2487
2488 if (d->is_Mach() && !d->is_MachSafePoint()) {
2489 // A node that fits in the delay slot was found, so we need to
2490 // set the appropriate bits in the bundle pipeline information so
2491 // that it correctly indicates resource usage. Later, when we
2492 // attempt to add this instruction to the bundle, we will skip
2493 // setting the resource usage.
2494 _unconditional_delay_slot = d;
2495 node_bundling(n)->set_use_unconditional_delay();
2496 node_bundling(d)->set_used_in_unconditional_delay();
2497 _bundle_use.add_usage(avail_pipeline->resourceUse());
2498 _current_latency[d->_idx] = _bundle_cycle_number;
2499 _next_node = d;
2500 ++_bundle_instr_count;
2501 #ifndef PRODUCT
2502 _unconditional_delays++;
2503 #endif
2504 break;
2505 }
2506 }
2507 }
2508 }
2509
2510 // No delay slot, add a nop to the usage
2511 if (!_unconditional_delay_slot) {
2512 // See if adding an instruction in the delay slot will overflow
2513 // the bundle.
2514 if (!NodeFitsInBundle(_nop)) {
2515 #ifndef PRODUCT
2516 if (_cfg->C->trace_opto_output())
2517 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2518 #endif
2519 step(1);
2520 }
2521
2522 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2523 _next_node = _nop;
2524 ++_bundle_instr_count;
2525 }
2526
2527 // See if the instruction in the delay slot requires a
2528 // step of the bundles
2529 if (!NodeFitsInBundle(n)) {
2530 #ifndef PRODUCT
2531 if (_cfg->C->trace_opto_output())
2532 tty->print("# *** STEP(branch won't fit) ***\n");
2533 #endif
2534 // Update the state information
2535 _bundle_instr_count = 0;
2536 _bundle_cycle_number += 1;
2537 _bundle_use.step(1);
2538 }
2539 }
2540
2541 // Get the number of instructions
2542 uint instruction_count = node_pipeline->instructionCount();
2543 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2544 instruction_count = 0;
2545
2546 // Compute the latency information
2547 uint delay = 0;
2548
2549 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2550 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2551 if (relative_latency < 0)
2552 relative_latency = 0;
2553
2554 delay = _bundle_use.full_latency(relative_latency, node_usage);
2555
2556 // Does not fit in this bundle, start a new one
2557 if (delay > 0) {
2558 step(delay);
2559
2560 #ifndef PRODUCT
2561 if (_cfg->C->trace_opto_output())
2562 tty->print("# *** STEP(%d) ***\n", delay);
2563 #endif
2564 }
2565 }
2566
2567 // If this was placed in the delay slot, ignore it
2568 if (n != _unconditional_delay_slot) {
2569
2570 if (delay == 0) {
2571 if (node_pipeline->hasMultipleBundles()) {
2572 #ifndef PRODUCT
2573 if (_cfg->C->trace_opto_output())
2574 tty->print("# *** STEP(multiple instructions) ***\n");
2575 #endif
2576 step(1);
2577 }
2578
2579 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2580 #ifndef PRODUCT
2581 if (_cfg->C->trace_opto_output())
2582 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2583 instruction_count + _bundle_instr_count,
2584 Pipeline::_max_instrs_per_cycle);
2585 #endif
2586 step(1);
2587 }
2588 }
2589
2590 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2591 _bundle_instr_count++;
2592
2593 // Set the node's latency
2594 _current_latency[n->_idx] = _bundle_cycle_number;
2595
2596 // Now merge the functional unit information
2597 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2598 _bundle_use.add_usage(node_usage);
2599
2600 // Increment the number of instructions in this bundle
2601 _bundle_instr_count += instruction_count;
2602
2603 // Remember this node for later
2604 if (n->is_Mach())
2605 _next_node = n;
2606 }
2607
2608 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2609 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2610 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2611 // into the block. All other scheduled nodes get put in the schedule here.
2612 int op = n->Opcode();
2613 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2614 (op != Op_Node && // Not an unused antidepedence node and
2615 // not an unallocated boxlock
2616 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2617
2618 // Push any trailing projections
2619 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2620 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2621 Node *foi = n->fast_out(i);
2622 if( foi->is_Proj() )
2623 _scheduled.push(foi);
2624 }
2625 }
2626
2627 // Put the instruction in the schedule list
2628 _scheduled.push(n);
2629 }
2630
2631 #ifndef PRODUCT
2632 if (_cfg->C->trace_opto_output())
2633 dump_available();
2634 #endif
2635
2636 // Walk all the definitions, decrementing use counts, and
2637 // if a definition has a 0 use count, place it in the available list.
2638 DecrementUseCounts(n,bb);
2639 }
2640
2641 // This method sets the use count within a basic block. We will ignore all
2642 // uses outside the current basic block. As we are doing a backwards walk,
2643 // any node we reach that has a use count of 0 may be scheduled. This also
2644 // avoids the problem of cyclic references from phi nodes, as long as phi
2645 // nodes are at the front of the basic block. This method also initializes
2646 // the available list to the set of instructions that have no uses within this
2647 // basic block.
2648 void Scheduling::ComputeUseCount(const Block *bb) {
2649 #ifndef PRODUCT
2650 if (_cfg->C->trace_opto_output())
2651 tty->print("# -> ComputeUseCount\n");
2652 #endif
2653
2654 // Clear the list of available and scheduled instructions, just in case
2655 _available.clear();
2656 _scheduled.clear();
2657
2658 // No delay slot specified
2659 _unconditional_delay_slot = nullptr;
2660
2661 #ifdef ASSERT
2662 for( uint i=0; i < bb->number_of_nodes(); i++ )
2663 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2664 #endif
2665
2666 // Force the _uses count to never go to zero for unscheduable pieces
2667 // of the block
2668 for( uint k = 0; k < _bb_start; k++ )
2669 _uses[bb->get_node(k)->_idx] = 1;
2670 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2671 _uses[bb->get_node(l)->_idx] = 1;
2672
2673 // Iterate backwards over the instructions in the block. Don't count the
2674 // branch projections at end or the block header instructions.
2675 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2676 Node *n = bb->get_node(j);
2677 if( n->is_Proj() ) continue; // Projections handled another way
2678
2679 // Account for all uses
2680 for ( uint k = 0; k < n->len(); k++ ) {
2681 Node *inp = n->in(k);
2682 if (!inp) continue;
2683 assert(inp != n, "no cycles allowed" );
2684 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2685 if (inp->is_Proj()) { // Skip through Proj's
2686 inp = inp->in(0);
2687 }
2688 ++_uses[inp->_idx]; // Count 1 block-local use
2689 }
2690 }
2691
2692 // If this instruction has a 0 use count, then it is available
2693 if (!_uses[n->_idx]) {
2694 _current_latency[n->_idx] = _bundle_cycle_number;
2695 AddNodeToAvailableList(n);
2696 }
2697
2698 #ifndef PRODUCT
2699 if (_cfg->C->trace_opto_output()) {
2700 tty->print("# uses: %3d: ", _uses[n->_idx]);
2701 n->dump();
2702 }
2703 #endif
2704 }
2705
2706 #ifndef PRODUCT
2707 if (_cfg->C->trace_opto_output())
2708 tty->print("# <- ComputeUseCount\n");
2709 #endif
2710 }
2711
2712 // This routine performs scheduling on each basic block in reverse order,
2713 // using instruction latencies and taking into account function unit
2714 // availability.
2715 void Scheduling::DoScheduling() {
2716 #ifndef PRODUCT
2717 if (_cfg->C->trace_opto_output())
2718 tty->print("# -> DoScheduling\n");
2719 #endif
2720
2721 Block *succ_bb = nullptr;
2722 Block *bb;
2723 Compile* C = Compile::current();
2724
2725 // Walk over all the basic blocks in reverse order
2726 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2727 bb = _cfg->get_block(i);
2728
2729 #ifndef PRODUCT
2730 if (_cfg->C->trace_opto_output()) {
2731 tty->print("# Schedule BB#%03d (initial)\n", i);
2732 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2733 bb->get_node(j)->dump();
2734 }
2735 }
2736 #endif
2737
2738 // On the head node, skip processing
2739 if (bb == _cfg->get_root_block()) {
2740 continue;
2741 }
2742
2743 // Skip empty, connector blocks
2744 if (bb->is_connector())
2745 continue;
2746
2747 // If the following block is not the sole successor of
2748 // this one, then reset the pipeline information
2749 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2750 #ifndef PRODUCT
2751 if (_cfg->C->trace_opto_output()) {
2752 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2753 _next_node->_idx, _bundle_instr_count);
2754 }
2755 #endif
2756 step_and_clear();
2757 }
2758
2759 // Leave untouched the starting instruction, any Phis, a CreateEx node
2760 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2761 _bb_end = bb->number_of_nodes()-1;
2762 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2763 Node *n = bb->get_node(_bb_start);
2764 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2765 // Also, MachIdealNodes do not get scheduled
2766 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2767 MachNode *mach = n->as_Mach();
2768 int iop = mach->ideal_Opcode();
2769 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2770 if( iop == Op_Con ) continue; // Do not schedule Top
2771 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2772 mach->pipeline() == MachNode::pipeline_class() &&
2773 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2774 continue;
2775 break; // Funny loop structure to be sure...
2776 }
2777 // Compute last "interesting" instruction in block - last instruction we
2778 // might schedule. _bb_end points just after last schedulable inst. We
2779 // normally schedule conditional branches (despite them being forced last
2780 // in the block), because they have delay slots we can fill. Calls all
2781 // have their delay slots filled in the template expansions, so we don't
2782 // bother scheduling them.
2783 Node *last = bb->get_node(_bb_end);
2784 // Ignore trailing NOPs.
2785 while (_bb_end > 0 && last->is_Mach() &&
2786 last->as_Mach()->ideal_Opcode() == Op_Con) {
2787 last = bb->get_node(--_bb_end);
2788 }
2789 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2790 if( last->is_Catch() ||
2791 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2792 // There might be a prior call. Skip it.
2793 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2794 } else if( last->is_MachNullCheck() ) {
2795 // Backup so the last null-checked memory instruction is
2796 // outside the schedulable range. Skip over the nullcheck,
2797 // projection, and the memory nodes.
2798 Node *mem = last->in(1);
2799 do {
2800 _bb_end--;
2801 } while (mem != bb->get_node(_bb_end));
2802 } else {
2803 // Set _bb_end to point after last schedulable inst.
2804 _bb_end++;
2805 }
2806
2807 assert( _bb_start <= _bb_end, "inverted block ends" );
2808
2809 // Compute the register antidependencies for the basic block
2810 ComputeRegisterAntidependencies(bb);
2811 if (C->failing()) return; // too many D-U pinch points
2812
2813 // Compute the usage within the block, and set the list of all nodes
2814 // in the block that have no uses within the block.
2815 ComputeUseCount(bb);
2816
2817 // Schedule the remaining instructions in the block
2818 while ( _available.size() > 0 ) {
2819 Node *n = ChooseNodeToBundle();
2820 guarantee(n != nullptr, "no nodes available");
2821 AddNodeToBundle(n,bb);
2822 }
2823
2824 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2825 #ifdef ASSERT
2826 for( uint l = _bb_start; l < _bb_end; l++ ) {
2827 Node *n = bb->get_node(l);
2828 uint m;
2829 for( m = 0; m < _bb_end-_bb_start; m++ )
2830 if( _scheduled[m] == n )
2831 break;
2832 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2833 }
2834 #endif
2835
2836 // Now copy the instructions (in reverse order) back to the block
2837 for ( uint k = _bb_start; k < _bb_end; k++ )
2838 bb->map_node(_scheduled[_bb_end-k-1], k);
2839
2840 #ifndef PRODUCT
2841 if (_cfg->C->trace_opto_output()) {
2842 tty->print("# Schedule BB#%03d (final)\n", i);
2843 uint current = 0;
2844 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2845 Node *n = bb->get_node(j);
2846 if( valid_bundle_info(n) ) {
2847 Bundle *bundle = node_bundling(n);
2848 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2849 tty->print("*** Bundle: ");
2850 bundle->dump();
2851 }
2852 n->dump();
2853 }
2854 }
2855 }
2856 #endif
2857 #ifdef ASSERT
2858 verify_good_schedule(bb,"after block local scheduling");
2859 #endif
2860 }
2861
2862 #ifndef PRODUCT
2863 if (_cfg->C->trace_opto_output())
2864 tty->print("# <- DoScheduling\n");
2865 #endif
2866
2867 // Record final node-bundling array location
2868 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2869
2870 } // end DoScheduling
2871
2872 // Verify that no live-range used in the block is killed in the block by a
2873 // wrong DEF. This doesn't verify live-ranges that span blocks.
2874
2875 // Check for edge existence. Used to avoid adding redundant precedence edges.
2876 static bool edge_from_to( Node *from, Node *to ) {
2877 for( uint i=0; i<from->len(); i++ )
2878 if( from->in(i) == to )
2879 return true;
2880 return false;
2881 }
2882
2883 #ifdef ASSERT
2884 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2885 // Check for bad kills
2886 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2887 Node *prior_use = _reg_node[def];
2888 if( prior_use && !edge_from_to(prior_use,n) ) {
2889 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2890 n->dump();
2891 tty->print_cr("...");
2892 prior_use->dump();
2893 assert(edge_from_to(prior_use,n), "%s", msg);
2894 }
2895 _reg_node.map(def,nullptr); // Kill live USEs
2896 }
2897 }
2898
2899 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2900
2901 // Zap to something reasonable for the verify code
2902 _reg_node.clear();
2903
2904 // Walk over the block backwards. Check to make sure each DEF doesn't
2905 // kill a live value (other than the one it's supposed to). Add each
2906 // USE to the live set.
2907 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2908 Node *n = b->get_node(i);
2909 int n_op = n->Opcode();
2910 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2911 // Fat-proj kills a slew of registers
2912 RegMaskIterator rmi(n->out_RegMask());
2913 while (rmi.has_next()) {
2914 OptoReg::Name kill = rmi.next();
2915 verify_do_def(n, kill, msg);
2916 }
2917 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2918 // Get DEF'd registers the normal way
2919 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2920 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2921 }
2922
2923 // Now make all USEs live
2924 for( uint i=1; i<n->req(); i++ ) {
2925 Node *def = n->in(i);
2926 assert(def != nullptr, "input edge required");
2927 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2928 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2929 if( OptoReg::is_valid(reg_lo) ) {
2930 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2931 _reg_node.map(reg_lo,n);
2932 }
2933 if( OptoReg::is_valid(reg_hi) ) {
2934 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2935 _reg_node.map(reg_hi,n);
2936 }
2937 }
2938
2939 }
2940
2941 // Zap to something reasonable for the Antidependence code
2942 _reg_node.clear();
2943 }
2944 #endif
2945
2946 // Conditionally add precedence edges. Avoid putting edges on Projs.
2947 static void add_prec_edge_from_to( Node *from, Node *to ) {
2948 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2949 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
2950 from = from->in(0);
2951 }
2952 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2953 !edge_from_to( from, to ) ) // Avoid duplicate edge
2954 from->add_prec(to);
2955 }
2956
2957 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2958 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2959 return;
2960
2961 if (OptoReg::is_reg(def_reg)) {
2962 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2963 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2964 // This is one of the high slots of a vector register.
2965 // ScheduleAndBundle already checked there are no live wide
2966 // vectors in this method so it can be safely ignored.
2967 return;
2968 }
2969 }
2970
2971 Node *pinch = _reg_node[def_reg]; // Get pinch point
2972 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2973 is_def ) { // Check for a true def (not a kill)
2974 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2975 return;
2976 }
2977
2978 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2979 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2980
2981 // After some number of kills there _may_ be a later def
2982 Node *later_def = nullptr;
2983
2984 Compile* C = Compile::current();
2985
2986 // Finding a kill requires a real pinch-point.
2987 // Check for not already having a pinch-point.
2988 // Pinch points are Op_Node's.
2989 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2990 later_def = pinch; // Must be def/kill as optimistic pinch-point
2991 if ( _pinch_free_list.size() > 0) {
2992 pinch = _pinch_free_list.pop();
2993 } else {
2994 pinch = new Node(1); // Pinch point to-be
2995 }
2996 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2997 DEBUG_ONLY( pinch->dump(); );
2998 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2999 _cfg->C->record_method_not_compilable("too many D-U pinch points");
3000 return;
3001 }
3002 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3003 _reg_node.map(def_reg,pinch); // Record pinch-point
3004 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3005 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3006 pinch->init_req(0, C->top()); // set not null for the next call
3007 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3008 later_def = nullptr; // and no later def
3009 }
3010 pinch->set_req(0,later_def); // Hook later def so we can find it
3011 } else { // Else have valid pinch point
3012 if( pinch->in(0) ) // If there is a later-def
3013 later_def = pinch->in(0); // Get it
3014 }
3015
3016 // Add output-dependence edge from later def to kill
3017 if( later_def ) // If there is some original def
3018 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3019
3020 // See if current kill is also a use, and so is forced to be the pinch-point.
3021 if( pinch->Opcode() == Op_Node ) {
3022 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3023 for( uint i=1; i<uses->req(); i++ ) {
3024 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3025 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3026 // Yes, found a use/kill pinch-point
3027 pinch->set_req(0,nullptr); //
3028 pinch->replace_by(kill); // Move anti-dep edges up
3029 pinch = kill;
3030 _reg_node.map(def_reg,pinch);
3031 return;
3032 }
3033 }
3034 }
3035
3036 // Add edge from kill to pinch-point
3037 add_prec_edge_from_to(kill,pinch);
3038 }
3039
3040 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3041 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3042 return;
3043 Node *pinch = _reg_node[use_reg]; // Get pinch point
3044 // Check for no later def_reg/kill in block
3045 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3046 // Use has to be block-local as well
3047 _cfg->get_block_for_node(use) == b) {
3048 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3049 pinch->req() == 1 ) { // pinch not yet in block?
3050 pinch->del_req(0); // yank pointer to later-def, also set flag
3051 // Insert the pinch-point in the block just after the last use
3052 b->insert_node(pinch, b->find_node(use) + 1);
3053 _bb_end++; // Increase size scheduled region in block
3054 }
3055
3056 add_prec_edge_from_to(pinch,use);
3057 }
3058 }
3059
3060 // We insert antidependences between the reads and following write of
3061 // allocated registers to prevent illegal code motion. Hopefully, the
3062 // number of added references should be fairly small, especially as we
3063 // are only adding references within the current basic block.
3064 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3065
3066 #ifdef ASSERT
3067 verify_good_schedule(b,"before block local scheduling");
3068 #endif
3069
3070 // A valid schedule, for each register independently, is an endless cycle
3071 // of: a def, then some uses (connected to the def by true dependencies),
3072 // then some kills (defs with no uses), finally the cycle repeats with a new
3073 // def. The uses are allowed to float relative to each other, as are the
3074 // kills. No use is allowed to slide past a kill (or def). This requires
3075 // antidependencies between all uses of a single def and all kills that
3076 // follow, up to the next def. More edges are redundant, because later defs
3077 // & kills are already serialized with true or antidependencies. To keep
3078 // the edge count down, we add a 'pinch point' node if there's more than
3079 // one use or more than one kill/def.
3080
3081 // We add dependencies in one bottom-up pass.
3082
3083 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3084
3085 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3086 // register. If not, we record the DEF/KILL in _reg_node, the
3087 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3088 // "pinch point", a new Node that's in the graph but not in the block.
3089 // We put edges from the prior and current DEF/KILLs to the pinch point.
3090 // We put the pinch point in _reg_node. If there's already a pinch point
3091 // we merely add an edge from the current DEF/KILL to the pinch point.
3092
3093 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3094 // put an edge from the pinch point to the USE.
3095
3096 // To be expedient, the _reg_node array is pre-allocated for the whole
3097 // compilation. _reg_node is lazily initialized; it either contains a null,
3098 // or a valid def/kill/pinch-point, or a leftover node from some prior
3099 // block. Leftover node from some prior block is treated like a null (no
3100 // prior def, so no anti-dependence needed). Valid def is distinguished by
3101 // it being in the current block.
3102 bool fat_proj_seen = false;
3103 uint last_safept = _bb_end-1;
3104 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3105 Node* last_safept_node = end_node;
3106 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3107 Node *n = b->get_node(i);
3108 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3109 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3110 // Fat-proj kills a slew of registers
3111 // This can add edges to 'n' and obscure whether or not it was a def,
3112 // hence the is_def flag.
3113 fat_proj_seen = true;
3114 RegMaskIterator rmi(n->out_RegMask());
3115 while (rmi.has_next()) {
3116 OptoReg::Name kill = rmi.next();
3117 anti_do_def(b, n, kill, is_def);
3118 }
3119 } else {
3120 // Get DEF'd registers the normal way
3121 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3122 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3123 }
3124
3125 // Kill projections on a branch should appear to occur on the
3126 // branch, not afterwards, so grab the masks from the projections
3127 // and process them.
3128 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3129 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3130 Node* use = n->fast_out(i);
3131 if (use->is_Proj()) {
3132 RegMaskIterator rmi(use->out_RegMask());
3133 while (rmi.has_next()) {
3134 OptoReg::Name kill = rmi.next();
3135 anti_do_def(b, n, kill, false);
3136 }
3137 }
3138 }
3139 }
3140
3141 // Check each register used by this instruction for a following DEF/KILL
3142 // that must occur afterward and requires an anti-dependence edge.
3143 for( uint j=0; j<n->req(); j++ ) {
3144 Node *def = n->in(j);
3145 if( def ) {
3146 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3147 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3148 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3149 }
3150 }
3151 // Do not allow defs of new derived values to float above GC
3152 // points unless the base is definitely available at the GC point.
3153
3154 Node *m = b->get_node(i);
3155
3156 // Add precedence edge from following safepoint to use of derived pointer
3157 if( last_safept_node != end_node &&
3158 m != last_safept_node) {
3159 for (uint k = 1; k < m->req(); k++) {
3160 const Type *t = m->in(k)->bottom_type();
3161 if( t->isa_oop_ptr() &&
3162 t->is_ptr()->offset() != 0 ) {
3163 last_safept_node->add_prec( m );
3164 break;
3165 }
3166 }
3167 }
3168
3169 if( n->jvms() ) { // Precedence edge from derived to safept
3170 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3171 if( b->get_node(last_safept) != last_safept_node ) {
3172 last_safept = b->find_node(last_safept_node);
3173 }
3174 for( uint j=last_safept; j > i; j-- ) {
3175 Node *mach = b->get_node(j);
3176 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3177 mach->add_prec( n );
3178 }
3179 last_safept = i;
3180 last_safept_node = m;
3181 }
3182 }
3183
3184 if (fat_proj_seen) {
3185 // Garbage collect pinch nodes that were not consumed.
3186 // They are usually created by a fat kill MachProj for a call.
3187 garbage_collect_pinch_nodes();
3188 }
3189 }
3190
3191 // Garbage collect pinch nodes for reuse by other blocks.
3192 //
3193 // The block scheduler's insertion of anti-dependence
3194 // edges creates many pinch nodes when the block contains
3195 // 2 or more Calls. A pinch node is used to prevent a
3196 // combinatorial explosion of edges. If a set of kills for a
3197 // register is anti-dependent on a set of uses (or defs), rather
3198 // than adding an edge in the graph between each pair of kill
3199 // and use (or def), a pinch is inserted between them:
3200 //
3201 // use1 use2 use3
3202 // \ | /
3203 // \ | /
3204 // pinch
3205 // / | \
3206 // / | \
3207 // kill1 kill2 kill3
3208 //
3209 // One pinch node is created per register killed when
3210 // the second call is encountered during a backwards pass
3211 // over the block. Most of these pinch nodes are never
3212 // wired into the graph because the register is never
3213 // used or def'ed in the block.
3214 //
3215 void Scheduling::garbage_collect_pinch_nodes() {
3216 #ifndef PRODUCT
3217 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3218 #endif
3219 int trace_cnt = 0;
3220 for (uint k = 0; k < _reg_node.max(); k++) {
3221 Node* pinch = _reg_node[k];
3222 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3223 // no predecence input edges
3224 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3225 cleanup_pinch(pinch);
3226 _pinch_free_list.push(pinch);
3227 _reg_node.map(k, nullptr);
3228 #ifndef PRODUCT
3229 if (_cfg->C->trace_opto_output()) {
3230 trace_cnt++;
3231 if (trace_cnt > 40) {
3232 tty->print("\n");
3233 trace_cnt = 0;
3234 }
3235 tty->print(" %d", pinch->_idx);
3236 }
3237 #endif
3238 }
3239 }
3240 #ifndef PRODUCT
3241 if (_cfg->C->trace_opto_output()) tty->print("\n");
3242 #endif
3243 }
3244
3245 // Clean up a pinch node for reuse.
3246 void Scheduling::cleanup_pinch( Node *pinch ) {
3247 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3248
3249 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3250 Node* use = pinch->last_out(i);
3251 uint uses_found = 0;
3252 for (uint j = use->req(); j < use->len(); j++) {
3253 if (use->in(j) == pinch) {
3254 use->rm_prec(j);
3255 uses_found++;
3256 }
3257 }
3258 assert(uses_found > 0, "must be a precedence edge");
3259 i -= uses_found; // we deleted 1 or more copies of this edge
3260 }
3261 // May have a later_def entry
3262 pinch->set_req(0, nullptr);
3263 }
3264
3265 #ifndef PRODUCT
3266
3267 void Scheduling::dump_available() const {
3268 tty->print("#Availist ");
3269 for (uint i = 0; i < _available.size(); i++)
3270 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3271 tty->cr();
3272 }
3273
3274 // Print Scheduling Statistics
3275 void Scheduling::print_statistics() {
3276 // Print the size added by nops for bundling
3277 tty->print("Nops added %d bytes to total of %d bytes",
3278 _total_nop_size, _total_method_size);
3279 if (_total_method_size > 0)
3280 tty->print(", for %.2f%%",
3281 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3282 tty->print("\n");
3283
3284 // Print the number of branch shadows filled
3285 if (Pipeline::_branch_has_delay_slot) {
3286 tty->print("Of %d branches, %d had unconditional delay slots filled",
3287 _total_branches, _total_unconditional_delays);
3288 if (_total_branches > 0)
3289 tty->print(", for %.2f%%",
3290 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3291 tty->print("\n");
3292 }
3293
3294 uint total_instructions = 0, total_bundles = 0;
3295
3296 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3297 uint bundle_count = _total_instructions_per_bundle[i];
3298 total_instructions += bundle_count * i;
3299 total_bundles += bundle_count;
3300 }
3301
3302 if (total_bundles > 0)
3303 tty->print("Average ILP (excluding nops) is %.2f\n",
3304 ((double)total_instructions) / ((double)total_bundles));
3305 }
3306 #endif
3307
3308 //-----------------------init_scratch_buffer_blob------------------------------
3309 // Construct a temporary BufferBlob and cache it for this compile.
3310 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3311 // If there is already a scratch buffer blob allocated and the
3312 // constant section is big enough, use it. Otherwise free the
3313 // current and allocate a new one.
3314 BufferBlob* blob = scratch_buffer_blob();
3315 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3316 // Use the current blob.
3317 } else {
3318 if (blob != nullptr) {
3319 BufferBlob::free(blob);
3320 }
3321
3322 ResourceMark rm;
3323 _scratch_const_size = const_size;
3324 int size = C2Compiler::initial_code_buffer_size(const_size);
3325 blob = BufferBlob::create("Compile::scratch_buffer", size);
3326 // Record the buffer blob for next time.
3327 set_scratch_buffer_blob(blob);
3328 // Have we run out of code space?
3329 if (scratch_buffer_blob() == nullptr) {
3330 // Let CompilerBroker disable further compilations.
3331 C->record_failure("Not enough space for scratch buffer in CodeCache");
3332 return;
3333 }
3334 }
3335
3336 // Initialize the relocation buffers
3337 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3338 set_scratch_locs_memory(locs_buf);
3339 }
3340
3341
3342 //-----------------------scratch_emit_size-------------------------------------
3343 // Helper function that computes size by emitting code
3344 uint PhaseOutput::scratch_emit_size(const Node* n) {
3345 // Start scratch_emit_size section.
3346 set_in_scratch_emit_size(true);
3347
3348 // Emit into a trash buffer and count bytes emitted.
3349 // This is a pretty expensive way to compute a size,
3350 // but it works well enough if seldom used.
3351 // All common fixed-size instructions are given a size
3352 // method by the AD file.
3353 // Note that the scratch buffer blob and locs memory are
3354 // allocated at the beginning of the compile task, and
3355 // may be shared by several calls to scratch_emit_size.
3356 // The allocation of the scratch buffer blob is particularly
3357 // expensive, since it has to grab the code cache lock.
3358 BufferBlob* blob = this->scratch_buffer_blob();
3359 assert(blob != nullptr, "Initialize BufferBlob at start");
3360 assert(blob->size() > MAX_inst_size, "sanity");
3361 relocInfo* locs_buf = scratch_locs_memory();
3362 address blob_begin = blob->content_begin();
3363 address blob_end = (address)locs_buf;
3364 assert(blob->contains(blob_end), "sanity");
3365 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3366 buf.initialize_consts_size(_scratch_const_size);
3367 buf.initialize_stubs_size(MAX_stubs_size);
3368 assert(locs_buf != nullptr, "sanity");
3369 int lsize = MAX_locs_size / 3;
3370 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3371 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3372 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3373 // Mark as scratch buffer.
3374 buf.consts()->set_scratch_emit();
3375 buf.insts()->set_scratch_emit();
3376 buf.stubs()->set_scratch_emit();
3377
3378 // Do the emission.
3379
3380 Label fakeL; // Fake label for branch instructions.
3381 Label* saveL = nullptr;
3382 uint save_bnum = 0;
3383 bool is_branch = n->is_MachBranch();
3384 C2_MacroAssembler masm(&buf);
3385 masm.bind(fakeL);
3386 if (is_branch) {
3387 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3388 n->as_MachBranch()->label_set(&fakeL, 0);
3389 }
3390 n->emit(&masm, C->regalloc());
3391
3392 // Emitting into the scratch buffer should not fail
3393 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3394
3395 if (is_branch) // Restore label.
3396 n->as_MachBranch()->label_set(saveL, save_bnum);
3397
3398 // End scratch_emit_size section.
3399 set_in_scratch_emit_size(false);
3400
3401 return buf.insts_size();
3402 }
3403
3404 void PhaseOutput::install() {
3405 if (!C->should_install_code()) {
3406 return;
3407 } else if (C->stub_function() != nullptr) {
3408 install_stub(C->stub_name());
3409 } else {
3410 install_code(C->method(),
3411 C->entry_bci(),
3412 CompileBroker::compiler2(),
3413 C->has_unsafe_access(),
3414 SharedRuntime::is_wide_vector(C->max_vector_size()));
3415 }
3416 }
3417
3418 void PhaseOutput::install_code(ciMethod* target,
3419 int entry_bci,
3420 AbstractCompiler* compiler,
3421 bool has_unsafe_access,
3422 bool has_wide_vectors) {
3423 // Check if we want to skip execution of all compiled code.
3424 {
3425 #ifndef PRODUCT
3426 if (OptoNoExecute) {
3427 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3428 return;
3429 }
3430 #endif
3431 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3432
3433 if (C->is_osr_compilation()) {
3434 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3435 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3436 } else {
3437 if (!target->is_static()) {
3438 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3439 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3440 // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3441 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3442 }
3443 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3444 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3445 }
3446
3447 C->env()->register_method(target,
3448 entry_bci,
3449 &_code_offsets,
3450 _orig_pc_slot_offset_in_bytes,
3451 code_buffer(),
3452 frame_size_in_words(),
3453 oop_map_set(),
3454 &_handler_table,
3455 inc_table(),
3456 compiler,
3457 has_unsafe_access,
3458 SharedRuntime::is_wide_vector(C->max_vector_size()),
3459 C->has_monitors(),
3460 C->has_scoped_access(),
3461 0);
3462
3463 if (C->log() != nullptr) { // Print code cache state into compiler log
3464 C->log()->code_cache_state();
3465 }
3466 }
3467 }
3468 void PhaseOutput::install_stub(const char* stub_name) {
3469 // Entry point will be accessed using stub_entry_point();
3470 if (code_buffer() == nullptr) {
3471 Matcher::soft_match_failure();
3472 } else {
3473 if (PrintAssembly && (WizardMode || Verbose))
3474 tty->print_cr("### Stub::%s", stub_name);
3475
3476 if (!C->failing()) {
3477 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3478
3479 // Make the NMethod
3480 // For now we mark the frame as never safe for profile stackwalking
3481 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3482 code_buffer(),
3483 CodeOffsets::frame_never_safe,
3484 // _code_offsets.value(CodeOffsets::Frame_Complete),
3485 frame_size_in_words(),
3486 oop_map_set(),
3487 false);
3488 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3489
3490 C->set_stub_entry_point(rs->entry_point());
3491 }
3492 }
3493 }
3494
3495 // Support for bundling info
3496 Bundle* PhaseOutput::node_bundling(const Node *n) {
3497 assert(valid_bundle_info(n), "oob");
3498 return &_node_bundling_base[n->_idx];
3499 }
3500
3501 bool PhaseOutput::valid_bundle_info(const Node *n) {
3502 return (_node_bundling_limit > n->_idx);
3503 }
3504
3505 //------------------------------frame_size_in_words-----------------------------
3506 // frame_slots in units of words
3507 int PhaseOutput::frame_size_in_words() const {
3508 // shift is 0 in LP32 and 1 in LP64
3509 const int shift = (LogBytesPerWord - LogBytesPerInt);
3510 int words = _frame_slots >> shift;
3511 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3512 return words;
3513 }
3514
3515 // To bang the stack of this compiled method we use the stack size
3516 // that the interpreter would need in case of a deoptimization. This
3517 // removes the need to bang the stack in the deoptimization blob which
3518 // in turn simplifies stack overflow handling.
3519 int PhaseOutput::bang_size_in_bytes() const {
3520 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3521 }
3522
3523 //------------------------------dump_asm---------------------------------------
3524 // Dump formatted assembly
3525 #if defined(SUPPORT_OPTO_ASSEMBLY)
3526 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3527
3528 int pc_digits = 3; // #chars required for pc
3529 int sb_chars = 3; // #chars for "start bundle" indicator
3530 int tab_size = 8;
3531 if (pcs != nullptr) {
3532 int max_pc = 0;
3533 for (uint i = 0; i < pc_limit; i++) {
3534 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3535 }
3536 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3537 }
3538 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3539
3540 bool cut_short = false;
3541 st->print_cr("#");
3542 st->print("# "); C->tf()->dump_on(st); st->cr();
3543 st->print_cr("#");
3544
3545 // For all blocks
3546 int pc = 0x0; // Program counter
3547 char starts_bundle = ' ';
3548 C->regalloc()->dump_frame();
3549
3550 Node *n = nullptr;
3551 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3552 if (VMThread::should_terminate()) {
3553 cut_short = true;
3554 break;
3555 }
3556 Block* block = C->cfg()->get_block(i);
3557 if (block->is_connector() && !Verbose) {
3558 continue;
3559 }
3560 n = block->head();
3561 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3562 pc = pcs[n->_idx];
3563 st->print("%*.*x", pc_digits, pc_digits, pc);
3564 }
3565 st->fill_to(prefix_len);
3566 block->dump_head(C->cfg(), st);
3567 if (block->is_connector()) {
3568 st->fill_to(prefix_len);
3569 st->print_cr("# Empty connector block");
3570 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3571 st->fill_to(prefix_len);
3572 st->print_cr("# Block is sole successor of call");
3573 }
3574
3575 // For all instructions
3576 Node *delay = nullptr;
3577 for (uint j = 0; j < block->number_of_nodes(); j++) {
3578 if (VMThread::should_terminate()) {
3579 cut_short = true;
3580 break;
3581 }
3582 n = block->get_node(j);
3583 if (valid_bundle_info(n)) {
3584 Bundle* bundle = node_bundling(n);
3585 if (bundle->used_in_unconditional_delay()) {
3586 delay = n;
3587 continue;
3588 }
3589 if (bundle->starts_bundle()) {
3590 starts_bundle = '+';
3591 }
3592 }
3593
3594 if (WizardMode) {
3595 n->dump();
3596 }
3597
3598 if( !n->is_Region() && // Dont print in the Assembly
3599 !n->is_Phi() && // a few noisely useless nodes
3600 !n->is_Proj() &&
3601 !n->is_MachTemp() &&
3602 !n->is_SafePointScalarObject() &&
3603 !n->is_Catch() && // Would be nice to print exception table targets
3604 !n->is_MergeMem() && // Not very interesting
3605 !n->is_top() && // Debug info table constants
3606 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3607 ) {
3608 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3609 pc = pcs[n->_idx];
3610 st->print("%*.*x", pc_digits, pc_digits, pc);
3611 } else {
3612 st->fill_to(pc_digits);
3613 }
3614 st->print(" %c ", starts_bundle);
3615 starts_bundle = ' ';
3616 st->fill_to(prefix_len);
3617 n->format(C->regalloc(), st);
3618 st->cr();
3619 }
3620
3621 // If we have an instruction with a delay slot, and have seen a delay,
3622 // then back up and print it
3623 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3624 // Coverity finding - Explicit null dereferenced.
3625 guarantee(delay != nullptr, "no unconditional delay instruction");
3626 if (WizardMode) delay->dump();
3627
3628 if (node_bundling(delay)->starts_bundle())
3629 starts_bundle = '+';
3630 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3631 pc = pcs[n->_idx];
3632 st->print("%*.*x", pc_digits, pc_digits, pc);
3633 } else {
3634 st->fill_to(pc_digits);
3635 }
3636 st->print(" %c ", starts_bundle);
3637 starts_bundle = ' ';
3638 st->fill_to(prefix_len);
3639 delay->format(C->regalloc(), st);
3640 st->cr();
3641 delay = nullptr;
3642 }
3643
3644 // Dump the exception table as well
3645 if( n->is_Catch() && (Verbose || WizardMode) ) {
3646 // Print the exception table for this offset
3647 _handler_table.print_subtable_for(pc);
3648 }
3649 st->bol(); // Make sure we start on a new line
3650 }
3651 st->cr(); // one empty line between blocks
3652 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3653 } // End of per-block dump
3654
3655 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3656 }
3657 #endif
3658
3659 #ifndef PRODUCT
3660 void PhaseOutput::print_statistics() {
3661 Scheduling::print_statistics();
3662 }
3663 #endif