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 fill_buffer(cb, blk_starts);
367 }
368
369 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
370 // Determine if we need to generate a stack overflow check.
371 // Do it if the method is not a stub function and
372 // has java calls or has frame size > vm_page_size/8.
373 // The debug VM checks that deoptimization doesn't trigger an
374 // unexpected stack overflow (compiled method stack banging should
375 // guarantee it doesn't happen) so we always need the stack bang in
376 // a debug VM.
377 return (C->stub_function() == nullptr &&
378 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
379 DEBUG_ONLY(|| true)));
380 }
381
382 bool PhaseOutput::need_register_stack_bang() const {
383 // Determine if we need to generate a register stack overflow check.
384 // This is only used on architectures which have split register
385 // and memory stacks (ie. IA64).
386 // Bang if the method is not a stub function and has java calls
387 return (C->stub_function() == nullptr && C->has_java_calls());
388 }
389
390
391 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
392 // of a loop. When aligning a loop we need to provide enough instructions
393 // in cpu's fetch buffer to feed decoders. The loop alignment could be
394 // avoided if we have enough instructions in fetch buffer at the head of a loop.
395 // By default, the size is set to 999999 by Block's constructor so that
396 // a loop will be aligned if the size is not reset here.
397 //
398 // Note: Mach instructions could contain several HW instructions
399 // so the size is estimated only.
400 //
401 void PhaseOutput::compute_loop_first_inst_sizes() {
402 // The next condition is used to gate the loop alignment optimization.
403 // Don't aligned a loop if there are enough instructions at the head of a loop
404 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
405 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
406 // equal to 11 bytes which is the largest address NOP instruction.
407 if (MaxLoopPad < OptoLoopAlignment - 1) {
408 uint last_block = C->cfg()->number_of_blocks() - 1;
409 for (uint i = 1; i <= last_block; i++) {
410 Block* block = C->cfg()->get_block(i);
411 // Check the first loop's block which requires an alignment.
412 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
413 uint sum_size = 0;
414 uint inst_cnt = NumberOfLoopInstrToAlign;
415 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
416
417 // Check subsequent fallthrough blocks if the loop's first
418 // block(s) does not have enough instructions.
419 Block *nb = block;
420 while(inst_cnt > 0 &&
421 i < last_block &&
422 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
423 !nb->has_successor(block)) {
424 i++;
425 nb = C->cfg()->get_block(i);
426 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
427 } // while( inst_cnt > 0 && i < last_block )
428
429 block->set_first_inst_size(sum_size);
430 } // f( b->head()->is_Loop() )
431 } // for( i <= last_block )
432 } // if( MaxLoopPad < OptoLoopAlignment-1 )
433 }
434
435 // The architecture description provides short branch variants for some long
436 // branch instructions. Replace eligible long branches with short branches.
437 void PhaseOutput::shorten_branches(uint* blk_starts) {
438
439 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
440
441 // Compute size of each block, method size, and relocation information size
442 uint nblocks = C->cfg()->number_of_blocks();
443
444 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
445 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
446 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
447
448 // Collect worst case block paddings
449 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
450 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
451
452 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
453 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
454
455 bool has_short_branch_candidate = false;
456
457 // Initialize the sizes to 0
458 int code_size = 0; // Size in bytes of generated code
459 int stub_size = 0; // Size in bytes of all stub entries
460 // Size in bytes of all relocation entries, including those in local stubs.
461 // Start with 2-bytes of reloc info for the unvalidated entry point
462 int reloc_size = 1; // Number of relocation entries
463
464 // Make three passes. The first computes pessimistic blk_starts,
465 // relative jmp_offset and reloc_size information. The second performs
466 // short branch substitution using the pessimistic sizing. The
467 // third inserts nops where needed.
468
469 // Step one, perform a pessimistic sizing pass.
470 uint last_call_adr = max_juint;
471 uint last_avoid_back_to_back_adr = max_juint;
472 uint nop_size = (new MachNopNode())->size(C->regalloc());
473 for (uint i = 0; i < nblocks; i++) { // For all blocks
474 Block* block = C->cfg()->get_block(i);
475 _block = block;
476
477 // During short branch replacement, we store the relative (to blk_starts)
478 // offset of jump in jmp_offset, rather than the absolute offset of jump.
479 // This is so that we do not need to recompute sizes of all nodes when
480 // we compute correct blk_starts in our next sizing pass.
481 jmp_offset[i] = 0;
482 jmp_size[i] = 0;
483 jmp_nidx[i] = -1;
484 DEBUG_ONLY( jmp_target[i] = 0; )
485 DEBUG_ONLY( jmp_rule[i] = 0; )
486
487 // Sum all instruction sizes to compute block size
488 uint last_inst = block->number_of_nodes();
489 uint blk_size = 0;
490 for (uint j = 0; j < last_inst; j++) {
491 _index = j;
492 Node* nj = block->get_node(_index);
493 // Handle machine instruction nodes
494 if (nj->is_Mach()) {
495 MachNode* mach = nj->as_Mach();
496 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
497 reloc_size += mach->reloc();
498 if (mach->is_MachCall()) {
499 // add size information for trampoline stub
500 // class CallStubImpl is platform-specific and defined in the *.ad files.
501 stub_size += CallStubImpl::size_call_trampoline();
502 reloc_size += CallStubImpl::reloc_call_trampoline();
503
504 MachCallNode *mcall = mach->as_MachCall();
505 // This destination address is NOT PC-relative
506
507 mcall->method_set((intptr_t)mcall->entry_point());
508
509 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
510 stub_size += CompiledDirectCall::to_interp_stub_size();
511 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
512 }
513 } else if (mach->is_MachSafePoint()) {
514 // If call/safepoint are adjacent, account for possible
515 // nop to disambiguate the two safepoints.
516 // ScheduleAndBundle() can rearrange nodes in a block,
517 // check for all offsets inside this block.
518 if (last_call_adr >= blk_starts[i]) {
519 blk_size += nop_size;
520 }
521 }
522 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
523 // Nop is inserted between "avoid back to back" instructions.
524 // ScheduleAndBundle() can rearrange nodes in a block,
525 // check for all offsets inside this block.
526 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
527 blk_size += nop_size;
528 }
529 }
530 if (mach->may_be_short_branch()) {
531 if (!nj->is_MachBranch()) {
532 #ifndef PRODUCT
533 nj->dump(3);
534 #endif
535 Unimplemented();
536 }
537 assert(jmp_nidx[i] == -1, "block should have only one branch");
538 jmp_offset[i] = blk_size;
539 jmp_size[i] = nj->size(C->regalloc());
540 jmp_nidx[i] = j;
541 has_short_branch_candidate = true;
542 }
543 }
544 blk_size += nj->size(C->regalloc());
545 // Remember end of call offset
546 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
547 last_call_adr = blk_starts[i]+blk_size;
548 }
549 // Remember end of avoid_back_to_back offset
550 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
551 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
552 }
553 }
554
555 // When the next block starts a loop, we may insert pad NOP
556 // instructions. Since we cannot know our future alignment,
557 // assume the worst.
558 if (i < nblocks - 1) {
559 Block* nb = C->cfg()->get_block(i + 1);
560 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
561 if (max_loop_pad > 0) {
562 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
563 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
564 // If either is the last instruction in this block, bump by
565 // max_loop_pad in lock-step with blk_size, so sizing
566 // calculations in subsequent blocks still can conservatively
567 // detect that it may the last instruction in this block.
568 if (last_call_adr == blk_starts[i]+blk_size) {
569 last_call_adr += max_loop_pad;
570 }
571 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
572 last_avoid_back_to_back_adr += max_loop_pad;
573 }
574 blk_size += max_loop_pad;
575 block_worst_case_pad[i + 1] = max_loop_pad;
576 }
577 }
578
579 // Save block size; update total method size
580 blk_starts[i+1] = blk_starts[i]+blk_size;
581 }
582
583 // Step two, replace eligible long jumps.
584 bool progress = true;
585 uint last_may_be_short_branch_adr = max_juint;
586 while (has_short_branch_candidate && progress) {
587 progress = false;
588 has_short_branch_candidate = false;
589 int adjust_block_start = 0;
590 for (uint i = 0; i < nblocks; i++) {
591 Block* block = C->cfg()->get_block(i);
592 int idx = jmp_nidx[i];
593 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
594 if (mach != nullptr && mach->may_be_short_branch()) {
595 #ifdef ASSERT
596 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
597 int j;
598 // Find the branch; ignore trailing NOPs.
599 for (j = block->number_of_nodes()-1; j>=0; j--) {
600 Node* n = block->get_node(j);
601 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
602 break;
603 }
604 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
605 #endif
606 int br_size = jmp_size[i];
607 int br_offs = blk_starts[i] + jmp_offset[i];
608
609 // This requires the TRUE branch target be in succs[0]
610 uint bnum = block->non_connector_successor(0)->_pre_order;
611 int offset = blk_starts[bnum] - br_offs;
612 if (bnum > i) { // adjust following block's offset
613 offset -= adjust_block_start;
614 }
615
616 // This block can be a loop header, account for the padding
617 // in the previous block.
618 int block_padding = block_worst_case_pad[i];
619 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
620 // In the following code a nop could be inserted before
621 // the branch which will increase the backward distance.
622 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
623 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
624
625 if (needs_padding && offset <= 0)
626 offset -= nop_size;
627
628 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
629 // We've got a winner. Replace this branch.
630 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
631
632 // Update the jmp_size.
633 int new_size = replacement->size(C->regalloc());
634 int diff = br_size - new_size;
635 assert(diff >= (int)nop_size, "short_branch size should be smaller");
636 // Conservatively take into account padding between
637 // avoid_back_to_back branches. Previous branch could be
638 // converted into avoid_back_to_back branch during next
639 // rounds.
640 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
641 jmp_offset[i] += nop_size;
642 diff -= nop_size;
643 }
644 adjust_block_start += diff;
645 block->map_node(replacement, idx);
646 mach->subsume_by(replacement, C);
647 mach = replacement;
648 progress = true;
649
650 jmp_size[i] = new_size;
651 DEBUG_ONLY( jmp_target[i] = bnum; );
652 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
653 } else {
654 // The jump distance is not short, try again during next iteration.
655 has_short_branch_candidate = true;
656 }
657 } // (mach->may_be_short_branch())
658 if (mach != nullptr && (mach->may_be_short_branch() ||
659 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
660 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
661 }
662 blk_starts[i+1] -= adjust_block_start;
663 }
664 }
665
666 #ifdef ASSERT
667 for (uint i = 0; i < nblocks; i++) { // For all blocks
668 if (jmp_target[i] != 0) {
669 int br_size = jmp_size[i];
670 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
671 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
672 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]);
673 }
674 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
675 }
676 }
677 #endif
678
679 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
680 // after ScheduleAndBundle().
681
682 // ------------------
683 // Compute size for code buffer
684 code_size = blk_starts[nblocks];
685
686 // Relocation records
687 reloc_size += 1; // Relo entry for exception handler
688
689 // Adjust reloc_size to number of record of relocation info
690 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
691 // a relocation index.
692 // The CodeBuffer will expand the locs array if this estimate is too low.
693 reloc_size *= 10 / sizeof(relocInfo);
694
695 _buf_sizes._reloc = reloc_size;
696 _buf_sizes._code = code_size;
697 _buf_sizes._stub = stub_size;
698 }
699
700 //------------------------------FillLocArray-----------------------------------
701 // Create a bit of debug info and append it to the array. The mapping is from
702 // Java local or expression stack to constant, register or stack-slot. For
703 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
704 // entry has been taken care of and caller should skip it).
705 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
706 // This should never have accepted Bad before
707 assert(OptoReg::is_valid(regnum), "location must be valid");
708 return (OptoReg::is_reg(regnum))
709 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
710 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
711 }
712
713
714 ObjectValue*
715 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
716 for (int i = 0; i < objs->length(); i++) {
717 assert(objs->at(i)->is_object(), "corrupt object cache");
718 ObjectValue* sv = (ObjectValue*) objs->at(i);
719 if (sv->id() == id) {
720 return sv;
721 }
722 }
723 // Otherwise..
724 return nullptr;
725 }
726
727 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
728 ObjectValue* sv ) {
729 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
730 objs->append(sv);
731 }
732
733
734 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
735 GrowableArray<ScopeValue*> *array,
736 GrowableArray<ScopeValue*> *objs ) {
737 assert( local, "use _top instead of null" );
738 if (array->length() != idx) {
739 assert(array->length() == idx + 1, "Unexpected array count");
740 // Old functionality:
741 // return
742 // New functionality:
743 // Assert if the local is not top. In product mode let the new node
744 // override the old entry.
745 assert(local == C->top(), "LocArray collision");
746 if (local == C->top()) {
747 return;
748 }
749 array->pop();
750 }
751 const Type *t = local->bottom_type();
752
753 // Is it a safepoint scalar object node?
754 if (local->is_SafePointScalarObject()) {
755 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
756
757 ObjectValue* sv = (ObjectValue*) sv_for_node_id(objs, spobj->_idx);
758 if (sv == nullptr) {
759 ciKlass* cik = t->is_oopptr()->exact_klass();
760 assert(cik->is_instance_klass() ||
761 cik->is_array_klass(), "Not supported allocation.");
762 sv = new ObjectValue(spobj->_idx,
763 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
764 set_sv_for_object_node(objs, sv);
765
766 uint first_ind = spobj->first_index(sfpt->jvms());
767 for (uint i = 0; i < spobj->n_fields(); i++) {
768 Node* fld_node = sfpt->in(first_ind+i);
769 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
770 }
771 }
772 array->append(sv);
773 return;
774 } else if (local->is_SafePointScalarMerge()) {
775 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
776 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
777
778 if (mv == nullptr) {
779 GrowableArray<ScopeValue*> deps;
780
781 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
782 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
783 assert(deps.length() == 1, "missing value");
784
785 int selector_idx = smerge->selector_idx(sfpt->jvms());
786 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
787 assert(deps.length() == 2, "missing value");
788
789 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
790 set_sv_for_object_node(objs, mv);
791
792 for (uint i = 1; i < smerge->req(); i++) {
793 Node* obj_node = smerge->in(i);
794 (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs);
795 }
796 }
797 array->append(mv);
798 return;
799 }
800
801 // Grab the register number for the local
802 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
803 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
804 // Record the double as two float registers.
805 // The register mask for such a value always specifies two adjacent
806 // float registers, with the lower register number even.
807 // Normally, the allocation of high and low words to these registers
808 // is irrelevant, because nearly all operations on register pairs
809 // (e.g., StoreD) treat them as a single unit.
810 // Here, we assume in addition that the words in these two registers
811 // stored "naturally" (by operations like StoreD and double stores
812 // within the interpreter) such that the lower-numbered register
813 // is written to the lower memory address. This may seem like
814 // a machine dependency, but it is not--it is a requirement on
815 // the author of the <arch>.ad file to ensure that, for every
816 // even/odd double-register pair to which a double may be allocated,
817 // the word in the even single-register is stored to the first
818 // memory word. (Note that register numbers are completely
819 // arbitrary, and are not tied to any machine-level encodings.)
820 #ifdef _LP64
821 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
822 array->append(new ConstantIntValue((jint)0));
823 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
824 } else if ( t->base() == Type::Long ) {
825 array->append(new ConstantIntValue((jint)0));
826 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
827 } else if ( t->base() == Type::RawPtr ) {
828 // jsr/ret return address which must be restored into the full
829 // width 64-bit stack slot.
830 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
831 }
832 #else //_LP64
833 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
834 // Repack the double/long as two jints.
835 // The convention the interpreter uses is that the second local
836 // holds the first raw word of the native double representation.
837 // This is actually reasonable, since locals and stack arrays
838 // grow downwards in all implementations.
839 // (If, on some machine, the interpreter's Java locals or stack
840 // were to grow upwards, the embedded doubles would be word-swapped.)
841 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
842 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
843 }
844 #endif //_LP64
845 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
846 OptoReg::is_reg(regnum) ) {
847 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
848 ? Location::float_in_dbl : Location::normal ));
849 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
850 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
851 ? Location::int_in_long : Location::normal ));
852 } else if( t->base() == Type::NarrowOop ) {
853 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
854 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
855 t->base() == Type::VectorD || t->base() == Type::VectorX ||
856 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
857 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
858 } else if (C->regalloc()->is_oop(local)) {
859 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
860 t->base() == Type::AryPtr,
861 "Unexpected type: %s", t->msg());
862 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
863 } else {
864 assert(t->base() == Type::Int || t->base() == Type::Half ||
865 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
866 "Unexpected type: %s", t->msg());
867 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
868 }
869 return;
870 }
871
872 // No register. It must be constant data.
873 switch (t->base()) {
874 case Type::Half: // Second half of a double
875 ShouldNotReachHere(); // Caller should skip 2nd halves
876 break;
877 case Type::AnyPtr:
878 array->append(new ConstantOopWriteValue(nullptr));
879 break;
880 case Type::AryPtr:
881 case Type::InstPtr: // fall through
882 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
883 break;
884 case Type::NarrowOop:
885 if (t == TypeNarrowOop::NULL_PTR) {
886 array->append(new ConstantOopWriteValue(nullptr));
887 } else {
888 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
889 }
890 break;
891 case Type::Int:
892 array->append(new ConstantIntValue(t->is_int()->get_con()));
893 break;
894 case Type::RawPtr:
895 // A return address (T_ADDRESS).
896 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
897 #ifdef _LP64
898 // Must be restored to the full-width 64-bit stack slot.
899 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
900 #else
901 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
902 #endif
903 break;
904 case Type::FloatCon: {
905 float f = t->is_float_constant()->getf();
906 array->append(new ConstantIntValue(jint_cast(f)));
907 break;
908 }
909 case Type::DoubleCon: {
910 jdouble d = t->is_double_constant()->getd();
911 #ifdef _LP64
912 array->append(new ConstantIntValue((jint)0));
913 array->append(new ConstantDoubleValue(d));
914 #else
915 // Repack the double as two jints.
916 // The convention the interpreter uses is that the second local
917 // holds the first raw word of the native double representation.
918 // This is actually reasonable, since locals and stack arrays
919 // grow downwards in all implementations.
920 // (If, on some machine, the interpreter's Java locals or stack
921 // were to grow upwards, the embedded doubles would be word-swapped.)
922 jlong_accessor acc;
923 acc.long_value = jlong_cast(d);
924 array->append(new ConstantIntValue(acc.words[1]));
925 array->append(new ConstantIntValue(acc.words[0]));
926 #endif
927 break;
928 }
929 case Type::Long: {
930 jlong d = t->is_long()->get_con();
931 #ifdef _LP64
932 array->append(new ConstantIntValue((jint)0));
933 array->append(new ConstantLongValue(d));
934 #else
935 // Repack the long as two jints.
936 // The convention the interpreter uses is that the second local
937 // holds the first raw word of the native double representation.
938 // This is actually reasonable, since locals and stack arrays
939 // grow downwards in all implementations.
940 // (If, on some machine, the interpreter's Java locals or stack
941 // were to grow upwards, the embedded doubles would be word-swapped.)
942 jlong_accessor acc;
943 acc.long_value = d;
944 array->append(new ConstantIntValue(acc.words[1]));
945 array->append(new ConstantIntValue(acc.words[0]));
946 #endif
947 break;
948 }
949 case Type::Top: // Add an illegal value here
950 array->append(new LocationValue(Location()));
951 break;
952 default:
953 ShouldNotReachHere();
954 break;
955 }
956 }
957
958 // Determine if this node starts a bundle
959 bool PhaseOutput::starts_bundle(const Node *n) const {
960 return (_node_bundling_limit > n->_idx &&
961 _node_bundling_base[n->_idx].starts_bundle());
962 }
963
964 // Determine if there is a monitor that has 'ov' as its owner.
965 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
966 for (int k = 0; k < monarray->length(); k++) {
967 MonitorValue* mv = monarray->at(k);
968 if (mv->owner() == ov) {
969 return true;
970 }
971 }
972
973 return false;
974 }
975
976 //--------------------------Process_OopMap_Node--------------------------------
977 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
978 // Handle special safepoint nodes for synchronization
979 MachSafePointNode *sfn = mach->as_MachSafePoint();
980 MachCallNode *mcall;
981
982 int safepoint_pc_offset = current_offset;
983 bool is_method_handle_invoke = false;
984 bool return_oop = false;
985 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
986 bool arg_escape = false;
987
988 // Add the safepoint in the DebugInfoRecorder
989 if( !mach->is_MachCall() ) {
990 mcall = nullptr;
991 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
992 } else {
993 mcall = mach->as_MachCall();
994
995 // Is the call a MethodHandle call?
996 if (mcall->is_MachCallJava()) {
997 if (mcall->as_MachCallJava()->_method_handle_invoke) {
998 assert(C->has_method_handle_invokes(), "must have been set during call generation");
999 is_method_handle_invoke = true;
1000 }
1001 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1002 }
1003
1004 // Check if a call returns an object.
1005 if (mcall->returns_pointer()) {
1006 return_oop = true;
1007 }
1008 safepoint_pc_offset += mcall->ret_addr_offset();
1009 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1010 }
1011
1012 // Loop over the JVMState list to add scope information
1013 // Do not skip safepoints with a null method, they need monitor info
1014 JVMState* youngest_jvms = sfn->jvms();
1015 int max_depth = youngest_jvms->depth();
1016
1017 // Allocate the object pool for scalar-replaced objects -- the map from
1018 // small-integer keys (which can be recorded in the local and ostack
1019 // arrays) to descriptions of the object state.
1020 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1021
1022 // Visit scopes from oldest to youngest.
1023 for (int depth = 1; depth <= max_depth; depth++) {
1024 JVMState* jvms = youngest_jvms->of_depth(depth);
1025 int idx;
1026 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1027 // Safepoints that do not have method() set only provide oop-map and monitor info
1028 // to support GC; these do not support deoptimization.
1029 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1030 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1031 int num_mon = jvms->nof_monitors();
1032 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1033 "JVMS local count must match that of the method");
1034
1035 // Add Local and Expression Stack Information
1036
1037 // Insert locals into the locarray
1038 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1039 for( idx = 0; idx < num_locs; idx++ ) {
1040 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1041 }
1042
1043 // Insert expression stack entries into the exparray
1044 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1045 for( idx = 0; idx < num_exps; idx++ ) {
1046 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1047 }
1048
1049 // Add in mappings of the monitors
1050 assert( !method ||
1051 !method->is_synchronized() ||
1052 method->is_native() ||
1053 num_mon > 0 ||
1054 !GenerateSynchronizationCode,
1055 "monitors must always exist for synchronized methods");
1056
1057 // Build the growable array of ScopeValues for exp stack
1058 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1059
1060 // Loop over monitors and insert into array
1061 for (idx = 0; idx < num_mon; idx++) {
1062 // Grab the node that defines this monitor
1063 Node* box_node = sfn->monitor_box(jvms, idx);
1064 Node* obj_node = sfn->monitor_obj(jvms, idx);
1065
1066 // Create ScopeValue for object
1067 ScopeValue *scval = nullptr;
1068
1069 if (obj_node->is_SafePointScalarObject()) {
1070 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1071 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1072 if (scval == nullptr) {
1073 const Type *t = spobj->bottom_type();
1074 ciKlass* cik = t->is_oopptr()->exact_klass();
1075 assert(cik->is_instance_klass() ||
1076 cik->is_array_klass(), "Not supported allocation.");
1077 ObjectValue* sv = new ObjectValue(spobj->_idx,
1078 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1079 PhaseOutput::set_sv_for_object_node(objs, sv);
1080
1081 uint first_ind = spobj->first_index(youngest_jvms);
1082 for (uint i = 0; i < spobj->n_fields(); i++) {
1083 Node* fld_node = sfn->in(first_ind+i);
1084 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1085 }
1086 scval = sv;
1087 }
1088 } else if (obj_node->is_SafePointScalarMerge()) {
1089 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1090 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1091
1092 if (mv == nullptr) {
1093 GrowableArray<ScopeValue*> deps;
1094
1095 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1096 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1097 assert(deps.length() == 1, "missing value");
1098
1099 int selector_idx = smerge->selector_idx(youngest_jvms);
1100 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1101 assert(deps.length() == 2, "missing value");
1102
1103 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1104 set_sv_for_object_node(objs, mv);
1105
1106 for (uint i = 1; i < smerge->req(); i++) {
1107 Node* obj_node = smerge->in(i);
1108 FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs);
1109 }
1110 }
1111 scval = mv;
1112 } else if (!obj_node->is_Con()) {
1113 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1114 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1115 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1116 } else {
1117 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1118 }
1119 } else {
1120 const TypePtr *tp = obj_node->get_ptr_type();
1121 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1122 }
1123
1124 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1125 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1126 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1127 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1128 }
1129
1130 // Mark ObjectValue nodes as root nodes if they are directly
1131 // referenced in the JVMS.
1132 for (int i = 0; i < objs->length(); i++) {
1133 ScopeValue* sv = objs->at(i);
1134 if (sv->is_object_merge()) {
1135 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1136
1137 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1138 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1139 bool is_root = locarray->contains(ov) || exparray->contains(ov) || contains_as_owner(monarray, ov);
1140 ov->set_root(is_root);
1141 }
1142 }
1143 }
1144
1145 // We dump the object pool first, since deoptimization reads it in first.
1146 C->debug_info()->dump_object_pool(objs);
1147
1148 // Build first class objects to pass to scope
1149 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1150 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1151 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1152
1153 // Make method available for all Safepoints
1154 ciMethod* scope_method = method ? method : C->method();
1155 // Describe the scope here
1156 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1157 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1158 // Now we can describe the scope.
1159 methodHandle null_mh;
1160 bool rethrow_exception = false;
1161 C->debug_info()->describe_scope(
1162 safepoint_pc_offset,
1163 null_mh,
1164 scope_method,
1165 jvms->bci(),
1166 jvms->should_reexecute(),
1167 rethrow_exception,
1168 is_method_handle_invoke,
1169 return_oop,
1170 has_ea_local_in_scope,
1171 arg_escape,
1172 locvals,
1173 expvals,
1174 monvals
1175 );
1176 } // End jvms loop
1177
1178 // Mark the end of the scope set.
1179 C->debug_info()->end_safepoint(safepoint_pc_offset);
1180 }
1181
1182
1183
1184 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1185 class NonSafepointEmitter {
1186 Compile* C;
1187 JVMState* _pending_jvms;
1188 int _pending_offset;
1189
1190 void emit_non_safepoint();
1191
1192 public:
1193 NonSafepointEmitter(Compile* compile) {
1194 this->C = compile;
1195 _pending_jvms = nullptr;
1196 _pending_offset = 0;
1197 }
1198
1199 void observe_instruction(Node* n, int pc_offset) {
1200 if (!C->debug_info()->recording_non_safepoints()) return;
1201
1202 Node_Notes* nn = C->node_notes_at(n->_idx);
1203 if (nn == nullptr || nn->jvms() == nullptr) return;
1204 if (_pending_jvms != nullptr &&
1205 _pending_jvms->same_calls_as(nn->jvms())) {
1206 // Repeated JVMS? Stretch it up here.
1207 _pending_offset = pc_offset;
1208 } else {
1209 if (_pending_jvms != nullptr &&
1210 _pending_offset < pc_offset) {
1211 emit_non_safepoint();
1212 }
1213 _pending_jvms = nullptr;
1214 if (pc_offset > C->debug_info()->last_pc_offset()) {
1215 // This is the only way _pending_jvms can become non-null:
1216 _pending_jvms = nn->jvms();
1217 _pending_offset = pc_offset;
1218 }
1219 }
1220 }
1221
1222 // Stay out of the way of real safepoints:
1223 void observe_safepoint(JVMState* jvms, int pc_offset) {
1224 if (_pending_jvms != nullptr &&
1225 !_pending_jvms->same_calls_as(jvms) &&
1226 _pending_offset < pc_offset) {
1227 emit_non_safepoint();
1228 }
1229 _pending_jvms = nullptr;
1230 }
1231
1232 void flush_at_end() {
1233 if (_pending_jvms != nullptr) {
1234 emit_non_safepoint();
1235 }
1236 _pending_jvms = nullptr;
1237 }
1238 };
1239
1240 void NonSafepointEmitter::emit_non_safepoint() {
1241 JVMState* youngest_jvms = _pending_jvms;
1242 int pc_offset = _pending_offset;
1243
1244 // Clear it now:
1245 _pending_jvms = nullptr;
1246
1247 DebugInformationRecorder* debug_info = C->debug_info();
1248 assert(debug_info->recording_non_safepoints(), "sanity");
1249
1250 debug_info->add_non_safepoint(pc_offset);
1251 int max_depth = youngest_jvms->depth();
1252
1253 // Visit scopes from oldest to youngest.
1254 for (int depth = 1; depth <= max_depth; depth++) {
1255 JVMState* jvms = youngest_jvms->of_depth(depth);
1256 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1257 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1258 methodHandle null_mh;
1259 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1260 }
1261
1262 // Mark the end of the scope set.
1263 debug_info->end_non_safepoint(pc_offset);
1264 }
1265
1266 //------------------------------init_buffer------------------------------------
1267 void PhaseOutput::estimate_buffer_size(int& const_req) {
1268
1269 // Set the initially allocated size
1270 const_req = initial_const_capacity;
1271
1272 // The extra spacing after the code is necessary on some platforms.
1273 // Sometimes we need to patch in a jump after the last instruction,
1274 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1275
1276 // Compute the byte offset where we can store the deopt pc.
1277 if (C->fixed_slots() != 0) {
1278 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1279 }
1280
1281 // Compute prolog code size
1282 _method_size = 0;
1283 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1284 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1285
1286 if (C->has_mach_constant_base_node()) {
1287 uint add_size = 0;
1288 // Fill the constant table.
1289 // Note: This must happen before shorten_branches.
1290 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1291 Block* b = C->cfg()->get_block(i);
1292
1293 for (uint j = 0; j < b->number_of_nodes(); j++) {
1294 Node* n = b->get_node(j);
1295
1296 // If the node is a MachConstantNode evaluate the constant
1297 // value section.
1298 if (n->is_MachConstant()) {
1299 MachConstantNode* machcon = n->as_MachConstant();
1300 machcon->eval_constant(C);
1301 } else if (n->is_Mach()) {
1302 // On Power there are more nodes that issue constants.
1303 add_size += (n->as_Mach()->ins_num_consts() * 8);
1304 }
1305 }
1306 }
1307
1308 // Calculate the offsets of the constants and the size of the
1309 // constant table (including the padding to the next section).
1310 constant_table().calculate_offsets_and_size();
1311 const_req = constant_table().size() + add_size;
1312 }
1313
1314 // Initialize the space for the BufferBlob used to find and verify
1315 // instruction size in MachNode::emit_size()
1316 init_scratch_buffer_blob(const_req);
1317 }
1318
1319 CodeBuffer* PhaseOutput::init_buffer() {
1320 int stub_req = _buf_sizes._stub;
1321 int code_req = _buf_sizes._code;
1322 int const_req = _buf_sizes._const;
1323
1324 int pad_req = NativeCall::instruction_size;
1325
1326 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1327 stub_req += bs->estimate_stub_size();
1328
1329 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1330 // class HandlerImpl is platform-specific and defined in the *.ad files.
1331 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1332 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1333 stub_req += MAX_stubs_size; // ensure per-stub margin
1334 code_req += MAX_inst_size; // ensure per-instruction margin
1335
1336 if (StressCodeBuffers)
1337 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1338
1339 int total_req =
1340 const_req +
1341 code_req +
1342 pad_req +
1343 stub_req +
1344 exception_handler_req +
1345 deopt_handler_req; // deopt handler
1346
1347 if (C->has_method_handle_invokes())
1348 total_req += deopt_handler_req; // deopt MH handler
1349
1350 CodeBuffer* cb = code_buffer();
1351 cb->initialize(total_req, _buf_sizes._reloc);
1352
1353 // Have we run out of code space?
1354 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1355 C->record_failure("CodeCache is full");
1356 return nullptr;
1357 }
1358 // Configure the code buffer.
1359 cb->initialize_consts_size(const_req);
1360 cb->initialize_stubs_size(stub_req);
1361 cb->initialize_oop_recorder(C->env()->oop_recorder());
1362
1363 // fill in the nop array for bundling computations
1364 MachNode *_nop_list[Bundle::_nop_count];
1365 Bundle::initialize_nops(_nop_list);
1366
1367 return cb;
1368 }
1369
1370 //------------------------------fill_buffer------------------------------------
1371 void PhaseOutput::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1372 // blk_starts[] contains offsets calculated during short branches processing,
1373 // offsets should not be increased during following steps.
1374
1375 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1376 // of a loop. It is used to determine the padding for loop alignment.
1377 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1378
1379 compute_loop_first_inst_sizes();
1380
1381 // Create oopmap set.
1382 _oop_map_set = new OopMapSet();
1383
1384 // !!!!! This preserves old handling of oopmaps for now
1385 C->debug_info()->set_oopmaps(_oop_map_set);
1386
1387 uint nblocks = C->cfg()->number_of_blocks();
1388 // Count and start of implicit null check instructions
1389 uint inct_cnt = 0;
1390 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1391
1392 // Count and start of calls
1393 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1394
1395 uint return_offset = 0;
1396 int nop_size = (new MachNopNode())->size(C->regalloc());
1397
1398 int previous_offset = 0;
1399 int current_offset = 0;
1400 int last_call_offset = -1;
1401 int last_avoid_back_to_back_offset = -1;
1402 #ifdef ASSERT
1403 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1404 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1405 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1406 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1407 #endif
1408
1409 // Create an array of unused labels, one for each basic block, if printing is enabled
1410 #if defined(SUPPORT_OPTO_ASSEMBLY)
1411 int* node_offsets = nullptr;
1412 uint node_offset_limit = C->unique();
1413
1414 if (C->print_assembly()) {
1415 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1416 }
1417 if (node_offsets != nullptr) {
1418 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1419 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1420 }
1421 #endif
1422
1423 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1424
1425 // Emit the constant table.
1426 if (C->has_mach_constant_base_node()) {
1427 if (!constant_table().emit(*cb)) {
1428 C->record_failure("consts section overflow");
1429 return;
1430 }
1431 }
1432
1433 // Create an array of labels, one for each basic block
1434 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1435 for (uint i = 0; i <= nblocks; i++) {
1436 blk_labels[i].init();
1437 }
1438
1439 // Now fill in the code buffer
1440 Node* delay_slot = nullptr;
1441 for (uint i = 0; i < nblocks; i++) {
1442 Block* block = C->cfg()->get_block(i);
1443 _block = block;
1444 Node* head = block->head();
1445
1446 // If this block needs to start aligned (i.e, can be reached other
1447 // than by falling-thru from the previous block), then force the
1448 // start of a new bundle.
1449 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1450 cb->flush_bundle(true);
1451 }
1452
1453 #ifdef ASSERT
1454 if (!block->is_connector()) {
1455 stringStream st;
1456 block->dump_head(C->cfg(), &st);
1457 MacroAssembler(cb).block_comment(st.freeze());
1458 }
1459 jmp_target[i] = 0;
1460 jmp_offset[i] = 0;
1461 jmp_size[i] = 0;
1462 jmp_rule[i] = 0;
1463 #endif
1464 int blk_offset = current_offset;
1465
1466 // Define the label at the beginning of the basic block
1467 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1468
1469 uint last_inst = block->number_of_nodes();
1470
1471 // Emit block normally, except for last instruction.
1472 // Emit means "dump code bits into code buffer".
1473 for (uint j = 0; j<last_inst; j++) {
1474 _index = j;
1475
1476 // Get the node
1477 Node* n = block->get_node(j);
1478
1479 // See if delay slots are supported
1480 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1481 assert(delay_slot == nullptr, "no use of delay slot node");
1482 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1483
1484 delay_slot = n;
1485 continue;
1486 }
1487
1488 // If this starts a new instruction group, then flush the current one
1489 // (but allow split bundles)
1490 if (Pipeline::requires_bundling() && starts_bundle(n))
1491 cb->flush_bundle(false);
1492
1493 // Special handling for SafePoint/Call Nodes
1494 bool is_mcall = false;
1495 if (n->is_Mach()) {
1496 MachNode *mach = n->as_Mach();
1497 is_mcall = n->is_MachCall();
1498 bool is_sfn = n->is_MachSafePoint();
1499
1500 // If this requires all previous instructions be flushed, then do so
1501 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1502 cb->flush_bundle(true);
1503 current_offset = cb->insts_size();
1504 }
1505
1506 // A padding may be needed again since a previous instruction
1507 // could be moved to delay slot.
1508
1509 // align the instruction if necessary
1510 int padding = mach->compute_padding(current_offset);
1511 // Make sure safepoint node for polling is distinct from a call's
1512 // return by adding a nop if needed.
1513 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1514 padding = nop_size;
1515 }
1516 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1517 current_offset == last_avoid_back_to_back_offset) {
1518 // Avoid back to back some instructions.
1519 padding = nop_size;
1520 }
1521
1522 if (padding > 0) {
1523 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1524 int nops_cnt = padding / nop_size;
1525 MachNode *nop = new MachNopNode(nops_cnt);
1526 block->insert_node(nop, j++);
1527 last_inst++;
1528 C->cfg()->map_node_to_block(nop, block);
1529 // Ensure enough space.
1530 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1531 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1532 C->record_failure("CodeCache is full");
1533 return;
1534 }
1535 nop->emit(*cb, C->regalloc());
1536 cb->flush_bundle(true);
1537 current_offset = cb->insts_size();
1538 }
1539
1540 bool observe_safepoint = is_sfn;
1541 // Remember the start of the last call in a basic block
1542 if (is_mcall) {
1543 MachCallNode *mcall = mach->as_MachCall();
1544
1545 // This destination address is NOT PC-relative
1546 mcall->method_set((intptr_t)mcall->entry_point());
1547
1548 // Save the return address
1549 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1550
1551 observe_safepoint = mcall->guaranteed_safepoint();
1552 }
1553
1554 // sfn will be valid whenever mcall is valid now because of inheritance
1555 if (observe_safepoint) {
1556 // Handle special safepoint nodes for synchronization
1557 if (!is_mcall) {
1558 MachSafePointNode *sfn = mach->as_MachSafePoint();
1559 // !!!!! Stubs only need an oopmap right now, so bail out
1560 if (sfn->jvms()->method() == nullptr) {
1561 // Write the oopmap directly to the code blob??!!
1562 continue;
1563 }
1564 } // End synchronization
1565
1566 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1567 current_offset);
1568 Process_OopMap_Node(mach, current_offset);
1569 } // End if safepoint
1570
1571 // If this is a null check, then add the start of the previous instruction to the list
1572 else if( mach->is_MachNullCheck() ) {
1573 inct_starts[inct_cnt++] = previous_offset;
1574 }
1575
1576 // If this is a branch, then fill in the label with the target BB's label
1577 else if (mach->is_MachBranch()) {
1578 // This requires the TRUE branch target be in succs[0]
1579 uint block_num = block->non_connector_successor(0)->_pre_order;
1580
1581 // Try to replace long branch if delay slot is not used,
1582 // it is mostly for back branches since forward branch's
1583 // distance is not updated yet.
1584 bool delay_slot_is_used = valid_bundle_info(n) &&
1585 C->output()->node_bundling(n)->use_unconditional_delay();
1586 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1587 assert(delay_slot == nullptr, "not expecting delay slot node");
1588 int br_size = n->size(C->regalloc());
1589 int offset = blk_starts[block_num] - current_offset;
1590 if (block_num >= i) {
1591 // Current and following block's offset are not
1592 // finalized yet, adjust distance by the difference
1593 // between calculated and final offsets of current block.
1594 offset -= (blk_starts[i] - blk_offset);
1595 }
1596 // In the following code a nop could be inserted before
1597 // the branch which will increase the backward distance.
1598 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1599 if (needs_padding && offset <= 0)
1600 offset -= nop_size;
1601
1602 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1603 // We've got a winner. Replace this branch.
1604 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1605
1606 // Update the jmp_size.
1607 int new_size = replacement->size(C->regalloc());
1608 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1609 // Insert padding between avoid_back_to_back branches.
1610 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1611 MachNode *nop = new MachNopNode();
1612 block->insert_node(nop, j++);
1613 C->cfg()->map_node_to_block(nop, block);
1614 last_inst++;
1615 nop->emit(*cb, C->regalloc());
1616 cb->flush_bundle(true);
1617 current_offset = cb->insts_size();
1618 }
1619 #ifdef ASSERT
1620 jmp_target[i] = block_num;
1621 jmp_offset[i] = current_offset - blk_offset;
1622 jmp_size[i] = new_size;
1623 jmp_rule[i] = mach->rule();
1624 #endif
1625 block->map_node(replacement, j);
1626 mach->subsume_by(replacement, C);
1627 n = replacement;
1628 mach = replacement;
1629 }
1630 }
1631 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1632 } else if (mach->ideal_Opcode() == Op_Jump) {
1633 for (uint h = 0; h < block->_num_succs; h++) {
1634 Block* succs_block = block->_succs[h];
1635 for (uint j = 1; j < succs_block->num_preds(); j++) {
1636 Node* jpn = succs_block->pred(j);
1637 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1638 uint block_num = succs_block->non_connector()->_pre_order;
1639 Label *blkLabel = &blk_labels[block_num];
1640 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1641 }
1642 }
1643 }
1644 }
1645 #ifdef ASSERT
1646 // Check that oop-store precedes the card-mark
1647 else if (mach->ideal_Opcode() == Op_StoreCM) {
1648 uint storeCM_idx = j;
1649 int count = 0;
1650 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1651 Node *oop_store = mach->in(prec); // Precedence edge
1652 if (oop_store == nullptr) continue;
1653 count++;
1654 uint i4;
1655 for (i4 = 0; i4 < last_inst; ++i4) {
1656 if (block->get_node(i4) == oop_store) {
1657 break;
1658 }
1659 }
1660 // Note: This test can provide a false failure if other precedence
1661 // edges have been added to the storeCMNode.
1662 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1663 }
1664 assert(count > 0, "storeCM expects at least one precedence edge");
1665 }
1666 #endif
1667 else if (!n->is_Proj()) {
1668 // Remember the beginning of the previous instruction, in case
1669 // it's followed by a flag-kill and a null-check. Happens on
1670 // Intel all the time, with add-to-memory kind of opcodes.
1671 previous_offset = current_offset;
1672 }
1673
1674 // Not an else-if!
1675 // If this is a trap based cmp then add its offset to the list.
1676 if (mach->is_TrapBasedCheckNode()) {
1677 inct_starts[inct_cnt++] = current_offset;
1678 }
1679 }
1680
1681 // Verify that there is sufficient space remaining
1682 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1683 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1684 C->record_failure("CodeCache is full");
1685 return;
1686 }
1687
1688 // Save the offset for the listing
1689 #if defined(SUPPORT_OPTO_ASSEMBLY)
1690 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1691 node_offsets[n->_idx] = cb->insts_size();
1692 }
1693 #endif
1694 assert(!C->failing(), "Should not reach here if failing.");
1695
1696 // "Normal" instruction case
1697 DEBUG_ONLY(uint instr_offset = cb->insts_size());
1698 n->emit(*cb, C->regalloc());
1699 current_offset = cb->insts_size();
1700
1701 // Above we only verified that there is enough space in the instruction section.
1702 // However, the instruction may emit stubs that cause code buffer expansion.
1703 // Bail out here if expansion failed due to a lack of code cache space.
1704 if (C->failing()) {
1705 return;
1706 }
1707
1708 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1709 "ret_addr_offset() not within emitted code");
1710
1711 #ifdef ASSERT
1712 uint n_size = n->size(C->regalloc());
1713 if (n_size < (current_offset-instr_offset)) {
1714 MachNode* mach = n->as_Mach();
1715 n->dump();
1716 mach->dump_format(C->regalloc(), tty);
1717 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1718 Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty);
1719 tty->print_cr(" ------------------- ");
1720 BufferBlob* blob = this->scratch_buffer_blob();
1721 address blob_begin = blob->content_begin();
1722 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1723 assert(false, "wrong size of mach node");
1724 }
1725 #endif
1726 non_safepoints.observe_instruction(n, current_offset);
1727
1728 // mcall is last "call" that can be a safepoint
1729 // record it so we can see if a poll will directly follow it
1730 // in which case we'll need a pad to make the PcDesc sites unique
1731 // see 5010568. This can be slightly inaccurate but conservative
1732 // in the case that return address is not actually at current_offset.
1733 // This is a small price to pay.
1734
1735 if (is_mcall) {
1736 last_call_offset = current_offset;
1737 }
1738
1739 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1740 // Avoid back to back some instructions.
1741 last_avoid_back_to_back_offset = current_offset;
1742 }
1743
1744 // See if this instruction has a delay slot
1745 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1746 guarantee(delay_slot != nullptr, "expecting delay slot node");
1747
1748 // Back up 1 instruction
1749 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1750
1751 // Save the offset for the listing
1752 #if defined(SUPPORT_OPTO_ASSEMBLY)
1753 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1754 node_offsets[delay_slot->_idx] = cb->insts_size();
1755 }
1756 #endif
1757
1758 // Support a SafePoint in the delay slot
1759 if (delay_slot->is_MachSafePoint()) {
1760 MachNode *mach = delay_slot->as_Mach();
1761 // !!!!! Stubs only need an oopmap right now, so bail out
1762 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1763 // Write the oopmap directly to the code blob??!!
1764 delay_slot = nullptr;
1765 continue;
1766 }
1767
1768 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1769 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1770 adjusted_offset);
1771 // Generate an OopMap entry
1772 Process_OopMap_Node(mach, adjusted_offset);
1773 }
1774
1775 // Insert the delay slot instruction
1776 delay_slot->emit(*cb, C->regalloc());
1777
1778 // Don't reuse it
1779 delay_slot = nullptr;
1780 }
1781
1782 } // End for all instructions in block
1783
1784 // If the next block is the top of a loop, pad this block out to align
1785 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1786 if (i < nblocks-1) {
1787 Block *nb = C->cfg()->get_block(i + 1);
1788 int padding = nb->alignment_padding(current_offset);
1789 if( padding > 0 ) {
1790 MachNode *nop = new MachNopNode(padding / nop_size);
1791 block->insert_node(nop, block->number_of_nodes());
1792 C->cfg()->map_node_to_block(nop, block);
1793 nop->emit(*cb, C->regalloc());
1794 current_offset = cb->insts_size();
1795 }
1796 }
1797 // Verify that the distance for generated before forward
1798 // short branches is still valid.
1799 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1800
1801 // Save new block start offset
1802 blk_starts[i] = blk_offset;
1803 } // End of for all blocks
1804 blk_starts[nblocks] = current_offset;
1805
1806 non_safepoints.flush_at_end();
1807
1808 // Offset too large?
1809 if (C->failing()) return;
1810
1811 // Define a pseudo-label at the end of the code
1812 MacroAssembler(cb).bind( blk_labels[nblocks] );
1813
1814 // Compute the size of the first block
1815 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1816
1817 #ifdef ASSERT
1818 for (uint i = 0; i < nblocks; i++) { // For all blocks
1819 if (jmp_target[i] != 0) {
1820 int br_size = jmp_size[i];
1821 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1822 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1823 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]);
1824 assert(false, "Displacement too large for short jmp");
1825 }
1826 }
1827 }
1828 #endif
1829
1830 if (!cb->finalize_stubs()) {
1831 C->record_failure("CodeCache is full");
1832 return;
1833 }
1834
1835 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1836 bs->emit_stubs(*cb);
1837 if (C->failing()) return;
1838
1839 // Fill in stubs.
1840 _stub_list.emit(*cb);
1841 if (C->failing()) return;
1842
1843 #ifndef PRODUCT
1844 // Information on the size of the method, without the extraneous code
1845 Scheduling::increment_method_size(cb->insts_size());
1846 #endif
1847
1848 // ------------------
1849 // Fill in exception table entries.
1850 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1851
1852 // Only java methods have exception handlers and deopt handlers
1853 // class HandlerImpl is platform-specific and defined in the *.ad files.
1854 if (C->method()) {
1855 // Emit the exception handler code.
1856 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1857 if (C->failing()) {
1858 return; // CodeBuffer::expand failed
1859 }
1860 // Emit the deopt handler code.
1861 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1862
1863 // Emit the MethodHandle deopt handler code (if required).
1864 if (C->has_method_handle_invokes() && !C->failing()) {
1865 // We can use the same code as for the normal deopt handler, we
1866 // just need a different entry point address.
1867 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1868 }
1869 }
1870
1871 // One last check for failed CodeBuffer::expand:
1872 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1873 C->record_failure("CodeCache is full");
1874 return;
1875 }
1876
1877 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1878 if (C->print_assembly()) {
1879 tty->cr();
1880 tty->print_cr("============================= C2-compiled nmethod ==============================");
1881 }
1882 #endif
1883
1884 #if defined(SUPPORT_OPTO_ASSEMBLY)
1885 // Dump the assembly code, including basic-block numbers
1886 if (C->print_assembly()) {
1887 ttyLocker ttyl; // keep the following output all in one block
1888 if (!VMThread::should_terminate()) { // test this under the tty lock
1889 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1890 // We call them first and write to a stringStream, then we retake the lock to
1891 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1892 ResourceMark rm;
1893 stringStream method_metadata_str;
1894 if (C->method() != nullptr) {
1895 C->method()->print_metadata(&method_metadata_str);
1896 }
1897 stringStream dump_asm_str;
1898 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1899
1900 NoSafepointVerifier nsv;
1901 ttyLocker ttyl2;
1902 // This output goes directly to the tty, not the compiler log.
1903 // To enable tools to match it up with the compilation activity,
1904 // be sure to tag this tty output with the compile ID.
1905 if (xtty != nullptr) {
1906 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1907 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1908 }
1909 if (C->method() != nullptr) {
1910 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1911 tty->print_raw(method_metadata_str.freeze());
1912 } else if (C->stub_name() != nullptr) {
1913 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1914 }
1915 tty->cr();
1916 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1917 tty->print_raw(dump_asm_str.freeze());
1918 tty->print_cr("--------------------------------------------------------------------------------");
1919 if (xtty != nullptr) {
1920 xtty->tail("opto_assembly");
1921 }
1922 }
1923 }
1924 #endif
1925 }
1926
1927 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1928 _inc_table.set_size(cnt);
1929
1930 uint inct_cnt = 0;
1931 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1932 Block* block = C->cfg()->get_block(i);
1933 Node *n = nullptr;
1934 int j;
1935
1936 // Find the branch; ignore trailing NOPs.
1937 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1938 n = block->get_node(j);
1939 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1940 break;
1941 }
1942 }
1943
1944 // If we didn't find anything, continue
1945 if (j < 0) {
1946 continue;
1947 }
1948
1949 // Compute ExceptionHandlerTable subtable entry and add it
1950 // (skip empty blocks)
1951 if (n->is_Catch()) {
1952
1953 // Get the offset of the return from the call
1954 uint call_return = call_returns[block->_pre_order];
1955 #ifdef ASSERT
1956 assert( call_return > 0, "no call seen for this basic block" );
1957 while (block->get_node(--j)->is_MachProj()) ;
1958 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1959 #endif
1960 // last instruction is a CatchNode, find it's CatchProjNodes
1961 int nof_succs = block->_num_succs;
1962 // allocate space
1963 GrowableArray<intptr_t> handler_bcis(nof_succs);
1964 GrowableArray<intptr_t> handler_pcos(nof_succs);
1965 // iterate through all successors
1966 for (int j = 0; j < nof_succs; j++) {
1967 Block* s = block->_succs[j];
1968 bool found_p = false;
1969 for (uint k = 1; k < s->num_preds(); k++) {
1970 Node* pk = s->pred(k);
1971 if (pk->is_CatchProj() && pk->in(0) == n) {
1972 const CatchProjNode* p = pk->as_CatchProj();
1973 found_p = true;
1974 // add the corresponding handler bci & pco information
1975 if (p->_con != CatchProjNode::fall_through_index) {
1976 // p leads to an exception handler (and is not fall through)
1977 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1978 // no duplicates, please
1979 if (!handler_bcis.contains(p->handler_bci())) {
1980 uint block_num = s->non_connector()->_pre_order;
1981 handler_bcis.append(p->handler_bci());
1982 handler_pcos.append(blk_labels[block_num].loc_pos());
1983 }
1984 }
1985 }
1986 }
1987 assert(found_p, "no matching predecessor found");
1988 // Note: Due to empty block removal, one block may have
1989 // several CatchProj inputs, from the same Catch.
1990 }
1991
1992 // Set the offset of the return from the call
1993 assert(handler_bcis.find(-1) != -1, "must have default handler");
1994 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
1995 continue;
1996 }
1997
1998 // Handle implicit null exception table updates
1999 if (n->is_MachNullCheck()) {
2000 uint block_num = block->non_connector_successor(0)->_pre_order;
2001 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2002 continue;
2003 }
2004 // Handle implicit exception table updates: trap instructions.
2005 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2006 uint block_num = block->non_connector_successor(0)->_pre_order;
2007 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2008 continue;
2009 }
2010 } // End of for all blocks fill in exception table entries
2011 }
2012
2013 // Static Variables
2014 #ifndef PRODUCT
2015 uint Scheduling::_total_nop_size = 0;
2016 uint Scheduling::_total_method_size = 0;
2017 uint Scheduling::_total_branches = 0;
2018 uint Scheduling::_total_unconditional_delays = 0;
2019 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2020 #endif
2021
2022 // Initializer for class Scheduling
2023
2024 Scheduling::Scheduling(Arena *arena, Compile &compile)
2025 : _arena(arena),
2026 _cfg(compile.cfg()),
2027 _regalloc(compile.regalloc()),
2028 _scheduled(arena),
2029 _available(arena),
2030 _reg_node(arena),
2031 _pinch_free_list(arena),
2032 _next_node(nullptr),
2033 _bundle_instr_count(0),
2034 _bundle_cycle_number(0),
2035 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2036 #ifndef PRODUCT
2037 , _branches(0)
2038 , _unconditional_delays(0)
2039 #endif
2040 {
2041 // Create a MachNopNode
2042 _nop = new MachNopNode();
2043
2044 // Now that the nops are in the array, save the count
2045 // (but allow entries for the nops)
2046 _node_bundling_limit = compile.unique();
2047 uint node_max = _regalloc->node_regs_max_index();
2048
2049 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2050
2051 // This one is persistent within the Compile class
2052 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2053
2054 // Allocate space for fixed-size arrays
2055 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2056 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2057
2058 // Clear the arrays
2059 for (uint i = 0; i < node_max; i++) {
2060 ::new (&_node_bundling_base[i]) Bundle();
2061 }
2062 memset(_uses, 0, node_max * sizeof(short));
2063 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2064
2065 // Clear the bundling information
2066 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2067
2068 // Get the last node
2069 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2070
2071 _next_node = block->get_node(block->number_of_nodes() - 1);
2072 }
2073
2074 #ifndef PRODUCT
2075 // Scheduling destructor
2076 Scheduling::~Scheduling() {
2077 _total_branches += _branches;
2078 _total_unconditional_delays += _unconditional_delays;
2079 }
2080 #endif
2081
2082 // Step ahead "i" cycles
2083 void Scheduling::step(uint i) {
2084
2085 Bundle *bundle = node_bundling(_next_node);
2086 bundle->set_starts_bundle();
2087
2088 // Update the bundle record, but leave the flags information alone
2089 if (_bundle_instr_count > 0) {
2090 bundle->set_instr_count(_bundle_instr_count);
2091 bundle->set_resources_used(_bundle_use.resourcesUsed());
2092 }
2093
2094 // Update the state information
2095 _bundle_instr_count = 0;
2096 _bundle_cycle_number += i;
2097 _bundle_use.step(i);
2098 }
2099
2100 void Scheduling::step_and_clear() {
2101 Bundle *bundle = node_bundling(_next_node);
2102 bundle->set_starts_bundle();
2103
2104 // Update the bundle record
2105 if (_bundle_instr_count > 0) {
2106 bundle->set_instr_count(_bundle_instr_count);
2107 bundle->set_resources_used(_bundle_use.resourcesUsed());
2108
2109 _bundle_cycle_number += 1;
2110 }
2111
2112 // Clear the bundling information
2113 _bundle_instr_count = 0;
2114 _bundle_use.reset();
2115
2116 memcpy(_bundle_use_elements,
2117 Pipeline_Use::elaborated_elements,
2118 sizeof(Pipeline_Use::elaborated_elements));
2119 }
2120
2121 // Perform instruction scheduling and bundling over the sequence of
2122 // instructions in backwards order.
2123 void PhaseOutput::ScheduleAndBundle() {
2124
2125 // Don't optimize this if it isn't a method
2126 if (!C->method())
2127 return;
2128
2129 // Don't optimize this if scheduling is disabled
2130 if (!C->do_scheduling())
2131 return;
2132
2133 // Scheduling code works only with pairs (8 bytes) maximum.
2134 // And when the scalable vector register is used, we may spill/unspill
2135 // the whole reg regardless of the max vector size.
2136 if (C->max_vector_size() > 8 ||
2137 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2138 return;
2139 }
2140
2141 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2142
2143 // Create a data structure for all the scheduling information
2144 Scheduling scheduling(Thread::current()->resource_area(), *C);
2145
2146 // Walk backwards over each basic block, computing the needed alignment
2147 // Walk over all the basic blocks
2148 scheduling.DoScheduling();
2149
2150 #ifndef PRODUCT
2151 if (C->trace_opto_output()) {
2152 // Buffer and print all at once
2153 ResourceMark rm;
2154 stringStream ss;
2155 ss.print("\n---- After ScheduleAndBundle ----\n");
2156 print_scheduling(&ss);
2157 tty->print("%s", ss.as_string());
2158 }
2159 #endif
2160 }
2161
2162 #ifndef PRODUCT
2163 // Separated out so that it can be called directly from debugger
2164 void PhaseOutput::print_scheduling() {
2165 print_scheduling(tty);
2166 }
2167
2168 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2169 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2170 output_stream->print("\nBB#%03d:\n", i);
2171 Block* block = C->cfg()->get_block(i);
2172 for (uint j = 0; j < block->number_of_nodes(); j++) {
2173 Node* n = block->get_node(j);
2174 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2175 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2176 n->dump("\n", false, output_stream);
2177 }
2178 }
2179 }
2180 #endif
2181
2182 // See if this node fits into the present instruction bundle
2183 bool Scheduling::NodeFitsInBundle(Node *n) {
2184 uint n_idx = n->_idx;
2185
2186 // If this is the unconditional delay instruction, then it fits
2187 if (n == _unconditional_delay_slot) {
2188 #ifndef PRODUCT
2189 if (_cfg->C->trace_opto_output())
2190 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2191 #endif
2192 return (true);
2193 }
2194
2195 // If the node cannot be scheduled this cycle, skip it
2196 if (_current_latency[n_idx] > _bundle_cycle_number) {
2197 #ifndef PRODUCT
2198 if (_cfg->C->trace_opto_output())
2199 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2200 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2201 #endif
2202 return (false);
2203 }
2204
2205 const Pipeline *node_pipeline = n->pipeline();
2206
2207 uint instruction_count = node_pipeline->instructionCount();
2208 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2209 instruction_count = 0;
2210 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2211 instruction_count++;
2212
2213 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2214 #ifndef PRODUCT
2215 if (_cfg->C->trace_opto_output())
2216 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2217 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2218 #endif
2219 return (false);
2220 }
2221
2222 // Don't allow non-machine nodes to be handled this way
2223 if (!n->is_Mach() && instruction_count == 0)
2224 return (false);
2225
2226 // See if there is any overlap
2227 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2228
2229 if (delay > 0) {
2230 #ifndef PRODUCT
2231 if (_cfg->C->trace_opto_output())
2232 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2233 #endif
2234 return false;
2235 }
2236
2237 #ifndef PRODUCT
2238 if (_cfg->C->trace_opto_output())
2239 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2240 #endif
2241
2242 return true;
2243 }
2244
2245 Node * Scheduling::ChooseNodeToBundle() {
2246 uint siz = _available.size();
2247
2248 if (siz == 0) {
2249
2250 #ifndef PRODUCT
2251 if (_cfg->C->trace_opto_output())
2252 tty->print("# ChooseNodeToBundle: null\n");
2253 #endif
2254 return (nullptr);
2255 }
2256
2257 // Fast path, if only 1 instruction in the bundle
2258 if (siz == 1) {
2259 #ifndef PRODUCT
2260 if (_cfg->C->trace_opto_output()) {
2261 tty->print("# ChooseNodeToBundle (only 1): ");
2262 _available[0]->dump();
2263 }
2264 #endif
2265 return (_available[0]);
2266 }
2267
2268 // Don't bother, if the bundle is already full
2269 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2270 for ( uint i = 0; i < siz; i++ ) {
2271 Node *n = _available[i];
2272
2273 // Skip projections, we'll handle them another way
2274 if (n->is_Proj())
2275 continue;
2276
2277 // This presupposed that instructions are inserted into the
2278 // available list in a legality order; i.e. instructions that
2279 // must be inserted first are at the head of the list
2280 if (NodeFitsInBundle(n)) {
2281 #ifndef PRODUCT
2282 if (_cfg->C->trace_opto_output()) {
2283 tty->print("# ChooseNodeToBundle: ");
2284 n->dump();
2285 }
2286 #endif
2287 return (n);
2288 }
2289 }
2290 }
2291
2292 // Nothing fits in this bundle, choose the highest priority
2293 #ifndef PRODUCT
2294 if (_cfg->C->trace_opto_output()) {
2295 tty->print("# ChooseNodeToBundle: ");
2296 _available[0]->dump();
2297 }
2298 #endif
2299
2300 return _available[0];
2301 }
2302
2303 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2304 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2305
2306 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2307 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2308 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2309 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2310
2311 // Comparison between stack -> reg and stack -> reg
2312 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2313 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2314 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2315 }
2316
2317 // Comparison between reg -> stack and reg -> stack
2318 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2319 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2320 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2321 }
2322
2323 return 0; // Not comparable
2324 }
2325
2326 void Scheduling::AddNodeToAvailableList(Node *n) {
2327 assert( !n->is_Proj(), "projections never directly made available" );
2328 #ifndef PRODUCT
2329 if (_cfg->C->trace_opto_output()) {
2330 tty->print("# AddNodeToAvailableList: ");
2331 n->dump();
2332 }
2333 #endif
2334
2335 int latency = _current_latency[n->_idx];
2336
2337 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2338 // for stack spilling or unspilling have the same latency, we sort
2339 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2340 // have more opportunities to do ld/st merging
2341 uint i;
2342 for (i = 0; i < _available.size(); i++) {
2343 if (_current_latency[_available[i]->_idx] > latency) {
2344 break;
2345 } else if (_current_latency[_available[i]->_idx] == latency &&
2346 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2347 compare_two_spill_nodes(n, _available[i]) > 0) {
2348 break;
2349 }
2350 }
2351
2352 // Special Check for compares following branches
2353 if( n->is_Mach() && _scheduled.size() > 0 ) {
2354 int op = n->as_Mach()->ideal_Opcode();
2355 Node *last = _scheduled[0];
2356 if( last->is_MachIf() && last->in(1) == n &&
2357 ( op == Op_CmpI ||
2358 op == Op_CmpU ||
2359 op == Op_CmpUL ||
2360 op == Op_CmpP ||
2361 op == Op_CmpF ||
2362 op == Op_CmpD ||
2363 op == Op_CmpL ) ) {
2364
2365 // Recalculate position, moving to front of same latency
2366 for ( i=0 ; i < _available.size(); i++ )
2367 if (_current_latency[_available[i]->_idx] >= latency)
2368 break;
2369 }
2370 }
2371
2372 // Insert the node in the available list
2373 _available.insert(i, n);
2374
2375 #ifndef PRODUCT
2376 if (_cfg->C->trace_opto_output())
2377 dump_available();
2378 #endif
2379 }
2380
2381 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2382 for ( uint i=0; i < n->len(); i++ ) {
2383 Node *def = n->in(i);
2384 if (!def) continue;
2385 if( def->is_Proj() ) // If this is a machine projection, then
2386 def = def->in(0); // propagate usage thru to the base instruction
2387
2388 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2389 continue;
2390 }
2391
2392 // Compute the latency
2393 uint l = _bundle_cycle_number + n->latency(i);
2394 if (_current_latency[def->_idx] < l)
2395 _current_latency[def->_idx] = l;
2396
2397 // If this does not have uses then schedule it
2398 if ((--_uses[def->_idx]) == 0)
2399 AddNodeToAvailableList(def);
2400 }
2401 }
2402
2403 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2404 #ifndef PRODUCT
2405 if (_cfg->C->trace_opto_output()) {
2406 tty->print("# AddNodeToBundle: ");
2407 n->dump();
2408 }
2409 #endif
2410
2411 // Remove this from the available list
2412 uint i;
2413 for (i = 0; i < _available.size(); i++)
2414 if (_available[i] == n)
2415 break;
2416 assert(i < _available.size(), "entry in _available list not found");
2417 _available.remove(i);
2418
2419 // See if this fits in the current bundle
2420 const Pipeline *node_pipeline = n->pipeline();
2421 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2422
2423 // Check for instructions to be placed in the delay slot. We
2424 // do this before we actually schedule the current instruction,
2425 // because the delay slot follows the current instruction.
2426 if (Pipeline::_branch_has_delay_slot &&
2427 node_pipeline->hasBranchDelay() &&
2428 !_unconditional_delay_slot) {
2429
2430 uint siz = _available.size();
2431
2432 // Conditional branches can support an instruction that
2433 // is unconditionally executed and not dependent by the
2434 // branch, OR a conditionally executed instruction if
2435 // the branch is taken. In practice, this means that
2436 // the first instruction at the branch target is
2437 // copied to the delay slot, and the branch goes to
2438 // the instruction after that at the branch target
2439 if ( n->is_MachBranch() ) {
2440
2441 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2442 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2443
2444 #ifndef PRODUCT
2445 _branches++;
2446 #endif
2447
2448 // At least 1 instruction is on the available list
2449 // that is not dependent on the branch
2450 for (uint i = 0; i < siz; i++) {
2451 Node *d = _available[i];
2452 const Pipeline *avail_pipeline = d->pipeline();
2453
2454 // Don't allow safepoints in the branch shadow, that will
2455 // cause a number of difficulties
2456 if ( avail_pipeline->instructionCount() == 1 &&
2457 !avail_pipeline->hasMultipleBundles() &&
2458 !avail_pipeline->hasBranchDelay() &&
2459 Pipeline::instr_has_unit_size() &&
2460 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2461 NodeFitsInBundle(d) &&
2462 !node_bundling(d)->used_in_delay()) {
2463
2464 if (d->is_Mach() && !d->is_MachSafePoint()) {
2465 // A node that fits in the delay slot was found, so we need to
2466 // set the appropriate bits in the bundle pipeline information so
2467 // that it correctly indicates resource usage. Later, when we
2468 // attempt to add this instruction to the bundle, we will skip
2469 // setting the resource usage.
2470 _unconditional_delay_slot = d;
2471 node_bundling(n)->set_use_unconditional_delay();
2472 node_bundling(d)->set_used_in_unconditional_delay();
2473 _bundle_use.add_usage(avail_pipeline->resourceUse());
2474 _current_latency[d->_idx] = _bundle_cycle_number;
2475 _next_node = d;
2476 ++_bundle_instr_count;
2477 #ifndef PRODUCT
2478 _unconditional_delays++;
2479 #endif
2480 break;
2481 }
2482 }
2483 }
2484 }
2485
2486 // No delay slot, add a nop to the usage
2487 if (!_unconditional_delay_slot) {
2488 // See if adding an instruction in the delay slot will overflow
2489 // the bundle.
2490 if (!NodeFitsInBundle(_nop)) {
2491 #ifndef PRODUCT
2492 if (_cfg->C->trace_opto_output())
2493 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2494 #endif
2495 step(1);
2496 }
2497
2498 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2499 _next_node = _nop;
2500 ++_bundle_instr_count;
2501 }
2502
2503 // See if the instruction in the delay slot requires a
2504 // step of the bundles
2505 if (!NodeFitsInBundle(n)) {
2506 #ifndef PRODUCT
2507 if (_cfg->C->trace_opto_output())
2508 tty->print("# *** STEP(branch won't fit) ***\n");
2509 #endif
2510 // Update the state information
2511 _bundle_instr_count = 0;
2512 _bundle_cycle_number += 1;
2513 _bundle_use.step(1);
2514 }
2515 }
2516
2517 // Get the number of instructions
2518 uint instruction_count = node_pipeline->instructionCount();
2519 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2520 instruction_count = 0;
2521
2522 // Compute the latency information
2523 uint delay = 0;
2524
2525 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2526 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2527 if (relative_latency < 0)
2528 relative_latency = 0;
2529
2530 delay = _bundle_use.full_latency(relative_latency, node_usage);
2531
2532 // Does not fit in this bundle, start a new one
2533 if (delay > 0) {
2534 step(delay);
2535
2536 #ifndef PRODUCT
2537 if (_cfg->C->trace_opto_output())
2538 tty->print("# *** STEP(%d) ***\n", delay);
2539 #endif
2540 }
2541 }
2542
2543 // If this was placed in the delay slot, ignore it
2544 if (n != _unconditional_delay_slot) {
2545
2546 if (delay == 0) {
2547 if (node_pipeline->hasMultipleBundles()) {
2548 #ifndef PRODUCT
2549 if (_cfg->C->trace_opto_output())
2550 tty->print("# *** STEP(multiple instructions) ***\n");
2551 #endif
2552 step(1);
2553 }
2554
2555 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2556 #ifndef PRODUCT
2557 if (_cfg->C->trace_opto_output())
2558 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2559 instruction_count + _bundle_instr_count,
2560 Pipeline::_max_instrs_per_cycle);
2561 #endif
2562 step(1);
2563 }
2564 }
2565
2566 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2567 _bundle_instr_count++;
2568
2569 // Set the node's latency
2570 _current_latency[n->_idx] = _bundle_cycle_number;
2571
2572 // Now merge the functional unit information
2573 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2574 _bundle_use.add_usage(node_usage);
2575
2576 // Increment the number of instructions in this bundle
2577 _bundle_instr_count += instruction_count;
2578
2579 // Remember this node for later
2580 if (n->is_Mach())
2581 _next_node = n;
2582 }
2583
2584 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2585 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2586 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2587 // into the block. All other scheduled nodes get put in the schedule here.
2588 int op = n->Opcode();
2589 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2590 (op != Op_Node && // Not an unused antidepedence node and
2591 // not an unallocated boxlock
2592 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2593
2594 // Push any trailing projections
2595 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2596 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2597 Node *foi = n->fast_out(i);
2598 if( foi->is_Proj() )
2599 _scheduled.push(foi);
2600 }
2601 }
2602
2603 // Put the instruction in the schedule list
2604 _scheduled.push(n);
2605 }
2606
2607 #ifndef PRODUCT
2608 if (_cfg->C->trace_opto_output())
2609 dump_available();
2610 #endif
2611
2612 // Walk all the definitions, decrementing use counts, and
2613 // if a definition has a 0 use count, place it in the available list.
2614 DecrementUseCounts(n,bb);
2615 }
2616
2617 // This method sets the use count within a basic block. We will ignore all
2618 // uses outside the current basic block. As we are doing a backwards walk,
2619 // any node we reach that has a use count of 0 may be scheduled. This also
2620 // avoids the problem of cyclic references from phi nodes, as long as phi
2621 // nodes are at the front of the basic block. This method also initializes
2622 // the available list to the set of instructions that have no uses within this
2623 // basic block.
2624 void Scheduling::ComputeUseCount(const Block *bb) {
2625 #ifndef PRODUCT
2626 if (_cfg->C->trace_opto_output())
2627 tty->print("# -> ComputeUseCount\n");
2628 #endif
2629
2630 // Clear the list of available and scheduled instructions, just in case
2631 _available.clear();
2632 _scheduled.clear();
2633
2634 // No delay slot specified
2635 _unconditional_delay_slot = nullptr;
2636
2637 #ifdef ASSERT
2638 for( uint i=0; i < bb->number_of_nodes(); i++ )
2639 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2640 #endif
2641
2642 // Force the _uses count to never go to zero for unscheduable pieces
2643 // of the block
2644 for( uint k = 0; k < _bb_start; k++ )
2645 _uses[bb->get_node(k)->_idx] = 1;
2646 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2647 _uses[bb->get_node(l)->_idx] = 1;
2648
2649 // Iterate backwards over the instructions in the block. Don't count the
2650 // branch projections at end or the block header instructions.
2651 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2652 Node *n = bb->get_node(j);
2653 if( n->is_Proj() ) continue; // Projections handled another way
2654
2655 // Account for all uses
2656 for ( uint k = 0; k < n->len(); k++ ) {
2657 Node *inp = n->in(k);
2658 if (!inp) continue;
2659 assert(inp != n, "no cycles allowed" );
2660 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2661 if (inp->is_Proj()) { // Skip through Proj's
2662 inp = inp->in(0);
2663 }
2664 ++_uses[inp->_idx]; // Count 1 block-local use
2665 }
2666 }
2667
2668 // If this instruction has a 0 use count, then it is available
2669 if (!_uses[n->_idx]) {
2670 _current_latency[n->_idx] = _bundle_cycle_number;
2671 AddNodeToAvailableList(n);
2672 }
2673
2674 #ifndef PRODUCT
2675 if (_cfg->C->trace_opto_output()) {
2676 tty->print("# uses: %3d: ", _uses[n->_idx]);
2677 n->dump();
2678 }
2679 #endif
2680 }
2681
2682 #ifndef PRODUCT
2683 if (_cfg->C->trace_opto_output())
2684 tty->print("# <- ComputeUseCount\n");
2685 #endif
2686 }
2687
2688 // This routine performs scheduling on each basic block in reverse order,
2689 // using instruction latencies and taking into account function unit
2690 // availability.
2691 void Scheduling::DoScheduling() {
2692 #ifndef PRODUCT
2693 if (_cfg->C->trace_opto_output())
2694 tty->print("# -> DoScheduling\n");
2695 #endif
2696
2697 Block *succ_bb = nullptr;
2698 Block *bb;
2699 Compile* C = Compile::current();
2700
2701 // Walk over all the basic blocks in reverse order
2702 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2703 bb = _cfg->get_block(i);
2704
2705 #ifndef PRODUCT
2706 if (_cfg->C->trace_opto_output()) {
2707 tty->print("# Schedule BB#%03d (initial)\n", i);
2708 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2709 bb->get_node(j)->dump();
2710 }
2711 }
2712 #endif
2713
2714 // On the head node, skip processing
2715 if (bb == _cfg->get_root_block()) {
2716 continue;
2717 }
2718
2719 // Skip empty, connector blocks
2720 if (bb->is_connector())
2721 continue;
2722
2723 // If the following block is not the sole successor of
2724 // this one, then reset the pipeline information
2725 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2726 #ifndef PRODUCT
2727 if (_cfg->C->trace_opto_output()) {
2728 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2729 _next_node->_idx, _bundle_instr_count);
2730 }
2731 #endif
2732 step_and_clear();
2733 }
2734
2735 // Leave untouched the starting instruction, any Phis, a CreateEx node
2736 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2737 _bb_end = bb->number_of_nodes()-1;
2738 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2739 Node *n = bb->get_node(_bb_start);
2740 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2741 // Also, MachIdealNodes do not get scheduled
2742 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2743 MachNode *mach = n->as_Mach();
2744 int iop = mach->ideal_Opcode();
2745 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2746 if( iop == Op_Con ) continue; // Do not schedule Top
2747 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2748 mach->pipeline() == MachNode::pipeline_class() &&
2749 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2750 continue;
2751 break; // Funny loop structure to be sure...
2752 }
2753 // Compute last "interesting" instruction in block - last instruction we
2754 // might schedule. _bb_end points just after last schedulable inst. We
2755 // normally schedule conditional branches (despite them being forced last
2756 // in the block), because they have delay slots we can fill. Calls all
2757 // have their delay slots filled in the template expansions, so we don't
2758 // bother scheduling them.
2759 Node *last = bb->get_node(_bb_end);
2760 // Ignore trailing NOPs.
2761 while (_bb_end > 0 && last->is_Mach() &&
2762 last->as_Mach()->ideal_Opcode() == Op_Con) {
2763 last = bb->get_node(--_bb_end);
2764 }
2765 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2766 if( last->is_Catch() ||
2767 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2768 // There might be a prior call. Skip it.
2769 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2770 } else if( last->is_MachNullCheck() ) {
2771 // Backup so the last null-checked memory instruction is
2772 // outside the schedulable range. Skip over the nullcheck,
2773 // projection, and the memory nodes.
2774 Node *mem = last->in(1);
2775 do {
2776 _bb_end--;
2777 } while (mem != bb->get_node(_bb_end));
2778 } else {
2779 // Set _bb_end to point after last schedulable inst.
2780 _bb_end++;
2781 }
2782
2783 assert( _bb_start <= _bb_end, "inverted block ends" );
2784
2785 // Compute the register antidependencies for the basic block
2786 ComputeRegisterAntidependencies(bb);
2787 if (C->failing()) return; // too many D-U pinch points
2788
2789 // Compute the usage within the block, and set the list of all nodes
2790 // in the block that have no uses within the block.
2791 ComputeUseCount(bb);
2792
2793 // Schedule the remaining instructions in the block
2794 while ( _available.size() > 0 ) {
2795 Node *n = ChooseNodeToBundle();
2796 guarantee(n != nullptr, "no nodes available");
2797 AddNodeToBundle(n,bb);
2798 }
2799
2800 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2801 #ifdef ASSERT
2802 for( uint l = _bb_start; l < _bb_end; l++ ) {
2803 Node *n = bb->get_node(l);
2804 uint m;
2805 for( m = 0; m < _bb_end-_bb_start; m++ )
2806 if( _scheduled[m] == n )
2807 break;
2808 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2809 }
2810 #endif
2811
2812 // Now copy the instructions (in reverse order) back to the block
2813 for ( uint k = _bb_start; k < _bb_end; k++ )
2814 bb->map_node(_scheduled[_bb_end-k-1], k);
2815
2816 #ifndef PRODUCT
2817 if (_cfg->C->trace_opto_output()) {
2818 tty->print("# Schedule BB#%03d (final)\n", i);
2819 uint current = 0;
2820 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2821 Node *n = bb->get_node(j);
2822 if( valid_bundle_info(n) ) {
2823 Bundle *bundle = node_bundling(n);
2824 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2825 tty->print("*** Bundle: ");
2826 bundle->dump();
2827 }
2828 n->dump();
2829 }
2830 }
2831 }
2832 #endif
2833 #ifdef ASSERT
2834 verify_good_schedule(bb,"after block local scheduling");
2835 #endif
2836 }
2837
2838 #ifndef PRODUCT
2839 if (_cfg->C->trace_opto_output())
2840 tty->print("# <- DoScheduling\n");
2841 #endif
2842
2843 // Record final node-bundling array location
2844 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2845
2846 } // end DoScheduling
2847
2848 // Verify that no live-range used in the block is killed in the block by a
2849 // wrong DEF. This doesn't verify live-ranges that span blocks.
2850
2851 // Check for edge existence. Used to avoid adding redundant precedence edges.
2852 static bool edge_from_to( Node *from, Node *to ) {
2853 for( uint i=0; i<from->len(); i++ )
2854 if( from->in(i) == to )
2855 return true;
2856 return false;
2857 }
2858
2859 #ifdef ASSERT
2860 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2861 // Check for bad kills
2862 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2863 Node *prior_use = _reg_node[def];
2864 if( prior_use && !edge_from_to(prior_use,n) ) {
2865 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2866 n->dump();
2867 tty->print_cr("...");
2868 prior_use->dump();
2869 assert(edge_from_to(prior_use,n), "%s", msg);
2870 }
2871 _reg_node.map(def,nullptr); // Kill live USEs
2872 }
2873 }
2874
2875 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2876
2877 // Zap to something reasonable for the verify code
2878 _reg_node.clear();
2879
2880 // Walk over the block backwards. Check to make sure each DEF doesn't
2881 // kill a live value (other than the one it's supposed to). Add each
2882 // USE to the live set.
2883 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2884 Node *n = b->get_node(i);
2885 int n_op = n->Opcode();
2886 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2887 // Fat-proj kills a slew of registers
2888 RegMaskIterator rmi(n->out_RegMask());
2889 while (rmi.has_next()) {
2890 OptoReg::Name kill = rmi.next();
2891 verify_do_def(n, kill, msg);
2892 }
2893 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2894 // Get DEF'd registers the normal way
2895 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2896 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2897 }
2898
2899 // Now make all USEs live
2900 for( uint i=1; i<n->req(); i++ ) {
2901 Node *def = n->in(i);
2902 assert(def != 0, "input edge required");
2903 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2904 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2905 if( OptoReg::is_valid(reg_lo) ) {
2906 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2907 _reg_node.map(reg_lo,n);
2908 }
2909 if( OptoReg::is_valid(reg_hi) ) {
2910 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2911 _reg_node.map(reg_hi,n);
2912 }
2913 }
2914
2915 }
2916
2917 // Zap to something reasonable for the Antidependence code
2918 _reg_node.clear();
2919 }
2920 #endif
2921
2922 // Conditionally add precedence edges. Avoid putting edges on Projs.
2923 static void add_prec_edge_from_to( Node *from, Node *to ) {
2924 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2925 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2926 from = from->in(0);
2927 }
2928 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2929 !edge_from_to( from, to ) ) // Avoid duplicate edge
2930 from->add_prec(to);
2931 }
2932
2933 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2934 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2935 return;
2936
2937 if (OptoReg::is_reg(def_reg)) {
2938 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2939 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2940 // This is one of the high slots of a vector register.
2941 // ScheduleAndBundle already checked there are no live wide
2942 // vectors in this method so it can be safely ignored.
2943 return;
2944 }
2945 }
2946
2947 Node *pinch = _reg_node[def_reg]; // Get pinch point
2948 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2949 is_def ) { // Check for a true def (not a kill)
2950 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2951 return;
2952 }
2953
2954 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2955 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2956
2957 // After some number of kills there _may_ be a later def
2958 Node *later_def = nullptr;
2959
2960 Compile* C = Compile::current();
2961
2962 // Finding a kill requires a real pinch-point.
2963 // Check for not already having a pinch-point.
2964 // Pinch points are Op_Node's.
2965 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2966 later_def = pinch; // Must be def/kill as optimistic pinch-point
2967 if ( _pinch_free_list.size() > 0) {
2968 pinch = _pinch_free_list.pop();
2969 } else {
2970 pinch = new Node(1); // Pinch point to-be
2971 }
2972 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2973 DEBUG_ONLY( pinch->dump(); );
2974 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
2975 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2976 return;
2977 }
2978 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2979 _reg_node.map(def_reg,pinch); // Record pinch-point
2980 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2981 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2982 pinch->init_req(0, C->top()); // set not null for the next call
2983 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2984 later_def = nullptr; // and no later def
2985 }
2986 pinch->set_req(0,later_def); // Hook later def so we can find it
2987 } else { // Else have valid pinch point
2988 if( pinch->in(0) ) // If there is a later-def
2989 later_def = pinch->in(0); // Get it
2990 }
2991
2992 // Add output-dependence edge from later def to kill
2993 if( later_def ) // If there is some original def
2994 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2995
2996 // See if current kill is also a use, and so is forced to be the pinch-point.
2997 if( pinch->Opcode() == Op_Node ) {
2998 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2999 for( uint i=1; i<uses->req(); i++ ) {
3000 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3001 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3002 // Yes, found a use/kill pinch-point
3003 pinch->set_req(0,nullptr); //
3004 pinch->replace_by(kill); // Move anti-dep edges up
3005 pinch = kill;
3006 _reg_node.map(def_reg,pinch);
3007 return;
3008 }
3009 }
3010 }
3011
3012 // Add edge from kill to pinch-point
3013 add_prec_edge_from_to(kill,pinch);
3014 }
3015
3016 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3017 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3018 return;
3019 Node *pinch = _reg_node[use_reg]; // Get pinch point
3020 // Check for no later def_reg/kill in block
3021 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3022 // Use has to be block-local as well
3023 _cfg->get_block_for_node(use) == b) {
3024 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3025 pinch->req() == 1 ) { // pinch not yet in block?
3026 pinch->del_req(0); // yank pointer to later-def, also set flag
3027 // Insert the pinch-point in the block just after the last use
3028 b->insert_node(pinch, b->find_node(use) + 1);
3029 _bb_end++; // Increase size scheduled region in block
3030 }
3031
3032 add_prec_edge_from_to(pinch,use);
3033 }
3034 }
3035
3036 // We insert antidependences between the reads and following write of
3037 // allocated registers to prevent illegal code motion. Hopefully, the
3038 // number of added references should be fairly small, especially as we
3039 // are only adding references within the current basic block.
3040 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3041
3042 #ifdef ASSERT
3043 verify_good_schedule(b,"before block local scheduling");
3044 #endif
3045
3046 // A valid schedule, for each register independently, is an endless cycle
3047 // of: a def, then some uses (connected to the def by true dependencies),
3048 // then some kills (defs with no uses), finally the cycle repeats with a new
3049 // def. The uses are allowed to float relative to each other, as are the
3050 // kills. No use is allowed to slide past a kill (or def). This requires
3051 // antidependencies between all uses of a single def and all kills that
3052 // follow, up to the next def. More edges are redundant, because later defs
3053 // & kills are already serialized with true or antidependencies. To keep
3054 // the edge count down, we add a 'pinch point' node if there's more than
3055 // one use or more than one kill/def.
3056
3057 // We add dependencies in one bottom-up pass.
3058
3059 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3060
3061 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3062 // register. If not, we record the DEF/KILL in _reg_node, the
3063 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3064 // "pinch point", a new Node that's in the graph but not in the block.
3065 // We put edges from the prior and current DEF/KILLs to the pinch point.
3066 // We put the pinch point in _reg_node. If there's already a pinch point
3067 // we merely add an edge from the current DEF/KILL to the pinch point.
3068
3069 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3070 // put an edge from the pinch point to the USE.
3071
3072 // To be expedient, the _reg_node array is pre-allocated for the whole
3073 // compilation. _reg_node is lazily initialized; it either contains a null,
3074 // or a valid def/kill/pinch-point, or a leftover node from some prior
3075 // block. Leftover node from some prior block is treated like a null (no
3076 // prior def, so no anti-dependence needed). Valid def is distinguished by
3077 // it being in the current block.
3078 bool fat_proj_seen = false;
3079 uint last_safept = _bb_end-1;
3080 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3081 Node* last_safept_node = end_node;
3082 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3083 Node *n = b->get_node(i);
3084 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3085 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3086 // Fat-proj kills a slew of registers
3087 // This can add edges to 'n' and obscure whether or not it was a def,
3088 // hence the is_def flag.
3089 fat_proj_seen = true;
3090 RegMaskIterator rmi(n->out_RegMask());
3091 while (rmi.has_next()) {
3092 OptoReg::Name kill = rmi.next();
3093 anti_do_def(b, n, kill, is_def);
3094 }
3095 } else {
3096 // Get DEF'd registers the normal way
3097 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3098 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3099 }
3100
3101 // Kill projections on a branch should appear to occur on the
3102 // branch, not afterwards, so grab the masks from the projections
3103 // and process them.
3104 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3105 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3106 Node* use = n->fast_out(i);
3107 if (use->is_Proj()) {
3108 RegMaskIterator rmi(use->out_RegMask());
3109 while (rmi.has_next()) {
3110 OptoReg::Name kill = rmi.next();
3111 anti_do_def(b, n, kill, false);
3112 }
3113 }
3114 }
3115 }
3116
3117 // Check each register used by this instruction for a following DEF/KILL
3118 // that must occur afterward and requires an anti-dependence edge.
3119 for( uint j=0; j<n->req(); j++ ) {
3120 Node *def = n->in(j);
3121 if( def ) {
3122 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3123 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3124 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3125 }
3126 }
3127 // Do not allow defs of new derived values to float above GC
3128 // points unless the base is definitely available at the GC point.
3129
3130 Node *m = b->get_node(i);
3131
3132 // Add precedence edge from following safepoint to use of derived pointer
3133 if( last_safept_node != end_node &&
3134 m != last_safept_node) {
3135 for (uint k = 1; k < m->req(); k++) {
3136 const Type *t = m->in(k)->bottom_type();
3137 if( t->isa_oop_ptr() &&
3138 t->is_ptr()->offset() != 0 ) {
3139 last_safept_node->add_prec( m );
3140 break;
3141 }
3142 }
3143 }
3144
3145 if( n->jvms() ) { // Precedence edge from derived to safept
3146 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3147 if( b->get_node(last_safept) != last_safept_node ) {
3148 last_safept = b->find_node(last_safept_node);
3149 }
3150 for( uint j=last_safept; j > i; j-- ) {
3151 Node *mach = b->get_node(j);
3152 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3153 mach->add_prec( n );
3154 }
3155 last_safept = i;
3156 last_safept_node = m;
3157 }
3158 }
3159
3160 if (fat_proj_seen) {
3161 // Garbage collect pinch nodes that were not consumed.
3162 // They are usually created by a fat kill MachProj for a call.
3163 garbage_collect_pinch_nodes();
3164 }
3165 }
3166
3167 // Garbage collect pinch nodes for reuse by other blocks.
3168 //
3169 // The block scheduler's insertion of anti-dependence
3170 // edges creates many pinch nodes when the block contains
3171 // 2 or more Calls. A pinch node is used to prevent a
3172 // combinatorial explosion of edges. If a set of kills for a
3173 // register is anti-dependent on a set of uses (or defs), rather
3174 // than adding an edge in the graph between each pair of kill
3175 // and use (or def), a pinch is inserted between them:
3176 //
3177 // use1 use2 use3
3178 // \ | /
3179 // \ | /
3180 // pinch
3181 // / | \
3182 // / | \
3183 // kill1 kill2 kill3
3184 //
3185 // One pinch node is created per register killed when
3186 // the second call is encountered during a backwards pass
3187 // over the block. Most of these pinch nodes are never
3188 // wired into the graph because the register is never
3189 // used or def'ed in the block.
3190 //
3191 void Scheduling::garbage_collect_pinch_nodes() {
3192 #ifndef PRODUCT
3193 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3194 #endif
3195 int trace_cnt = 0;
3196 for (uint k = 0; k < _reg_node.max(); k++) {
3197 Node* pinch = _reg_node[k];
3198 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3199 // no predecence input edges
3200 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3201 cleanup_pinch(pinch);
3202 _pinch_free_list.push(pinch);
3203 _reg_node.map(k, nullptr);
3204 #ifndef PRODUCT
3205 if (_cfg->C->trace_opto_output()) {
3206 trace_cnt++;
3207 if (trace_cnt > 40) {
3208 tty->print("\n");
3209 trace_cnt = 0;
3210 }
3211 tty->print(" %d", pinch->_idx);
3212 }
3213 #endif
3214 }
3215 }
3216 #ifndef PRODUCT
3217 if (_cfg->C->trace_opto_output()) tty->print("\n");
3218 #endif
3219 }
3220
3221 // Clean up a pinch node for reuse.
3222 void Scheduling::cleanup_pinch( Node *pinch ) {
3223 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3224
3225 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3226 Node* use = pinch->last_out(i);
3227 uint uses_found = 0;
3228 for (uint j = use->req(); j < use->len(); j++) {
3229 if (use->in(j) == pinch) {
3230 use->rm_prec(j);
3231 uses_found++;
3232 }
3233 }
3234 assert(uses_found > 0, "must be a precedence edge");
3235 i -= uses_found; // we deleted 1 or more copies of this edge
3236 }
3237 // May have a later_def entry
3238 pinch->set_req(0, nullptr);
3239 }
3240
3241 #ifndef PRODUCT
3242
3243 void Scheduling::dump_available() const {
3244 tty->print("#Availist ");
3245 for (uint i = 0; i < _available.size(); i++)
3246 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3247 tty->cr();
3248 }
3249
3250 // Print Scheduling Statistics
3251 void Scheduling::print_statistics() {
3252 // Print the size added by nops for bundling
3253 tty->print("Nops added %d bytes to total of %d bytes",
3254 _total_nop_size, _total_method_size);
3255 if (_total_method_size > 0)
3256 tty->print(", for %.2f%%",
3257 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3258 tty->print("\n");
3259
3260 // Print the number of branch shadows filled
3261 if (Pipeline::_branch_has_delay_slot) {
3262 tty->print("Of %d branches, %d had unconditional delay slots filled",
3263 _total_branches, _total_unconditional_delays);
3264 if (_total_branches > 0)
3265 tty->print(", for %.2f%%",
3266 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3267 tty->print("\n");
3268 }
3269
3270 uint total_instructions = 0, total_bundles = 0;
3271
3272 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3273 uint bundle_count = _total_instructions_per_bundle[i];
3274 total_instructions += bundle_count * i;
3275 total_bundles += bundle_count;
3276 }
3277
3278 if (total_bundles > 0)
3279 tty->print("Average ILP (excluding nops) is %.2f\n",
3280 ((double)total_instructions) / ((double)total_bundles));
3281 }
3282 #endif
3283
3284 //-----------------------init_scratch_buffer_blob------------------------------
3285 // Construct a temporary BufferBlob and cache it for this compile.
3286 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3287 // If there is already a scratch buffer blob allocated and the
3288 // constant section is big enough, use it. Otherwise free the
3289 // current and allocate a new one.
3290 BufferBlob* blob = scratch_buffer_blob();
3291 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3292 // Use the current blob.
3293 } else {
3294 if (blob != nullptr) {
3295 BufferBlob::free(blob);
3296 }
3297
3298 ResourceMark rm;
3299 _scratch_const_size = const_size;
3300 int size = C2Compiler::initial_code_buffer_size(const_size);
3301 blob = BufferBlob::create("Compile::scratch_buffer", size);
3302 // Record the buffer blob for next time.
3303 set_scratch_buffer_blob(blob);
3304 // Have we run out of code space?
3305 if (scratch_buffer_blob() == nullptr) {
3306 // Let CompilerBroker disable further compilations.
3307 C->record_failure("Not enough space for scratch buffer in CodeCache");
3308 return;
3309 }
3310 }
3311
3312 // Initialize the relocation buffers
3313 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3314 set_scratch_locs_memory(locs_buf);
3315 }
3316
3317
3318 //-----------------------scratch_emit_size-------------------------------------
3319 // Helper function that computes size by emitting code
3320 uint PhaseOutput::scratch_emit_size(const Node* n) {
3321 // Start scratch_emit_size section.
3322 set_in_scratch_emit_size(true);
3323
3324 // Emit into a trash buffer and count bytes emitted.
3325 // This is a pretty expensive way to compute a size,
3326 // but it works well enough if seldom used.
3327 // All common fixed-size instructions are given a size
3328 // method by the AD file.
3329 // Note that the scratch buffer blob and locs memory are
3330 // allocated at the beginning of the compile task, and
3331 // may be shared by several calls to scratch_emit_size.
3332 // The allocation of the scratch buffer blob is particularly
3333 // expensive, since it has to grab the code cache lock.
3334 BufferBlob* blob = this->scratch_buffer_blob();
3335 assert(blob != nullptr, "Initialize BufferBlob at start");
3336 assert(blob->size() > MAX_inst_size, "sanity");
3337 relocInfo* locs_buf = scratch_locs_memory();
3338 address blob_begin = blob->content_begin();
3339 address blob_end = (address)locs_buf;
3340 assert(blob->contains(blob_end), "sanity");
3341 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3342 buf.initialize_consts_size(_scratch_const_size);
3343 buf.initialize_stubs_size(MAX_stubs_size);
3344 assert(locs_buf != nullptr, "sanity");
3345 int lsize = MAX_locs_size / 3;
3346 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3347 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3348 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3349 // Mark as scratch buffer.
3350 buf.consts()->set_scratch_emit();
3351 buf.insts()->set_scratch_emit();
3352 buf.stubs()->set_scratch_emit();
3353
3354 // Do the emission.
3355
3356 Label fakeL; // Fake label for branch instructions.
3357 Label* saveL = nullptr;
3358 uint save_bnum = 0;
3359 bool is_branch = n->is_MachBranch();
3360 if (is_branch) {
3361 MacroAssembler masm(&buf);
3362 masm.bind(fakeL);
3363 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3364 n->as_MachBranch()->label_set(&fakeL, 0);
3365 }
3366 n->emit(buf, C->regalloc());
3367
3368 // Emitting into the scratch buffer should not fail
3369 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3370
3371 if (is_branch) // Restore label.
3372 n->as_MachBranch()->label_set(saveL, save_bnum);
3373
3374 // End scratch_emit_size section.
3375 set_in_scratch_emit_size(false);
3376
3377 return buf.insts_size();
3378 }
3379
3380 void PhaseOutput::install() {
3381 if (!C->should_install_code()) {
3382 return;
3383 } else if (C->stub_function() != nullptr) {
3384 install_stub(C->stub_name());
3385 } else {
3386 install_code(C->method(),
3387 C->entry_bci(),
3388 CompileBroker::compiler2(),
3389 C->has_unsafe_access(),
3390 SharedRuntime::is_wide_vector(C->max_vector_size()),
3391 C->rtm_state());
3392 }
3393 }
3394
3395 void PhaseOutput::install_code(ciMethod* target,
3396 int entry_bci,
3397 AbstractCompiler* compiler,
3398 bool has_unsafe_access,
3399 bool has_wide_vectors,
3400 RTMState rtm_state) {
3401 // Check if we want to skip execution of all compiled code.
3402 {
3403 #ifndef PRODUCT
3404 if (OptoNoExecute) {
3405 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3406 return;
3407 }
3408 #endif
3409 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3410
3411 if (C->is_osr_compilation()) {
3412 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3413 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3414 } else {
3415 if (!target->is_static()) {
3416 // The UEP of an nmethod ensures that the VEP is padded. However, the padding of the UEP is placed
3417 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
3418 // through the UEP, yet we can ensure that the VEP is aligned appropriately.
3419 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size - MacroAssembler::ic_check_size());
3420 }
3421 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3422 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3423 }
3424
3425 C->env()->register_method(target,
3426 entry_bci,
3427 &_code_offsets,
3428 _orig_pc_slot_offset_in_bytes,
3429 code_buffer(),
3430 frame_size_in_words(),
3431 oop_map_set(),
3432 &_handler_table,
3433 inc_table(),
3434 compiler,
3435 has_unsafe_access,
3436 SharedRuntime::is_wide_vector(C->max_vector_size()),
3437 C->has_monitors(),
3438 0,
3439 C->rtm_state());
3440
3441 if (C->log() != nullptr) { // Print code cache state into compiler log
3442 C->log()->code_cache_state();
3443 }
3444 }
3445 }
3446 void PhaseOutput::install_stub(const char* stub_name) {
3447 // Entry point will be accessed using stub_entry_point();
3448 if (code_buffer() == nullptr) {
3449 Matcher::soft_match_failure();
3450 } else {
3451 if (PrintAssembly && (WizardMode || Verbose))
3452 tty->print_cr("### Stub::%s", stub_name);
3453
3454 if (!C->failing()) {
3455 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3456
3457 // Make the NMethod
3458 // For now we mark the frame as never safe for profile stackwalking
3459 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3460 code_buffer(),
3461 CodeOffsets::frame_never_safe,
3462 // _code_offsets.value(CodeOffsets::Frame_Complete),
3463 frame_size_in_words(),
3464 oop_map_set(),
3465 false);
3466 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3467
3468 C->set_stub_entry_point(rs->entry_point());
3469 }
3470 }
3471 }
3472
3473 // Support for bundling info
3474 Bundle* PhaseOutput::node_bundling(const Node *n) {
3475 assert(valid_bundle_info(n), "oob");
3476 return &_node_bundling_base[n->_idx];
3477 }
3478
3479 bool PhaseOutput::valid_bundle_info(const Node *n) {
3480 return (_node_bundling_limit > n->_idx);
3481 }
3482
3483 //------------------------------frame_size_in_words-----------------------------
3484 // frame_slots in units of words
3485 int PhaseOutput::frame_size_in_words() const {
3486 // shift is 0 in LP32 and 1 in LP64
3487 const int shift = (LogBytesPerWord - LogBytesPerInt);
3488 int words = _frame_slots >> shift;
3489 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3490 return words;
3491 }
3492
3493 // To bang the stack of this compiled method we use the stack size
3494 // that the interpreter would need in case of a deoptimization. This
3495 // removes the need to bang the stack in the deoptimization blob which
3496 // in turn simplifies stack overflow handling.
3497 int PhaseOutput::bang_size_in_bytes() const {
3498 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3499 }
3500
3501 //------------------------------dump_asm---------------------------------------
3502 // Dump formatted assembly
3503 #if defined(SUPPORT_OPTO_ASSEMBLY)
3504 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3505
3506 int pc_digits = 3; // #chars required for pc
3507 int sb_chars = 3; // #chars for "start bundle" indicator
3508 int tab_size = 8;
3509 if (pcs != nullptr) {
3510 int max_pc = 0;
3511 for (uint i = 0; i < pc_limit; i++) {
3512 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3513 }
3514 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3515 }
3516 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3517
3518 bool cut_short = false;
3519 st->print_cr("#");
3520 st->print("# "); C->tf()->dump_on(st); st->cr();
3521 st->print_cr("#");
3522
3523 // For all blocks
3524 int pc = 0x0; // Program counter
3525 char starts_bundle = ' ';
3526 C->regalloc()->dump_frame();
3527
3528 Node *n = nullptr;
3529 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3530 if (VMThread::should_terminate()) {
3531 cut_short = true;
3532 break;
3533 }
3534 Block* block = C->cfg()->get_block(i);
3535 if (block->is_connector() && !Verbose) {
3536 continue;
3537 }
3538 n = block->head();
3539 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3540 pc = pcs[n->_idx];
3541 st->print("%*.*x", pc_digits, pc_digits, pc);
3542 }
3543 st->fill_to(prefix_len);
3544 block->dump_head(C->cfg(), st);
3545 if (block->is_connector()) {
3546 st->fill_to(prefix_len);
3547 st->print_cr("# Empty connector block");
3548 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3549 st->fill_to(prefix_len);
3550 st->print_cr("# Block is sole successor of call");
3551 }
3552
3553 // For all instructions
3554 Node *delay = nullptr;
3555 for (uint j = 0; j < block->number_of_nodes(); j++) {
3556 if (VMThread::should_terminate()) {
3557 cut_short = true;
3558 break;
3559 }
3560 n = block->get_node(j);
3561 if (valid_bundle_info(n)) {
3562 Bundle* bundle = node_bundling(n);
3563 if (bundle->used_in_unconditional_delay()) {
3564 delay = n;
3565 continue;
3566 }
3567 if (bundle->starts_bundle()) {
3568 starts_bundle = '+';
3569 }
3570 }
3571
3572 if (WizardMode) {
3573 n->dump();
3574 }
3575
3576 if( !n->is_Region() && // Dont print in the Assembly
3577 !n->is_Phi() && // a few noisely useless nodes
3578 !n->is_Proj() &&
3579 !n->is_MachTemp() &&
3580 !n->is_SafePointScalarObject() &&
3581 !n->is_Catch() && // Would be nice to print exception table targets
3582 !n->is_MergeMem() && // Not very interesting
3583 !n->is_top() && // Debug info table constants
3584 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3585 ) {
3586 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3587 pc = pcs[n->_idx];
3588 st->print("%*.*x", pc_digits, pc_digits, pc);
3589 } else {
3590 st->fill_to(pc_digits);
3591 }
3592 st->print(" %c ", starts_bundle);
3593 starts_bundle = ' ';
3594 st->fill_to(prefix_len);
3595 n->format(C->regalloc(), st);
3596 st->cr();
3597 }
3598
3599 // If we have an instruction with a delay slot, and have seen a delay,
3600 // then back up and print it
3601 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3602 // Coverity finding - Explicit null dereferenced.
3603 guarantee(delay != nullptr, "no unconditional delay instruction");
3604 if (WizardMode) delay->dump();
3605
3606 if (node_bundling(delay)->starts_bundle())
3607 starts_bundle = '+';
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 delay->format(C->regalloc(), st);
3618 st->cr();
3619 delay = nullptr;
3620 }
3621
3622 // Dump the exception table as well
3623 if( n->is_Catch() && (Verbose || WizardMode) ) {
3624 // Print the exception table for this offset
3625 _handler_table.print_subtable_for(pc);
3626 }
3627 st->bol(); // Make sure we start on a new line
3628 }
3629 st->cr(); // one empty line between blocks
3630 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3631 } // End of per-block dump
3632
3633 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3634 }
3635 #endif
3636
3637 #ifndef PRODUCT
3638 void PhaseOutput::print_statistics() {
3639 Scheduling::print_statistics();
3640 }
3641 #endif