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/gc_globals.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/allocation.inline.hpp"
39 #include "memory/allocation.hpp"
40 #include "opto/ad.hpp"
41 #include "opto/block.hpp"
42 #include "opto/c2compiler.hpp"
43 #include "opto/c2_MacroAssembler.hpp"
44 #include "opto/callnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/node.hpp"
49 #include "opto/optoreg.hpp"
50 #include "opto/output.hpp"
51 #include "opto/regalloc.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/subnode.hpp"
54 #include "opto/type.hpp"
55 #include "runtime/handles.inline.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "utilities/macros.hpp"
58 #include "utilities/powerOfTwo.hpp"
59 #include "utilities/xmlstream.hpp"
60
61 #ifndef PRODUCT
62 #define DEBUG_ARG(x) , x
63 #else
64 #define DEBUG_ARG(x)
65 #endif
66
67 //------------------------------Scheduling----------------------------------
68 // This class contains all the information necessary to implement instruction
69 // scheduling and bundling.
70 class Scheduling {
71
72 private:
73 // Arena to use
74 Arena *_arena;
75
76 // Control-Flow Graph info
77 PhaseCFG *_cfg;
78
79 // Register Allocation info
80 PhaseRegAlloc *_regalloc;
81
82 // Number of nodes in the method
83 uint _node_bundling_limit;
84
85 // List of scheduled nodes. Generated in reverse order
86 Node_List _scheduled;
87
88 // List of nodes currently available for choosing for scheduling
89 Node_List _available;
90
91 // For each instruction beginning a bundle, the number of following
92 // nodes to be bundled with it.
93 Bundle *_node_bundling_base;
94
95 // Mapping from register to Node
96 Node_List _reg_node;
97
98 // Free list for pinch nodes.
99 Node_List _pinch_free_list;
100
101 // Number of uses of this node within the containing basic block.
102 short *_uses;
103
104 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
105 // front, branch+proj at end. Also skips Catch/CProj (same as
106 // branch-at-end), plus just-prior exception-throwing call.
107 uint _bb_start, _bb_end;
108
109 // Latency from the end of the basic block as scheduled
110 unsigned short *_current_latency;
111
112 // Remember the next node
113 Node *_next_node;
114
115 // Use this for an unconditional branch delay slot
116 Node *_unconditional_delay_slot;
117
118 // Pointer to a Nop
119 MachNopNode *_nop;
120
121 // Length of the current bundle, in instructions
122 uint _bundle_instr_count;
123
124 // Current Cycle number, for computing latencies and bundling
125 uint _bundle_cycle_number;
126
127 // Bundle information
128 Pipeline_Use_Element _bundle_use_elements[resource_count];
129 Pipeline_Use _bundle_use;
130
131 // Dump the available list
132 void dump_available() const;
133
134 public:
135 Scheduling(Arena *arena, Compile &compile);
136
137 // Destructor
138 NOT_PRODUCT( ~Scheduling(); )
139
140 // Step ahead "i" cycles
141 void step(uint i);
142
143 // Step ahead 1 cycle, and clear the bundle state (for example,
144 // at a branch target)
145 void step_and_clear();
146
147 Bundle* node_bundling(const Node *n) {
148 assert(valid_bundle_info(n), "oob");
149 return (&_node_bundling_base[n->_idx]);
150 }
151
152 bool valid_bundle_info(const Node *n) const {
153 return (_node_bundling_limit > n->_idx);
154 }
155
156 bool starts_bundle(const Node *n) const {
157 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
158 }
159
160 // Do the scheduling
161 void DoScheduling();
162
163 // Compute the register antidependencies within a basic block
164 void ComputeRegisterAntidependencies(Block *bb);
165 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
166 void verify_good_schedule( Block *b, const char *msg );
167 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
168 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
169
170 // Add a node to the current bundle
171 void AddNodeToBundle(Node *n, const Block *bb);
172
173 // Return an integer less than, equal to, or greater than zero
174 // if the stack offset of the first argument is respectively
175 // less than, equal to, or greater than the second.
176 int compare_two_spill_nodes(Node* first, Node* second);
177
178 // Add a node to the list of available nodes
179 void AddNodeToAvailableList(Node *n);
180
181 // Compute the local use count for the nodes in a block, and compute
182 // the list of instructions with no uses in the block as available
183 void ComputeUseCount(const Block *bb);
184
185 // Choose an instruction from the available list to add to the bundle
186 Node * ChooseNodeToBundle();
187
188 // See if this Node fits into the currently accumulating bundle
189 bool NodeFitsInBundle(Node *n);
190
191 // Decrement the use count for a node
192 void DecrementUseCounts(Node *n, const Block *bb);
193
194 // Garbage collect pinch nodes for reuse by other blocks.
195 void garbage_collect_pinch_nodes();
196 // Clean up a pinch node for reuse (helper for above).
197 void cleanup_pinch( Node *pinch );
198
199 // Information for statistics gathering
200 #ifndef PRODUCT
201 private:
202 // Gather information on size of nops relative to total
203 uint _branches, _unconditional_delays;
204
205 static uint _total_nop_size, _total_method_size;
206 static uint _total_branches, _total_unconditional_delays;
207 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
208
209 public:
210 static void print_statistics();
211
212 static void increment_instructions_per_bundle(uint i) {
213 _total_instructions_per_bundle[i]++;
214 }
215
216 static void increment_nop_size(uint s) {
217 _total_nop_size += s;
218 }
219
220 static void increment_method_size(uint s) {
221 _total_method_size += s;
222 }
223 #endif
224
225 };
226
227 PhaseOutput::PhaseOutput()
228 : Phase(Phase::Output),
229 _code_buffer("Compile::Fill_buffer"),
230 _first_block_size(0),
231 _handler_table(),
232 _inc_table(),
233 _stub_list(),
234 _oop_map_set(nullptr),
235 _scratch_buffer_blob(nullptr),
236 _scratch_locs_memory(nullptr),
237 _scratch_const_size(-1),
238 _in_scratch_emit_size(false),
239 _frame_slots(0),
240 _code_offsets(),
241 _node_bundling_limit(0),
242 _node_bundling_base(nullptr),
243 _orig_pc_slot(0),
244 _orig_pc_slot_offset_in_bytes(0),
245 _buf_sizes(),
246 _block(nullptr),
247 _index(0) {
248 C->set_output(this);
249 if (C->stub_name() == nullptr) {
250 int fixed_slots = C->fixed_slots();
251 if (C->needs_stack_repair()) {
252 fixed_slots -= 2;
253 }
254 // TODO 8284443 Only reserve extra slot if needed
255 if (InlineTypeReturnedAsFields) {
256 fixed_slots -= 2;
257 }
258 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
259 }
260 }
261
262 PhaseOutput::~PhaseOutput() {
263 C->set_output(nullptr);
264 if (_scratch_buffer_blob != nullptr) {
265 BufferBlob::free(_scratch_buffer_blob);
266 }
267 }
268
269 void PhaseOutput::perform_mach_node_analysis() {
270 // Late barrier analysis must be done after schedule and bundle
271 // Otherwise liveness based spilling will fail
272 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
273 bs->late_barrier_analysis();
274
275 pd_perform_mach_node_analysis();
276
277 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
278 }
279
280 // Convert Nodes to instruction bits and pass off to the VM
281 void PhaseOutput::Output() {
282 // RootNode goes
283 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
284
285 // The number of new nodes (mostly MachNop) is proportional to
286 // the number of java calls and inner loops which are aligned.
287 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
288 C->inner_loops()*(OptoLoopAlignment-1)),
289 "out of nodes before code generation" ) ) {
290 return;
291 }
292 // Make sure I can find the Start Node
293 Block *entry = C->cfg()->get_block(1);
294 Block *broot = C->cfg()->get_root_block();
295
296 const StartNode *start = entry->head()->as_Start();
297
298 // Replace StartNode with prolog
299 Label verified_entry;
300 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
301 entry->map_node(prolog, 0);
302 C->cfg()->map_node_to_block(prolog, entry);
303 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
304
305 // Virtual methods need an unverified entry point
306 if (C->is_osr_compilation()) {
307 if (PoisonOSREntry) {
308 // TODO: Should use a ShouldNotReachHereNode...
309 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
310 }
311 } else {
312 if (C->method()) {
313 if (C->method()->has_scalarized_args()) {
314 // Add entry point to unpack all inline type arguments
315 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
316 if (!C->method()->is_static()) {
317 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
318 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
319 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
320 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
321 }
322 } else if (!C->method()->is_static()) {
323 // Insert unvalidated entry point
324 C->cfg()->insert(broot, 0, new MachUEPNode());
325 }
326 }
327 }
328
329 // Break before main entry point
330 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
331 (OptoBreakpoint && C->is_method_compilation()) ||
332 (OptoBreakpointOSR && C->is_osr_compilation()) ||
333 (OptoBreakpointC2R && !C->method()) ) {
334 // checking for C->method() means that OptoBreakpoint does not apply to
335 // runtime stubs or frame converters
336 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
337 }
338
339 // Insert epilogs before every return
340 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
341 Block* block = C->cfg()->get_block(i);
342 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
343 Node* m = block->end();
344 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
345 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
346 block->add_inst(epilog);
347 C->cfg()->map_node_to_block(epilog, block);
348 }
349 }
350 }
351
352 // Keeper of sizing aspects
353 _buf_sizes = BufferSizingData();
354
355 // Initialize code buffer
356 estimate_buffer_size(_buf_sizes._const);
357 if (C->failing()) return;
358
359 // Pre-compute the length of blocks and replace
360 // long branches with short if machine supports it.
361 // Must be done before ScheduleAndBundle due to SPARC delay slots
362 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
363 blk_starts[0] = 0;
364 shorten_branches(blk_starts);
365
366 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
367 // Compute the offsets of the entry points required by the inline type calling convention
368 if (!C->method()->is_static()) {
369 // We have entries at the beginning of the method, implemented by the first 4 nodes.
370 // Entry (unverified) @ offset 0
371 // Verified_Inline_Entry_RO
372 // Inline_Entry (unverified)
373 // Verified_Inline_Entry
374 uint offset = 0;
375 _code_offsets.set_value(CodeOffsets::Entry, offset);
376
377 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
378 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
379
380 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
381 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
382
383 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
384 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
385 } else {
386 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
387 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
388 }
389 }
390
391 ScheduleAndBundle();
392 if (C->failing()) {
393 return;
394 }
395
396 perform_mach_node_analysis();
397
398 // Complete sizing of codebuffer
399 CodeBuffer* cb = init_buffer();
400 if (cb == nullptr || C->failing()) {
401 return;
402 }
403
404 BuildOopMaps();
405
406 if (C->failing()) {
407 return;
408 }
409
410 C2_MacroAssembler masm(cb);
411 fill_buffer(&masm, blk_starts);
412 }
413
414 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
415 // Determine if we need to generate a stack overflow check.
416 // Do it if the method is not a stub function and
417 // has java calls or has frame size > vm_page_size/8.
418 // The debug VM checks that deoptimization doesn't trigger an
419 // unexpected stack overflow (compiled method stack banging should
420 // guarantee it doesn't happen) so we always need the stack bang in
421 // a debug VM.
422 return (C->stub_function() == nullptr &&
423 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
424 DEBUG_ONLY(|| true)));
425 }
426
427 bool PhaseOutput::need_register_stack_bang() const {
428 // Determine if we need to generate a register stack overflow check.
429 // This is only used on architectures which have split register
430 // and memory stacks (ie. IA64).
431 // Bang if the method is not a stub function and has java calls
432 return (C->stub_function() == nullptr && C->has_java_calls());
433 }
434
435
436 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
437 // of a loop. When aligning a loop we need to provide enough instructions
438 // in cpu's fetch buffer to feed decoders. The loop alignment could be
439 // avoided if we have enough instructions in fetch buffer at the head of a loop.
440 // By default, the size is set to 999999 by Block's constructor so that
441 // a loop will be aligned if the size is not reset here.
442 //
443 // Note: Mach instructions could contain several HW instructions
444 // so the size is estimated only.
445 //
446 void PhaseOutput::compute_loop_first_inst_sizes() {
447 // The next condition is used to gate the loop alignment optimization.
448 // Don't aligned a loop if there are enough instructions at the head of a loop
449 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
450 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
451 // equal to 11 bytes which is the largest address NOP instruction.
452 if (MaxLoopPad < OptoLoopAlignment - 1) {
453 uint last_block = C->cfg()->number_of_blocks() - 1;
454 for (uint i = 1; i <= last_block; i++) {
455 Block* block = C->cfg()->get_block(i);
456 // Check the first loop's block which requires an alignment.
457 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
458 uint sum_size = 0;
459 uint inst_cnt = NumberOfLoopInstrToAlign;
460 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
461
462 // Check subsequent fallthrough blocks if the loop's first
463 // block(s) does not have enough instructions.
464 Block *nb = block;
465 while(inst_cnt > 0 &&
466 i < last_block &&
467 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
468 !nb->has_successor(block)) {
469 i++;
470 nb = C->cfg()->get_block(i);
471 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
472 } // while( inst_cnt > 0 && i < last_block )
473
474 block->set_first_inst_size(sum_size);
475 } // f( b->head()->is_Loop() )
476 } // for( i <= last_block )
477 } // if( MaxLoopPad < OptoLoopAlignment-1 )
478 }
479
480 // The architecture description provides short branch variants for some long
481 // branch instructions. Replace eligible long branches with short branches.
482 void PhaseOutput::shorten_branches(uint* blk_starts) {
483
484 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
485
486 // Compute size of each block, method size, and relocation information size
487 uint nblocks = C->cfg()->number_of_blocks();
488
489 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
490 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
491 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
492
493 // Collect worst case block paddings
494 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
495 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
496
497 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
498 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
499
500 bool has_short_branch_candidate = false;
501
502 // Initialize the sizes to 0
503 int code_size = 0; // Size in bytes of generated code
504 int stub_size = 0; // Size in bytes of all stub entries
505 // Size in bytes of all relocation entries, including those in local stubs.
506 // Start with 2-bytes of reloc info for the unvalidated entry point
507 int reloc_size = 1; // Number of relocation entries
508
509 // Make three passes. The first computes pessimistic blk_starts,
510 // relative jmp_offset and reloc_size information. The second performs
511 // short branch substitution using the pessimistic sizing. The
512 // third inserts nops where needed.
513
514 // Step one, perform a pessimistic sizing pass.
515 uint last_call_adr = max_juint;
516 uint last_avoid_back_to_back_adr = max_juint;
517 uint nop_size = (new MachNopNode())->size(C->regalloc());
518 for (uint i = 0; i < nblocks; i++) { // For all blocks
519 Block* block = C->cfg()->get_block(i);
520 _block = block;
521
522 // During short branch replacement, we store the relative (to blk_starts)
523 // offset of jump in jmp_offset, rather than the absolute offset of jump.
524 // This is so that we do not need to recompute sizes of all nodes when
525 // we compute correct blk_starts in our next sizing pass.
526 jmp_offset[i] = 0;
527 jmp_size[i] = 0;
528 jmp_nidx[i] = -1;
529 DEBUG_ONLY( jmp_target[i] = 0; )
530 DEBUG_ONLY( jmp_rule[i] = 0; )
531
532 // Sum all instruction sizes to compute block size
533 uint last_inst = block->number_of_nodes();
534 uint blk_size = 0;
535 for (uint j = 0; j < last_inst; j++) {
536 _index = j;
537 Node* nj = block->get_node(_index);
538 // Handle machine instruction nodes
539 if (nj->is_Mach()) {
540 MachNode* mach = nj->as_Mach();
541 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
542 reloc_size += mach->reloc();
543 if (mach->is_MachCall()) {
544 // add size information for trampoline stub
545 // class CallStubImpl is platform-specific and defined in the *.ad files.
546 stub_size += CallStubImpl::size_call_trampoline();
547 reloc_size += CallStubImpl::reloc_call_trampoline();
548
549 MachCallNode *mcall = mach->as_MachCall();
550 // This destination address is NOT PC-relative
551
552 if (mcall->entry_point() != nullptr) {
553 mcall->method_set((intptr_t)mcall->entry_point());
554 }
555
556 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
557 stub_size += CompiledDirectCall::to_interp_stub_size();
558 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
559 }
560 } else if (mach->is_MachSafePoint()) {
561 // If call/safepoint are adjacent, account for possible
562 // nop to disambiguate the two safepoints.
563 // ScheduleAndBundle() can rearrange nodes in a block,
564 // check for all offsets inside this block.
565 if (last_call_adr >= blk_starts[i]) {
566 blk_size += nop_size;
567 }
568 }
569 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
570 // Nop is inserted between "avoid back to back" instructions.
571 // ScheduleAndBundle() can rearrange nodes in a block,
572 // check for all offsets inside this block.
573 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
574 blk_size += nop_size;
575 }
576 }
577 if (mach->may_be_short_branch()) {
578 if (!nj->is_MachBranch()) {
579 #ifndef PRODUCT
580 nj->dump(3);
581 #endif
582 Unimplemented();
583 }
584 assert(jmp_nidx[i] == -1, "block should have only one branch");
585 jmp_offset[i] = blk_size;
586 jmp_size[i] = nj->size(C->regalloc());
587 jmp_nidx[i] = j;
588 has_short_branch_candidate = true;
589 }
590 }
591 blk_size += nj->size(C->regalloc());
592 // Remember end of call offset
593 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
594 last_call_adr = blk_starts[i]+blk_size;
595 }
596 // Remember end of avoid_back_to_back offset
597 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
598 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
599 }
600 }
601
602 // When the next block starts a loop, we may insert pad NOP
603 // instructions. Since we cannot know our future alignment,
604 // assume the worst.
605 if (i < nblocks - 1) {
606 Block* nb = C->cfg()->get_block(i + 1);
607 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
608 if (max_loop_pad > 0) {
609 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
610 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
611 // If either is the last instruction in this block, bump by
612 // max_loop_pad in lock-step with blk_size, so sizing
613 // calculations in subsequent blocks still can conservatively
614 // detect that it may the last instruction in this block.
615 if (last_call_adr == blk_starts[i]+blk_size) {
616 last_call_adr += max_loop_pad;
617 }
618 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
619 last_avoid_back_to_back_adr += max_loop_pad;
620 }
621 blk_size += max_loop_pad;
622 block_worst_case_pad[i + 1] = max_loop_pad;
623 }
624 }
625
626 // Save block size; update total method size
627 blk_starts[i+1] = blk_starts[i]+blk_size;
628 }
629
630 // Step two, replace eligible long jumps.
631 bool progress = true;
632 uint last_may_be_short_branch_adr = max_juint;
633 while (has_short_branch_candidate && progress) {
634 progress = false;
635 has_short_branch_candidate = false;
636 int adjust_block_start = 0;
637 for (uint i = 0; i < nblocks; i++) {
638 Block* block = C->cfg()->get_block(i);
639 int idx = jmp_nidx[i];
640 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
641 if (mach != nullptr && mach->may_be_short_branch()) {
642 #ifdef ASSERT
643 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
644 int j;
645 // Find the branch; ignore trailing NOPs.
646 for (j = block->number_of_nodes()-1; j>=0; j--) {
647 Node* n = block->get_node(j);
648 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
649 break;
650 }
651 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
652 #endif
653 int br_size = jmp_size[i];
654 int br_offs = blk_starts[i] + jmp_offset[i];
655
656 // This requires the TRUE branch target be in succs[0]
657 uint bnum = block->non_connector_successor(0)->_pre_order;
658 int offset = blk_starts[bnum] - br_offs;
659 if (bnum > i) { // adjust following block's offset
660 offset -= adjust_block_start;
661 }
662
663 // This block can be a loop header, account for the padding
664 // in the previous block.
665 int block_padding = block_worst_case_pad[i];
666 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
667 // In the following code a nop could be inserted before
668 // the branch which will increase the backward distance.
669 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
670 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
671
672 if (needs_padding && offset <= 0)
673 offset -= nop_size;
674
675 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
676 // We've got a winner. Replace this branch.
677 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
678
679 // Update the jmp_size.
680 int new_size = replacement->size(C->regalloc());
681 int diff = br_size - new_size;
682 assert(diff >= (int)nop_size, "short_branch size should be smaller");
683 // Conservatively take into account padding between
684 // avoid_back_to_back branches. Previous branch could be
685 // converted into avoid_back_to_back branch during next
686 // rounds.
687 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
688 jmp_offset[i] += nop_size;
689 diff -= nop_size;
690 }
691 adjust_block_start += diff;
692 block->map_node(replacement, idx);
693 mach->subsume_by(replacement, C);
694 mach = replacement;
695 progress = true;
696
697 jmp_size[i] = new_size;
698 DEBUG_ONLY( jmp_target[i] = bnum; );
699 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
700 } else {
701 // The jump distance is not short, try again during next iteration.
702 has_short_branch_candidate = true;
703 }
704 } // (mach->may_be_short_branch())
705 if (mach != nullptr && (mach->may_be_short_branch() ||
706 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
707 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
708 }
709 blk_starts[i+1] -= adjust_block_start;
710 }
711 }
712
713 #ifdef ASSERT
714 for (uint i = 0; i < nblocks; i++) { // For all blocks
715 if (jmp_target[i] != 0) {
716 int br_size = jmp_size[i];
717 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
718 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
719 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]);
720 }
721 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
722 }
723 }
724 #endif
725
726 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
727 // after ScheduleAndBundle().
728
729 // ------------------
730 // Compute size for code buffer
731 code_size = blk_starts[nblocks];
732
733 // Relocation records
734 reloc_size += 1; // Relo entry for exception handler
735
736 // Adjust reloc_size to number of record of relocation info
737 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
738 // a relocation index.
739 // The CodeBuffer will expand the locs array if this estimate is too low.
740 reloc_size *= 10 / sizeof(relocInfo);
741
742 _buf_sizes._reloc = reloc_size;
743 _buf_sizes._code = code_size;
744 _buf_sizes._stub = stub_size;
745 }
746
747 //------------------------------FillLocArray-----------------------------------
748 // Create a bit of debug info and append it to the array. The mapping is from
749 // Java local or expression stack to constant, register or stack-slot. For
750 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
751 // entry has been taken care of and caller should skip it).
752 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
753 // This should never have accepted Bad before
754 assert(OptoReg::is_valid(regnum), "location must be valid");
755 return (OptoReg::is_reg(regnum))
756 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
757 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
758 }
759
760
761 ObjectValue*
762 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
763 for (int i = 0; i < objs->length(); i++) {
764 assert(objs->at(i)->is_object(), "corrupt object cache");
765 ObjectValue* sv = objs->at(i)->as_ObjectValue();
766 if (sv->id() == id) {
767 return sv;
768 }
769 }
770 // Otherwise..
771 return nullptr;
772 }
773
774 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
775 ObjectValue* sv ) {
776 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
777 objs->append(sv);
778 }
779
780
781 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
782 GrowableArray<ScopeValue*> *array,
783 GrowableArray<ScopeValue*> *objs ) {
784 assert( local, "use _top instead of null" );
785 if (array->length() != idx) {
786 assert(array->length() == idx + 1, "Unexpected array count");
787 // Old functionality:
788 // return
789 // New functionality:
790 // Assert if the local is not top. In product mode let the new node
791 // override the old entry.
792 assert(local == C->top(), "LocArray collision");
793 if (local == C->top()) {
794 return;
795 }
796 array->pop();
797 }
798 const Type *t = local->bottom_type();
799
800 // Is it a safepoint scalar object node?
801 if (local->is_SafePointScalarObject()) {
802 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
803
804 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
805 if (sv == nullptr) {
806 ciKlass* cik = t->is_oopptr()->exact_klass();
807 assert(cik->is_instance_klass() ||
808 cik->is_array_klass(), "Not supported allocation.");
809 uint first_ind = spobj->first_index(sfpt->jvms());
810 // Nullable, scalarized inline types have an is_init input
811 // that needs to be checked before using the field values.
812 ScopeValue* is_init = nullptr;
813 if (cik->is_inlinetype()) {
814 Node* init_node = sfpt->in(first_ind++);
815 assert(init_node != nullptr, "is_init node not found");
816 if (!init_node->is_top()) {
817 const TypeInt* init_type = init_node->bottom_type()->is_int();
818 if (init_node->is_Con()) {
819 is_init = new ConstantIntValue(init_type->get_con());
820 } else {
821 OptoReg::Name init_reg = C->regalloc()->get_reg_first(init_node);
822 is_init = new_loc_value(C->regalloc(), init_reg, Location::normal);
823 }
824 }
825 }
826 sv = new ObjectValue(spobj->_idx,
827 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, is_init);
828 set_sv_for_object_node(objs, sv);
829
830 for (uint i = 0; i < spobj->n_fields(); i++) {
831 Node* fld_node = sfpt->in(first_ind+i);
832 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
833 }
834 }
835 array->append(sv);
836 return;
837 } else if (local->is_SafePointScalarMerge()) {
838 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
839 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
840
841 if (mv == nullptr) {
842 GrowableArray<ScopeValue*> deps;
843
844 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
845 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
846 assert(deps.length() == 1, "missing value");
847
848 int selector_idx = smerge->selector_idx(sfpt->jvms());
849 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
850 assert(deps.length() == 2, "missing value");
851
852 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
853 set_sv_for_object_node(objs, mv);
854
855 for (uint i = 1; i < smerge->req(); i++) {
856 Node* obj_node = smerge->in(i);
857 int idx = mv->possible_objects()->length();
858 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
859
860 // By default ObjectValues that are in 'possible_objects' are not root objects.
861 // They will be marked as root later if they are directly referenced in a JVMS.
862 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
863 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
864 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
865 }
866 }
867 array->append(mv);
868 return;
869 }
870
871 // Grab the register number for the local
872 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
873 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
874 // Record the double as two float registers.
875 // The register mask for such a value always specifies two adjacent
876 // float registers, with the lower register number even.
877 // Normally, the allocation of high and low words to these registers
878 // is irrelevant, because nearly all operations on register pairs
879 // (e.g., StoreD) treat them as a single unit.
880 // Here, we assume in addition that the words in these two registers
881 // stored "naturally" (by operations like StoreD and double stores
882 // within the interpreter) such that the lower-numbered register
883 // is written to the lower memory address. This may seem like
884 // a machine dependency, but it is not--it is a requirement on
885 // the author of the <arch>.ad file to ensure that, for every
886 // even/odd double-register pair to which a double may be allocated,
887 // the word in the even single-register is stored to the first
888 // memory word. (Note that register numbers are completely
889 // arbitrary, and are not tied to any machine-level encodings.)
890 #ifdef _LP64
891 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
892 array->append(new ConstantIntValue((jint)0));
893 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
894 } else if ( t->base() == Type::Long ) {
895 array->append(new ConstantIntValue((jint)0));
896 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
897 } else if ( t->base() == Type::RawPtr ) {
898 // jsr/ret return address which must be restored into the full
899 // width 64-bit stack slot.
900 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
901 }
902 #else //_LP64
903 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
904 // Repack the double/long as two jints.
905 // The convention the interpreter uses is that the second local
906 // holds the first raw word of the native double representation.
907 // This is actually reasonable, since locals and stack arrays
908 // grow downwards in all implementations.
909 // (If, on some machine, the interpreter's Java locals or stack
910 // were to grow upwards, the embedded doubles would be word-swapped.)
911 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
912 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
913 }
914 #endif //_LP64
915 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
916 OptoReg::is_reg(regnum) ) {
917 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
918 ? Location::float_in_dbl : Location::normal ));
919 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
920 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
921 ? Location::int_in_long : Location::normal ));
922 } else if( t->base() == Type::NarrowOop ) {
923 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
924 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
925 t->base() == Type::VectorD || t->base() == Type::VectorX ||
926 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
927 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
928 } else if (C->regalloc()->is_oop(local)) {
929 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
930 t->base() == Type::AryPtr,
931 "Unexpected type: %s", t->msg());
932 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
933 } else {
934 assert(t->base() == Type::Int || t->base() == Type::Half ||
935 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
936 "Unexpected type: %s", t->msg());
937 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
938 }
939 return;
940 }
941
942 // No register. It must be constant data.
943 switch (t->base()) {
944 case Type::Half: // Second half of a double
945 ShouldNotReachHere(); // Caller should skip 2nd halves
946 break;
947 case Type::AnyPtr:
948 array->append(new ConstantOopWriteValue(nullptr));
949 break;
950 case Type::AryPtr:
951 case Type::InstPtr: // fall through
952 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
953 break;
954 case Type::NarrowOop:
955 if (t == TypeNarrowOop::NULL_PTR) {
956 array->append(new ConstantOopWriteValue(nullptr));
957 } else {
958 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
959 }
960 break;
961 case Type::Int:
962 array->append(new ConstantIntValue(t->is_int()->get_con()));
963 break;
964 case Type::RawPtr:
965 // A return address (T_ADDRESS).
966 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
967 #ifdef _LP64
968 // Must be restored to the full-width 64-bit stack slot.
969 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
970 #else
971 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
972 #endif
973 break;
974 case Type::FloatCon: {
975 float f = t->is_float_constant()->getf();
976 array->append(new ConstantIntValue(jint_cast(f)));
977 break;
978 }
979 case Type::DoubleCon: {
980 jdouble d = t->is_double_constant()->getd();
981 #ifdef _LP64
982 array->append(new ConstantIntValue((jint)0));
983 array->append(new ConstantDoubleValue(d));
984 #else
985 // Repack the double as two jints.
986 // The convention the interpreter uses is that the second local
987 // holds the first raw word of the native double representation.
988 // This is actually reasonable, since locals and stack arrays
989 // grow downwards in all implementations.
990 // (If, on some machine, the interpreter's Java locals or stack
991 // were to grow upwards, the embedded doubles would be word-swapped.)
992 jlong_accessor acc;
993 acc.long_value = jlong_cast(d);
994 array->append(new ConstantIntValue(acc.words[1]));
995 array->append(new ConstantIntValue(acc.words[0]));
996 #endif
997 break;
998 }
999 case Type::Long: {
1000 jlong d = t->is_long()->get_con();
1001 #ifdef _LP64
1002 array->append(new ConstantIntValue((jint)0));
1003 array->append(new ConstantLongValue(d));
1004 #else
1005 // Repack the long as two jints.
1006 // The convention the interpreter uses is that the second local
1007 // holds the first raw word of the native double representation.
1008 // This is actually reasonable, since locals and stack arrays
1009 // grow downwards in all implementations.
1010 // (If, on some machine, the interpreter's Java locals or stack
1011 // were to grow upwards, the embedded doubles would be word-swapped.)
1012 jlong_accessor acc;
1013 acc.long_value = d;
1014 array->append(new ConstantIntValue(acc.words[1]));
1015 array->append(new ConstantIntValue(acc.words[0]));
1016 #endif
1017 break;
1018 }
1019 case Type::Top: // Add an illegal value here
1020 array->append(new LocationValue(Location()));
1021 break;
1022 default:
1023 ShouldNotReachHere();
1024 break;
1025 }
1026 }
1027
1028 // Determine if this node starts a bundle
1029 bool PhaseOutput::starts_bundle(const Node *n) const {
1030 return (_node_bundling_limit > n->_idx &&
1031 _node_bundling_base[n->_idx].starts_bundle());
1032 }
1033
1034 // Determine if there is a monitor that has 'ov' as its owner.
1035 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
1036 for (int k = 0; k < monarray->length(); k++) {
1037 MonitorValue* mv = monarray->at(k);
1038 if (mv->owner() == ov) {
1039 return true;
1040 }
1041 }
1042
1043 return false;
1044 }
1045
1046 // Determine if there is a scalar replaced object description represented by 'ov'.
1047 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
1048 GrowableArray<ScopeValue*>* objs,
1049 ObjectValue* ov) const {
1050 for (int i = 0; i < jvms->scl_size(); i++) {
1051 Node* n = sfn->scalarized_obj(jvms, i);
1052 // Other kinds of nodes that we may encounter here, for instance constants
1053 // representing values of fields of objects scalarized, aren't relevant for
1054 // us, since they don't map to ObjectValue.
1055 if (!n->is_SafePointScalarObject()) {
1056 continue;
1057 }
1058
1059 ObjectValue* other = sv_for_node_id(objs, n->_idx);
1060 if (ov == other) {
1061 return true;
1062 }
1063 }
1064 return false;
1065 }
1066
1067 //--------------------------Process_OopMap_Node--------------------------------
1068 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1069 // Handle special safepoint nodes for synchronization
1070 MachSafePointNode *sfn = mach->as_MachSafePoint();
1071 MachCallNode *mcall;
1072
1073 int safepoint_pc_offset = current_offset;
1074 bool is_method_handle_invoke = false;
1075 bool return_oop = false;
1076 bool return_scalarized = false;
1077 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1078 bool arg_escape = false;
1079
1080 // Add the safepoint in the DebugInfoRecorder
1081 if( !mach->is_MachCall() ) {
1082 mcall = nullptr;
1083 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1084 } else {
1085 mcall = mach->as_MachCall();
1086
1087 // Is the call a MethodHandle call?
1088 if (mcall->is_MachCallJava()) {
1089 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1090 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1091 is_method_handle_invoke = true;
1092 }
1093 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1094 }
1095
1096 // Check if a call returns an object.
1097 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1098 return_oop = true;
1099 }
1100 if (mcall->returns_scalarized()) {
1101 return_scalarized = true;
1102 }
1103 safepoint_pc_offset += mcall->ret_addr_offset();
1104 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1105 }
1106
1107 // Loop over the JVMState list to add scope information
1108 // Do not skip safepoints with a null method, they need monitor info
1109 JVMState* youngest_jvms = sfn->jvms();
1110 int max_depth = youngest_jvms->depth();
1111
1112 // Allocate the object pool for scalar-replaced objects -- the map from
1113 // small-integer keys (which can be recorded in the local and ostack
1114 // arrays) to descriptions of the object state.
1115 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1116
1117 // Visit scopes from oldest to youngest.
1118 for (int depth = 1; depth <= max_depth; depth++) {
1119 JVMState* jvms = youngest_jvms->of_depth(depth);
1120 int idx;
1121 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1122 // Safepoints that do not have method() set only provide oop-map and monitor info
1123 // to support GC; these do not support deoptimization.
1124 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1125 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1126 int num_mon = jvms->nof_monitors();
1127 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1128 "JVMS local count must match that of the method");
1129
1130 // Add Local and Expression Stack Information
1131
1132 // Insert locals into the locarray
1133 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1134 for( idx = 0; idx < num_locs; idx++ ) {
1135 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1136 }
1137
1138 // Insert expression stack entries into the exparray
1139 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1140 for( idx = 0; idx < num_exps; idx++ ) {
1141 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1142 }
1143
1144 // Add in mappings of the monitors
1145 assert( !method ||
1146 !method->is_synchronized() ||
1147 method->is_native() ||
1148 num_mon > 0 ||
1149 !GenerateSynchronizationCode,
1150 "monitors must always exist for synchronized methods");
1151
1152 // Build the growable array of ScopeValues for exp stack
1153 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1154
1155 // Loop over monitors and insert into array
1156 for (idx = 0; idx < num_mon; idx++) {
1157 // Grab the node that defines this monitor
1158 Node* box_node = sfn->monitor_box(jvms, idx);
1159 Node* obj_node = sfn->monitor_obj(jvms, idx);
1160
1161 // Create ScopeValue for object
1162 ScopeValue *scval = nullptr;
1163
1164 if (obj_node->is_SafePointScalarObject()) {
1165 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1166 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1167 if (scval == nullptr) {
1168 const Type *t = spobj->bottom_type();
1169 ciKlass* cik = t->is_oopptr()->exact_klass();
1170 assert(cik->is_instance_klass() ||
1171 cik->is_array_klass(), "Not supported allocation.");
1172 ObjectValue* sv = new ObjectValue(spobj->_idx,
1173 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1174 PhaseOutput::set_sv_for_object_node(objs, sv);
1175
1176 uint first_ind = spobj->first_index(youngest_jvms);
1177 for (uint i = 0; i < spobj->n_fields(); i++) {
1178 Node* fld_node = sfn->in(first_ind+i);
1179 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1180 }
1181 scval = sv;
1182 }
1183 } else if (obj_node->is_SafePointScalarMerge()) {
1184 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1185 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1186
1187 if (mv == nullptr) {
1188 GrowableArray<ScopeValue*> deps;
1189
1190 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1191 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1192 assert(deps.length() == 1, "missing value");
1193
1194 int selector_idx = smerge->selector_idx(youngest_jvms);
1195 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1196 assert(deps.length() == 2, "missing value");
1197
1198 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1199 set_sv_for_object_node(objs, mv);
1200
1201 for (uint i = 1; i < smerge->req(); i++) {
1202 Node* obj_node = smerge->in(i);
1203 int idx = mv->possible_objects()->length();
1204 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1205
1206 // By default ObjectValues that are in 'possible_objects' are not root objects.
1207 // They will be marked as root later if they are directly referenced in a JVMS.
1208 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1209 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1210 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1211 }
1212 }
1213 scval = mv;
1214 } else if (!obj_node->is_Con()) {
1215 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1216 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1217 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1218 } else {
1219 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1220 }
1221 } else {
1222 const TypePtr *tp = obj_node->get_ptr_type();
1223 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1224 }
1225
1226 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1227 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1228 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1229 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1230 }
1231
1232 // Mark ObjectValue nodes as root nodes if they are directly
1233 // referenced in the JVMS.
1234 for (int i = 0; i < objs->length(); i++) {
1235 ScopeValue* sv = objs->at(i);
1236 if (sv->is_object_merge()) {
1237 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1238
1239 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1240 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1241 if (ov->is_root()) {
1242 // Already flagged as 'root' by something else. We shouldn't change it
1243 // to non-root in a younger JVMS because it may need to be alive in
1244 // a younger JVMS.
1245 } else {
1246 bool is_root = locarray->contains(ov) ||
1247 exparray->contains(ov) ||
1248 contains_as_owner(monarray, ov) ||
1249 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1250 ov->set_root(is_root);
1251 }
1252 }
1253 }
1254 }
1255
1256 // We dump the object pool first, since deoptimization reads it in first.
1257 C->debug_info()->dump_object_pool(objs);
1258
1259 // Build first class objects to pass to scope
1260 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1261 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1262 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1263
1264 // Make method available for all Safepoints
1265 ciMethod* scope_method = method ? method : C->method();
1266 // Describe the scope here
1267 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1268 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1269 // Now we can describe the scope.
1270 methodHandle null_mh;
1271 bool rethrow_exception = false;
1272 C->debug_info()->describe_scope(
1273 safepoint_pc_offset,
1274 null_mh,
1275 scope_method,
1276 jvms->bci(),
1277 jvms->should_reexecute(),
1278 rethrow_exception,
1279 is_method_handle_invoke,
1280 return_oop,
1281 return_scalarized,
1282 has_ea_local_in_scope,
1283 arg_escape,
1284 locvals,
1285 expvals,
1286 monvals
1287 );
1288 } // End jvms loop
1289
1290 // Mark the end of the scope set.
1291 C->debug_info()->end_safepoint(safepoint_pc_offset);
1292 }
1293
1294
1295
1296 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1297 class NonSafepointEmitter {
1298 Compile* C;
1299 JVMState* _pending_jvms;
1300 int _pending_offset;
1301
1302 void emit_non_safepoint();
1303
1304 public:
1305 NonSafepointEmitter(Compile* compile) {
1306 this->C = compile;
1307 _pending_jvms = nullptr;
1308 _pending_offset = 0;
1309 }
1310
1311 void observe_instruction(Node* n, int pc_offset) {
1312 if (!C->debug_info()->recording_non_safepoints()) return;
1313
1314 Node_Notes* nn = C->node_notes_at(n->_idx);
1315 if (nn == nullptr || nn->jvms() == nullptr) return;
1316 if (_pending_jvms != nullptr &&
1317 _pending_jvms->same_calls_as(nn->jvms())) {
1318 // Repeated JVMS? Stretch it up here.
1319 _pending_offset = pc_offset;
1320 } else {
1321 if (_pending_jvms != nullptr &&
1322 _pending_offset < pc_offset) {
1323 emit_non_safepoint();
1324 }
1325 _pending_jvms = nullptr;
1326 if (pc_offset > C->debug_info()->last_pc_offset()) {
1327 // This is the only way _pending_jvms can become non-null:
1328 _pending_jvms = nn->jvms();
1329 _pending_offset = pc_offset;
1330 }
1331 }
1332 }
1333
1334 // Stay out of the way of real safepoints:
1335 void observe_safepoint(JVMState* jvms, int pc_offset) {
1336 if (_pending_jvms != nullptr &&
1337 !_pending_jvms->same_calls_as(jvms) &&
1338 _pending_offset < pc_offset) {
1339 emit_non_safepoint();
1340 }
1341 _pending_jvms = nullptr;
1342 }
1343
1344 void flush_at_end() {
1345 if (_pending_jvms != nullptr) {
1346 emit_non_safepoint();
1347 }
1348 _pending_jvms = nullptr;
1349 }
1350 };
1351
1352 void NonSafepointEmitter::emit_non_safepoint() {
1353 JVMState* youngest_jvms = _pending_jvms;
1354 int pc_offset = _pending_offset;
1355
1356 // Clear it now:
1357 _pending_jvms = nullptr;
1358
1359 DebugInformationRecorder* debug_info = C->debug_info();
1360 assert(debug_info->recording_non_safepoints(), "sanity");
1361
1362 debug_info->add_non_safepoint(pc_offset);
1363 int max_depth = youngest_jvms->depth();
1364
1365 // Visit scopes from oldest to youngest.
1366 for (int depth = 1; depth <= max_depth; depth++) {
1367 JVMState* jvms = youngest_jvms->of_depth(depth);
1368 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1369 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1370 methodHandle null_mh;
1371 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1372 }
1373
1374 // Mark the end of the scope set.
1375 debug_info->end_non_safepoint(pc_offset);
1376 }
1377
1378 //------------------------------init_buffer------------------------------------
1379 void PhaseOutput::estimate_buffer_size(int& const_req) {
1380
1381 // Set the initially allocated size
1382 const_req = initial_const_capacity;
1383
1384 // The extra spacing after the code is necessary on some platforms.
1385 // Sometimes we need to patch in a jump after the last instruction,
1386 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1387
1388 // Compute the byte offset where we can store the deopt pc.
1389 if (C->fixed_slots() != 0) {
1390 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1391 }
1392
1393 // Compute prolog code size
1394 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1395 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1396
1397 if (C->has_mach_constant_base_node()) {
1398 uint add_size = 0;
1399 // Fill the constant table.
1400 // Note: This must happen before shorten_branches.
1401 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1402 Block* b = C->cfg()->get_block(i);
1403
1404 for (uint j = 0; j < b->number_of_nodes(); j++) {
1405 Node* n = b->get_node(j);
1406
1407 // If the node is a MachConstantNode evaluate the constant
1408 // value section.
1409 if (n->is_MachConstant()) {
1410 MachConstantNode* machcon = n->as_MachConstant();
1411 machcon->eval_constant(C);
1412 } else if (n->is_Mach()) {
1413 // On Power there are more nodes that issue constants.
1414 add_size += (n->as_Mach()->ins_num_consts() * 8);
1415 }
1416 }
1417 }
1418
1419 // Calculate the offsets of the constants and the size of the
1420 // constant table (including the padding to the next section).
1421 constant_table().calculate_offsets_and_size();
1422 const_req = constant_table().size() + add_size;
1423 }
1424
1425 // Initialize the space for the BufferBlob used to find and verify
1426 // instruction size in MachNode::emit_size()
1427 init_scratch_buffer_blob(const_req);
1428 }
1429
1430 CodeBuffer* PhaseOutput::init_buffer() {
1431 int stub_req = _buf_sizes._stub;
1432 int code_req = _buf_sizes._code;
1433 int const_req = _buf_sizes._const;
1434
1435 int pad_req = NativeCall::byte_size();
1436
1437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1438 stub_req += bs->estimate_stub_size();
1439
1440 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1441 // class HandlerImpl is platform-specific and defined in the *.ad files.
1442 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1443 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1444 stub_req += MAX_stubs_size; // ensure per-stub margin
1445 code_req += MAX_inst_size; // ensure per-instruction margin
1446
1447 if (StressCodeBuffers)
1448 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1449
1450 int total_req =
1451 const_req +
1452 code_req +
1453 pad_req +
1454 stub_req +
1455 exception_handler_req +
1456 deopt_handler_req; // deopt handler
1457
1458 if (C->has_method_handle_invokes())
1459 total_req += deopt_handler_req; // deopt MH handler
1460
1461 CodeBuffer* cb = code_buffer();
1462 cb->initialize(total_req, _buf_sizes._reloc);
1463
1464 // Have we run out of code space?
1465 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1466 C->record_failure("CodeCache is full");
1467 return nullptr;
1468 }
1469 // Configure the code buffer.
1470 cb->initialize_consts_size(const_req);
1471 cb->initialize_stubs_size(stub_req);
1472 cb->initialize_oop_recorder(C->env()->oop_recorder());
1473
1474 // fill in the nop array for bundling computations
1475 MachNode *_nop_list[Bundle::_nop_count];
1476 Bundle::initialize_nops(_nop_list);
1477
1478 return cb;
1479 }
1480
1481 //------------------------------fill_buffer------------------------------------
1482 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1483 // blk_starts[] contains offsets calculated during short branches processing,
1484 // offsets should not be increased during following steps.
1485
1486 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1487 // of a loop. It is used to determine the padding for loop alignment.
1488 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1489
1490 compute_loop_first_inst_sizes();
1491
1492 // Create oopmap set.
1493 _oop_map_set = new OopMapSet();
1494
1495 // !!!!! This preserves old handling of oopmaps for now
1496 C->debug_info()->set_oopmaps(_oop_map_set);
1497
1498 uint nblocks = C->cfg()->number_of_blocks();
1499 // Count and start of implicit null check instructions
1500 uint inct_cnt = 0;
1501 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1502
1503 // Count and start of calls
1504 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1505
1506 uint return_offset = 0;
1507 int nop_size = (new MachNopNode())->size(C->regalloc());
1508
1509 int previous_offset = 0;
1510 int current_offset = 0;
1511 int last_call_offset = -1;
1512 int last_avoid_back_to_back_offset = -1;
1513 #ifdef ASSERT
1514 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1515 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1516 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1517 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1518 #endif
1519
1520 // Create an array of unused labels, one for each basic block, if printing is enabled
1521 #if defined(SUPPORT_OPTO_ASSEMBLY)
1522 int* node_offsets = nullptr;
1523 uint node_offset_limit = C->unique();
1524
1525 if (C->print_assembly()) {
1526 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1527 }
1528 if (node_offsets != nullptr) {
1529 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1530 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1531 }
1532 #endif
1533
1534 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1535
1536 // Emit the constant table.
1537 if (C->has_mach_constant_base_node()) {
1538 if (!constant_table().emit(masm)) {
1539 C->record_failure("consts section overflow");
1540 return;
1541 }
1542 }
1543
1544 // Create an array of labels, one for each basic block
1545 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1546 for (uint i = 0; i <= nblocks; i++) {
1547 blk_labels[i].init();
1548 }
1549
1550 // Now fill in the code buffer
1551 Node* delay_slot = nullptr;
1552 for (uint i = 0; i < nblocks; i++) {
1553 Block* block = C->cfg()->get_block(i);
1554 _block = block;
1555 Node* head = block->head();
1556
1557 // If this block needs to start aligned (i.e, can be reached other
1558 // than by falling-thru from the previous block), then force the
1559 // start of a new bundle.
1560 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1561 masm->code()->flush_bundle(true);
1562 }
1563
1564 #ifdef ASSERT
1565 if (!block->is_connector()) {
1566 stringStream st;
1567 block->dump_head(C->cfg(), &st);
1568 masm->block_comment(st.freeze());
1569 }
1570 jmp_target[i] = 0;
1571 jmp_offset[i] = 0;
1572 jmp_size[i] = 0;
1573 jmp_rule[i] = 0;
1574 #endif
1575 int blk_offset = current_offset;
1576
1577 // Define the label at the beginning of the basic block
1578 masm->bind(blk_labels[block->_pre_order]);
1579
1580 uint last_inst = block->number_of_nodes();
1581
1582 // Emit block normally, except for last instruction.
1583 // Emit means "dump code bits into code buffer".
1584 for (uint j = 0; j<last_inst; j++) {
1585 _index = j;
1586
1587 // Get the node
1588 Node* n = block->get_node(j);
1589
1590 // See if delay slots are supported
1591 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1592 assert(delay_slot == nullptr, "no use of delay slot node");
1593 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1594
1595 delay_slot = n;
1596 continue;
1597 }
1598
1599 // If this starts a new instruction group, then flush the current one
1600 // (but allow split bundles)
1601 if (Pipeline::requires_bundling() && starts_bundle(n))
1602 masm->code()->flush_bundle(false);
1603
1604 // Special handling for SafePoint/Call Nodes
1605 bool is_mcall = false;
1606 if (n->is_Mach()) {
1607 MachNode *mach = n->as_Mach();
1608 is_mcall = n->is_MachCall();
1609 bool is_sfn = n->is_MachSafePoint();
1610
1611 // If this requires all previous instructions be flushed, then do so
1612 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1613 masm->code()->flush_bundle(true);
1614 current_offset = masm->offset();
1615 }
1616
1617 // A padding may be needed again since a previous instruction
1618 // could be moved to delay slot.
1619
1620 // align the instruction if necessary
1621 int padding = mach->compute_padding(current_offset);
1622 // Make sure safepoint node for polling is distinct from a call's
1623 // return by adding a nop if needed.
1624 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1625 padding = nop_size;
1626 }
1627 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1628 current_offset == last_avoid_back_to_back_offset) {
1629 // Avoid back to back some instructions.
1630 padding = nop_size;
1631 }
1632
1633 if (padding > 0) {
1634 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1635 int nops_cnt = padding / nop_size;
1636 MachNode *nop = new MachNopNode(nops_cnt);
1637 block->insert_node(nop, j++);
1638 last_inst++;
1639 C->cfg()->map_node_to_block(nop, block);
1640 // Ensure enough space.
1641 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1642 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1643 C->record_failure("CodeCache is full");
1644 return;
1645 }
1646 nop->emit(masm, C->regalloc());
1647 masm->code()->flush_bundle(true);
1648 current_offset = masm->offset();
1649 }
1650
1651 bool observe_safepoint = is_sfn;
1652 // Remember the start of the last call in a basic block
1653 if (is_mcall) {
1654 MachCallNode *mcall = mach->as_MachCall();
1655
1656 if (mcall->entry_point() != nullptr) {
1657 // This destination address is NOT PC-relative
1658 mcall->method_set((intptr_t)mcall->entry_point());
1659 }
1660
1661 // Save the return address
1662 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1663
1664 observe_safepoint = mcall->guaranteed_safepoint();
1665 }
1666
1667 // sfn will be valid whenever mcall is valid now because of inheritance
1668 if (observe_safepoint) {
1669 // Handle special safepoint nodes for synchronization
1670 if (!is_mcall) {
1671 MachSafePointNode *sfn = mach->as_MachSafePoint();
1672 // !!!!! Stubs only need an oopmap right now, so bail out
1673 if (sfn->jvms()->method() == nullptr) {
1674 // Write the oopmap directly to the code blob??!!
1675 continue;
1676 }
1677 } // End synchronization
1678
1679 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1680 current_offset);
1681 Process_OopMap_Node(mach, current_offset);
1682 } // End if safepoint
1683
1684 // If this is a null check, then add the start of the previous instruction to the list
1685 else if( mach->is_MachNullCheck() ) {
1686 inct_starts[inct_cnt++] = previous_offset;
1687 }
1688
1689 // If this is a branch, then fill in the label with the target BB's label
1690 else if (mach->is_MachBranch()) {
1691 // This requires the TRUE branch target be in succs[0]
1692 uint block_num = block->non_connector_successor(0)->_pre_order;
1693
1694 // Try to replace long branch if delay slot is not used,
1695 // it is mostly for back branches since forward branch's
1696 // distance is not updated yet.
1697 bool delay_slot_is_used = valid_bundle_info(n) &&
1698 C->output()->node_bundling(n)->use_unconditional_delay();
1699 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1700 assert(delay_slot == nullptr, "not expecting delay slot node");
1701 int br_size = n->size(C->regalloc());
1702 int offset = blk_starts[block_num] - current_offset;
1703 if (block_num >= i) {
1704 // Current and following block's offset are not
1705 // finalized yet, adjust distance by the difference
1706 // between calculated and final offsets of current block.
1707 offset -= (blk_starts[i] - blk_offset);
1708 }
1709 // In the following code a nop could be inserted before
1710 // the branch which will increase the backward distance.
1711 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1712 if (needs_padding && offset <= 0)
1713 offset -= nop_size;
1714
1715 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1716 // We've got a winner. Replace this branch.
1717 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1718
1719 // Update the jmp_size.
1720 int new_size = replacement->size(C->regalloc());
1721 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1722 // Insert padding between avoid_back_to_back branches.
1723 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1724 MachNode *nop = new MachNopNode();
1725 block->insert_node(nop, j++);
1726 C->cfg()->map_node_to_block(nop, block);
1727 last_inst++;
1728 nop->emit(masm, C->regalloc());
1729 masm->code()->flush_bundle(true);
1730 current_offset = masm->offset();
1731 }
1732 #ifdef ASSERT
1733 jmp_target[i] = block_num;
1734 jmp_offset[i] = current_offset - blk_offset;
1735 jmp_size[i] = new_size;
1736 jmp_rule[i] = mach->rule();
1737 #endif
1738 block->map_node(replacement, j);
1739 mach->subsume_by(replacement, C);
1740 n = replacement;
1741 mach = replacement;
1742 }
1743 }
1744 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1745 } else if (mach->ideal_Opcode() == Op_Jump) {
1746 for (uint h = 0; h < block->_num_succs; h++) {
1747 Block* succs_block = block->_succs[h];
1748 for (uint j = 1; j < succs_block->num_preds(); j++) {
1749 Node* jpn = succs_block->pred(j);
1750 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1751 uint block_num = succs_block->non_connector()->_pre_order;
1752 Label *blkLabel = &blk_labels[block_num];
1753 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1754 }
1755 }
1756 }
1757 } else if (!n->is_Proj()) {
1758 // Remember the beginning of the previous instruction, in case
1759 // it's followed by a flag-kill and a null-check. Happens on
1760 // Intel all the time, with add-to-memory kind of opcodes.
1761 previous_offset = current_offset;
1762 }
1763
1764 // Not an else-if!
1765 // If this is a trap based cmp then add its offset to the list.
1766 if (mach->is_TrapBasedCheckNode()) {
1767 inct_starts[inct_cnt++] = current_offset;
1768 }
1769 }
1770
1771 // Verify that there is sufficient space remaining
1772 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1773 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1774 C->record_failure("CodeCache is full");
1775 return;
1776 }
1777
1778 // Save the offset for the listing
1779 #if defined(SUPPORT_OPTO_ASSEMBLY)
1780 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1781 node_offsets[n->_idx] = masm->offset();
1782 }
1783 #endif
1784 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1785
1786 // "Normal" instruction case
1787 DEBUG_ONLY(uint instr_offset = masm->offset());
1788 n->emit(masm, C->regalloc());
1789 current_offset = masm->offset();
1790
1791 // Above we only verified that there is enough space in the instruction section.
1792 // However, the instruction may emit stubs that cause code buffer expansion.
1793 // Bail out here if expansion failed due to a lack of code cache space.
1794 if (C->failing()) {
1795 return;
1796 }
1797
1798 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1799 "ret_addr_offset() not within emitted code");
1800 #ifdef ASSERT
1801 uint n_size = n->size(C->regalloc());
1802 if (n_size < (current_offset-instr_offset)) {
1803 MachNode* mach = n->as_Mach();
1804 n->dump();
1805 mach->dump_format(C->regalloc(), tty);
1806 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1807 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1808 tty->print_cr(" ------------------- ");
1809 BufferBlob* blob = this->scratch_buffer_blob();
1810 address blob_begin = blob->content_begin();
1811 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1812 assert(false, "wrong size of mach node");
1813 }
1814 #endif
1815 non_safepoints.observe_instruction(n, current_offset);
1816
1817 // mcall is last "call" that can be a safepoint
1818 // record it so we can see if a poll will directly follow it
1819 // in which case we'll need a pad to make the PcDesc sites unique
1820 // see 5010568. This can be slightly inaccurate but conservative
1821 // in the case that return address is not actually at current_offset.
1822 // This is a small price to pay.
1823
1824 if (is_mcall) {
1825 last_call_offset = current_offset;
1826 }
1827
1828 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1829 // Avoid back to back some instructions.
1830 last_avoid_back_to_back_offset = current_offset;
1831 }
1832
1833 // See if this instruction has a delay slot
1834 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1835 guarantee(delay_slot != nullptr, "expecting delay slot node");
1836
1837 // Back up 1 instruction
1838 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1839
1840 // Save the offset for the listing
1841 #if defined(SUPPORT_OPTO_ASSEMBLY)
1842 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1843 node_offsets[delay_slot->_idx] = masm->offset();
1844 }
1845 #endif
1846
1847 // Support a SafePoint in the delay slot
1848 if (delay_slot->is_MachSafePoint()) {
1849 MachNode *mach = delay_slot->as_Mach();
1850 // !!!!! Stubs only need an oopmap right now, so bail out
1851 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1852 // Write the oopmap directly to the code blob??!!
1853 delay_slot = nullptr;
1854 continue;
1855 }
1856
1857 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1858 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1859 adjusted_offset);
1860 // Generate an OopMap entry
1861 Process_OopMap_Node(mach, adjusted_offset);
1862 }
1863
1864 // Insert the delay slot instruction
1865 delay_slot->emit(masm, C->regalloc());
1866
1867 // Don't reuse it
1868 delay_slot = nullptr;
1869 }
1870
1871 } // End for all instructions in block
1872
1873 // If the next block is the top of a loop, pad this block out to align
1874 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1875 if (i < nblocks-1) {
1876 Block *nb = C->cfg()->get_block(i + 1);
1877 int padding = nb->alignment_padding(current_offset);
1878 if( padding > 0 ) {
1879 MachNode *nop = new MachNopNode(padding / nop_size);
1880 block->insert_node(nop, block->number_of_nodes());
1881 C->cfg()->map_node_to_block(nop, block);
1882 nop->emit(masm, C->regalloc());
1883 current_offset = masm->offset();
1884 }
1885 }
1886 // Verify that the distance for generated before forward
1887 // short branches is still valid.
1888 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1889
1890 // Save new block start offset
1891 blk_starts[i] = blk_offset;
1892 } // End of for all blocks
1893 blk_starts[nblocks] = current_offset;
1894
1895 non_safepoints.flush_at_end();
1896
1897 // Offset too large?
1898 if (C->failing()) return;
1899
1900 // Define a pseudo-label at the end of the code
1901 masm->bind( blk_labels[nblocks] );
1902
1903 // Compute the size of the first block
1904 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1905
1906 #ifdef ASSERT
1907 for (uint i = 0; i < nblocks; i++) { // For all blocks
1908 if (jmp_target[i] != 0) {
1909 int br_size = jmp_size[i];
1910 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1911 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1912 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]);
1913 assert(false, "Displacement too large for short jmp");
1914 }
1915 }
1916 }
1917 #endif
1918
1919 if (!masm->code()->finalize_stubs()) {
1920 C->record_failure("CodeCache is full");
1921 return;
1922 }
1923
1924 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1925 bs->emit_stubs(*masm->code());
1926 if (C->failing()) return;
1927
1928 // Fill in stubs.
1929 assert(masm->inst_mark() == nullptr, "should be.");
1930 _stub_list.emit(*masm);
1931 if (C->failing()) return;
1932
1933 #ifndef PRODUCT
1934 // Information on the size of the method, without the extraneous code
1935 Scheduling::increment_method_size(masm->offset());
1936 #endif
1937
1938 // ------------------
1939 // Fill in exception table entries.
1940 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1941
1942 // Only java methods have exception handlers and deopt handlers
1943 // class HandlerImpl is platform-specific and defined in the *.ad files.
1944 if (C->method()) {
1945 // Emit the exception handler code.
1946 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1947 if (C->failing()) {
1948 return; // CodeBuffer::expand failed
1949 }
1950 // Emit the deopt handler code.
1951 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1952
1953 // Emit the MethodHandle deopt handler code (if required).
1954 if (C->has_method_handle_invokes() && !C->failing()) {
1955 // We can use the same code as for the normal deopt handler, we
1956 // just need a different entry point address.
1957 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1958 }
1959 }
1960
1961 // One last check for failed CodeBuffer::expand:
1962 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1963 C->record_failure("CodeCache is full");
1964 return;
1965 }
1966
1967 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1968 if (C->print_assembly()) {
1969 tty->cr();
1970 tty->print_cr("============================= C2-compiled nmethod ==============================");
1971 }
1972 #endif
1973
1974 #if defined(SUPPORT_OPTO_ASSEMBLY)
1975 // Dump the assembly code, including basic-block numbers
1976 if (C->print_assembly()) {
1977 ttyLocker ttyl; // keep the following output all in one block
1978 if (!VMThread::should_terminate()) { // test this under the tty lock
1979 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1980 // We call them first and write to a stringStream, then we retake the lock to
1981 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1982 ResourceMark rm;
1983 stringStream method_metadata_str;
1984 if (C->method() != nullptr) {
1985 C->method()->print_metadata(&method_metadata_str);
1986 }
1987 stringStream dump_asm_str;
1988 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1989
1990 NoSafepointVerifier nsv;
1991 ttyLocker ttyl2;
1992 // This output goes directly to the tty, not the compiler log.
1993 // To enable tools to match it up with the compilation activity,
1994 // be sure to tag this tty output with the compile ID.
1995 if (xtty != nullptr) {
1996 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1997 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1998 }
1999 if (C->method() != nullptr) {
2000 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
2001 tty->print_raw(method_metadata_str.freeze());
2002 } else if (C->stub_name() != nullptr) {
2003 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
2004 }
2005 tty->cr();
2006 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
2007 tty->print_raw(dump_asm_str.freeze());
2008 tty->print_cr("--------------------------------------------------------------------------------");
2009 if (xtty != nullptr) {
2010 xtty->tail("opto_assembly");
2011 }
2012 }
2013 }
2014 #endif
2015 }
2016
2017 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
2018 _inc_table.set_size(cnt);
2019
2020 uint inct_cnt = 0;
2021 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2022 Block* block = C->cfg()->get_block(i);
2023 Node *n = nullptr;
2024 int j;
2025
2026 // Find the branch; ignore trailing NOPs.
2027 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
2028 n = block->get_node(j);
2029 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
2030 break;
2031 }
2032 }
2033
2034 // If we didn't find anything, continue
2035 if (j < 0) {
2036 continue;
2037 }
2038
2039 // Compute ExceptionHandlerTable subtable entry and add it
2040 // (skip empty blocks)
2041 if (n->is_Catch()) {
2042
2043 // Get the offset of the return from the call
2044 uint call_return = call_returns[block->_pre_order];
2045 #ifdef ASSERT
2046 assert( call_return > 0, "no call seen for this basic block" );
2047 while (block->get_node(--j)->is_MachProj()) ;
2048 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
2049 #endif
2050 // last instruction is a CatchNode, find it's CatchProjNodes
2051 int nof_succs = block->_num_succs;
2052 // allocate space
2053 GrowableArray<intptr_t> handler_bcis(nof_succs);
2054 GrowableArray<intptr_t> handler_pcos(nof_succs);
2055 // iterate through all successors
2056 for (int j = 0; j < nof_succs; j++) {
2057 Block* s = block->_succs[j];
2058 bool found_p = false;
2059 for (uint k = 1; k < s->num_preds(); k++) {
2060 Node* pk = s->pred(k);
2061 if (pk->is_CatchProj() && pk->in(0) == n) {
2062 const CatchProjNode* p = pk->as_CatchProj();
2063 found_p = true;
2064 // add the corresponding handler bci & pco information
2065 if (p->_con != CatchProjNode::fall_through_index) {
2066 // p leads to an exception handler (and is not fall through)
2067 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2068 // no duplicates, please
2069 if (!handler_bcis.contains(p->handler_bci())) {
2070 uint block_num = s->non_connector()->_pre_order;
2071 handler_bcis.append(p->handler_bci());
2072 handler_pcos.append(blk_labels[block_num].loc_pos());
2073 }
2074 }
2075 }
2076 }
2077 assert(found_p, "no matching predecessor found");
2078 // Note: Due to empty block removal, one block may have
2079 // several CatchProj inputs, from the same Catch.
2080 }
2081
2082 // Set the offset of the return from the call
2083 assert(handler_bcis.find(-1) != -1, "must have default handler");
2084 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2085 continue;
2086 }
2087
2088 // Handle implicit null exception table updates
2089 if (n->is_MachNullCheck()) {
2090 assert(n->in(1)->as_Mach()->barrier_data() == 0,
2091 "Implicit null checks on memory accesses with barriers are not yet supported");
2092 uint block_num = block->non_connector_successor(0)->_pre_order;
2093 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2094 continue;
2095 }
2096 // Handle implicit exception table updates: trap instructions.
2097 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2098 uint block_num = block->non_connector_successor(0)->_pre_order;
2099 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2100 continue;
2101 }
2102 } // End of for all blocks fill in exception table entries
2103 }
2104
2105 // Static Variables
2106 #ifndef PRODUCT
2107 uint Scheduling::_total_nop_size = 0;
2108 uint Scheduling::_total_method_size = 0;
2109 uint Scheduling::_total_branches = 0;
2110 uint Scheduling::_total_unconditional_delays = 0;
2111 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2112 #endif
2113
2114 // Initializer for class Scheduling
2115
2116 Scheduling::Scheduling(Arena *arena, Compile &compile)
2117 : _arena(arena),
2118 _cfg(compile.cfg()),
2119 _regalloc(compile.regalloc()),
2120 _scheduled(arena),
2121 _available(arena),
2122 _reg_node(arena),
2123 _pinch_free_list(arena),
2124 _next_node(nullptr),
2125 _bundle_instr_count(0),
2126 _bundle_cycle_number(0),
2127 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2128 #ifndef PRODUCT
2129 , _branches(0)
2130 , _unconditional_delays(0)
2131 #endif
2132 {
2133 // Create a MachNopNode
2134 _nop = new MachNopNode();
2135
2136 // Now that the nops are in the array, save the count
2137 // (but allow entries for the nops)
2138 _node_bundling_limit = compile.unique();
2139 uint node_max = _regalloc->node_regs_max_index();
2140
2141 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2142
2143 // This one is persistent within the Compile class
2144 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2145
2146 // Allocate space for fixed-size arrays
2147 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2148 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2149
2150 // Clear the arrays
2151 for (uint i = 0; i < node_max; i++) {
2152 ::new (&_node_bundling_base[i]) Bundle();
2153 }
2154 memset(_uses, 0, node_max * sizeof(short));
2155 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2156
2157 // Clear the bundling information
2158 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2159
2160 // Get the last node
2161 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2162
2163 _next_node = block->get_node(block->number_of_nodes() - 1);
2164 }
2165
2166 #ifndef PRODUCT
2167 // Scheduling destructor
2168 Scheduling::~Scheduling() {
2169 _total_branches += _branches;
2170 _total_unconditional_delays += _unconditional_delays;
2171 }
2172 #endif
2173
2174 // Step ahead "i" cycles
2175 void Scheduling::step(uint i) {
2176
2177 Bundle *bundle = node_bundling(_next_node);
2178 bundle->set_starts_bundle();
2179
2180 // Update the bundle record, but leave the flags information alone
2181 if (_bundle_instr_count > 0) {
2182 bundle->set_instr_count(_bundle_instr_count);
2183 bundle->set_resources_used(_bundle_use.resourcesUsed());
2184 }
2185
2186 // Update the state information
2187 _bundle_instr_count = 0;
2188 _bundle_cycle_number += i;
2189 _bundle_use.step(i);
2190 }
2191
2192 void Scheduling::step_and_clear() {
2193 Bundle *bundle = node_bundling(_next_node);
2194 bundle->set_starts_bundle();
2195
2196 // Update the bundle record
2197 if (_bundle_instr_count > 0) {
2198 bundle->set_instr_count(_bundle_instr_count);
2199 bundle->set_resources_used(_bundle_use.resourcesUsed());
2200
2201 _bundle_cycle_number += 1;
2202 }
2203
2204 // Clear the bundling information
2205 _bundle_instr_count = 0;
2206 _bundle_use.reset();
2207
2208 memcpy(_bundle_use_elements,
2209 Pipeline_Use::elaborated_elements,
2210 sizeof(Pipeline_Use::elaborated_elements));
2211 }
2212
2213 // Perform instruction scheduling and bundling over the sequence of
2214 // instructions in backwards order.
2215 void PhaseOutput::ScheduleAndBundle() {
2216
2217 // Don't optimize this if it isn't a method
2218 if (!C->method())
2219 return;
2220
2221 // Don't optimize this if scheduling is disabled
2222 if (!C->do_scheduling())
2223 return;
2224
2225 // Scheduling code works only with pairs (8 bytes) maximum.
2226 // And when the scalable vector register is used, we may spill/unspill
2227 // the whole reg regardless of the max vector size.
2228 if (C->max_vector_size() > 8 ||
2229 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2230 return;
2231 }
2232
2233 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2234
2235 // Create a data structure for all the scheduling information
2236 Scheduling scheduling(Thread::current()->resource_area(), *C);
2237
2238 // Walk backwards over each basic block, computing the needed alignment
2239 // Walk over all the basic blocks
2240 scheduling.DoScheduling();
2241
2242 #ifndef PRODUCT
2243 if (C->trace_opto_output()) {
2244 // Buffer and print all at once
2245 ResourceMark rm;
2246 stringStream ss;
2247 ss.print("\n---- After ScheduleAndBundle ----\n");
2248 print_scheduling(&ss);
2249 tty->print("%s", ss.as_string());
2250 }
2251 #endif
2252 }
2253
2254 #ifndef PRODUCT
2255 // Separated out so that it can be called directly from debugger
2256 void PhaseOutput::print_scheduling() {
2257 print_scheduling(tty);
2258 }
2259
2260 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2261 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2262 output_stream->print("\nBB#%03d:\n", i);
2263 Block* block = C->cfg()->get_block(i);
2264 for (uint j = 0; j < block->number_of_nodes(); j++) {
2265 Node* n = block->get_node(j);
2266 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2267 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2268 n->dump("\n", false, output_stream);
2269 }
2270 }
2271 }
2272 #endif
2273
2274 // See if this node fits into the present instruction bundle
2275 bool Scheduling::NodeFitsInBundle(Node *n) {
2276 uint n_idx = n->_idx;
2277
2278 // If this is the unconditional delay instruction, then it fits
2279 if (n == _unconditional_delay_slot) {
2280 #ifndef PRODUCT
2281 if (_cfg->C->trace_opto_output())
2282 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2283 #endif
2284 return (true);
2285 }
2286
2287 // If the node cannot be scheduled this cycle, skip it
2288 if (_current_latency[n_idx] > _bundle_cycle_number) {
2289 #ifndef PRODUCT
2290 if (_cfg->C->trace_opto_output())
2291 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2292 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2293 #endif
2294 return (false);
2295 }
2296
2297 const Pipeline *node_pipeline = n->pipeline();
2298
2299 uint instruction_count = node_pipeline->instructionCount();
2300 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2301 instruction_count = 0;
2302 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2303 instruction_count++;
2304
2305 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2306 #ifndef PRODUCT
2307 if (_cfg->C->trace_opto_output())
2308 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2309 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2310 #endif
2311 return (false);
2312 }
2313
2314 // Don't allow non-machine nodes to be handled this way
2315 if (!n->is_Mach() && instruction_count == 0)
2316 return (false);
2317
2318 // See if there is any overlap
2319 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2320
2321 if (delay > 0) {
2322 #ifndef PRODUCT
2323 if (_cfg->C->trace_opto_output())
2324 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2325 #endif
2326 return false;
2327 }
2328
2329 #ifndef PRODUCT
2330 if (_cfg->C->trace_opto_output())
2331 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2332 #endif
2333
2334 return true;
2335 }
2336
2337 Node * Scheduling::ChooseNodeToBundle() {
2338 uint siz = _available.size();
2339
2340 if (siz == 0) {
2341
2342 #ifndef PRODUCT
2343 if (_cfg->C->trace_opto_output())
2344 tty->print("# ChooseNodeToBundle: null\n");
2345 #endif
2346 return (nullptr);
2347 }
2348
2349 // Fast path, if only 1 instruction in the bundle
2350 if (siz == 1) {
2351 #ifndef PRODUCT
2352 if (_cfg->C->trace_opto_output()) {
2353 tty->print("# ChooseNodeToBundle (only 1): ");
2354 _available[0]->dump();
2355 }
2356 #endif
2357 return (_available[0]);
2358 }
2359
2360 // Don't bother, if the bundle is already full
2361 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2362 for ( uint i = 0; i < siz; i++ ) {
2363 Node *n = _available[i];
2364
2365 // Skip projections, we'll handle them another way
2366 if (n->is_Proj())
2367 continue;
2368
2369 // This presupposed that instructions are inserted into the
2370 // available list in a legality order; i.e. instructions that
2371 // must be inserted first are at the head of the list
2372 if (NodeFitsInBundle(n)) {
2373 #ifndef PRODUCT
2374 if (_cfg->C->trace_opto_output()) {
2375 tty->print("# ChooseNodeToBundle: ");
2376 n->dump();
2377 }
2378 #endif
2379 return (n);
2380 }
2381 }
2382 }
2383
2384 // Nothing fits in this bundle, choose the highest priority
2385 #ifndef PRODUCT
2386 if (_cfg->C->trace_opto_output()) {
2387 tty->print("# ChooseNodeToBundle: ");
2388 _available[0]->dump();
2389 }
2390 #endif
2391
2392 return _available[0];
2393 }
2394
2395 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2396 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2397
2398 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2399 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2400 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2401 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2402
2403 // Comparison between stack -> reg and stack -> reg
2404 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2405 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2406 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2407 }
2408
2409 // Comparison between reg -> stack and reg -> stack
2410 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2411 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2412 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2413 }
2414
2415 return 0; // Not comparable
2416 }
2417
2418 void Scheduling::AddNodeToAvailableList(Node *n) {
2419 assert( !n->is_Proj(), "projections never directly made available" );
2420 #ifndef PRODUCT
2421 if (_cfg->C->trace_opto_output()) {
2422 tty->print("# AddNodeToAvailableList: ");
2423 n->dump();
2424 }
2425 #endif
2426
2427 int latency = _current_latency[n->_idx];
2428
2429 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2430 // for stack spilling or unspilling have the same latency, we sort
2431 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2432 // have more opportunities to do ld/st merging
2433 uint i;
2434 for (i = 0; i < _available.size(); i++) {
2435 if (_current_latency[_available[i]->_idx] > latency) {
2436 break;
2437 } else if (_current_latency[_available[i]->_idx] == latency &&
2438 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2439 compare_two_spill_nodes(n, _available[i]) > 0) {
2440 break;
2441 }
2442 }
2443
2444 // Special Check for compares following branches
2445 if( n->is_Mach() && _scheduled.size() > 0 ) {
2446 int op = n->as_Mach()->ideal_Opcode();
2447 Node *last = _scheduled[0];
2448 if( last->is_MachIf() && last->in(1) == n &&
2449 ( op == Op_CmpI ||
2450 op == Op_CmpU ||
2451 op == Op_CmpUL ||
2452 op == Op_CmpP ||
2453 op == Op_CmpF ||
2454 op == Op_CmpD ||
2455 op == Op_CmpL ) ) {
2456
2457 // Recalculate position, moving to front of same latency
2458 for ( i=0 ; i < _available.size(); i++ )
2459 if (_current_latency[_available[i]->_idx] >= latency)
2460 break;
2461 }
2462 }
2463
2464 // Insert the node in the available list
2465 _available.insert(i, n);
2466
2467 #ifndef PRODUCT
2468 if (_cfg->C->trace_opto_output())
2469 dump_available();
2470 #endif
2471 }
2472
2473 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2474 for ( uint i=0; i < n->len(); i++ ) {
2475 Node *def = n->in(i);
2476 if (!def) continue;
2477 if( def->is_Proj() ) // If this is a machine projection, then
2478 def = def->in(0); // propagate usage thru to the base instruction
2479
2480 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2481 continue;
2482 }
2483
2484 // Compute the latency
2485 uint l = _bundle_cycle_number + n->latency(i);
2486 if (_current_latency[def->_idx] < l)
2487 _current_latency[def->_idx] = l;
2488
2489 // If this does not have uses then schedule it
2490 if ((--_uses[def->_idx]) == 0)
2491 AddNodeToAvailableList(def);
2492 }
2493 }
2494
2495 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2496 #ifndef PRODUCT
2497 if (_cfg->C->trace_opto_output()) {
2498 tty->print("# AddNodeToBundle: ");
2499 n->dump();
2500 }
2501 #endif
2502
2503 // Remove this from the available list
2504 uint i;
2505 for (i = 0; i < _available.size(); i++)
2506 if (_available[i] == n)
2507 break;
2508 assert(i < _available.size(), "entry in _available list not found");
2509 _available.remove(i);
2510
2511 // See if this fits in the current bundle
2512 const Pipeline *node_pipeline = n->pipeline();
2513 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2514
2515 // Check for instructions to be placed in the delay slot. We
2516 // do this before we actually schedule the current instruction,
2517 // because the delay slot follows the current instruction.
2518 if (Pipeline::_branch_has_delay_slot &&
2519 node_pipeline->hasBranchDelay() &&
2520 !_unconditional_delay_slot) {
2521
2522 uint siz = _available.size();
2523
2524 // Conditional branches can support an instruction that
2525 // is unconditionally executed and not dependent by the
2526 // branch, OR a conditionally executed instruction if
2527 // the branch is taken. In practice, this means that
2528 // the first instruction at the branch target is
2529 // copied to the delay slot, and the branch goes to
2530 // the instruction after that at the branch target
2531 if ( n->is_MachBranch() ) {
2532
2533 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2534 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2535
2536 #ifndef PRODUCT
2537 _branches++;
2538 #endif
2539
2540 // At least 1 instruction is on the available list
2541 // that is not dependent on the branch
2542 for (uint i = 0; i < siz; i++) {
2543 Node *d = _available[i];
2544 const Pipeline *avail_pipeline = d->pipeline();
2545
2546 // Don't allow safepoints in the branch shadow, that will
2547 // cause a number of difficulties
2548 if ( avail_pipeline->instructionCount() == 1 &&
2549 !avail_pipeline->hasMultipleBundles() &&
2550 !avail_pipeline->hasBranchDelay() &&
2551 Pipeline::instr_has_unit_size() &&
2552 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2553 NodeFitsInBundle(d) &&
2554 !node_bundling(d)->used_in_delay()) {
2555
2556 if (d->is_Mach() && !d->is_MachSafePoint()) {
2557 // A node that fits in the delay slot was found, so we need to
2558 // set the appropriate bits in the bundle pipeline information so
2559 // that it correctly indicates resource usage. Later, when we
2560 // attempt to add this instruction to the bundle, we will skip
2561 // setting the resource usage.
2562 _unconditional_delay_slot = d;
2563 node_bundling(n)->set_use_unconditional_delay();
2564 node_bundling(d)->set_used_in_unconditional_delay();
2565 _bundle_use.add_usage(avail_pipeline->resourceUse());
2566 _current_latency[d->_idx] = _bundle_cycle_number;
2567 _next_node = d;
2568 ++_bundle_instr_count;
2569 #ifndef PRODUCT
2570 _unconditional_delays++;
2571 #endif
2572 break;
2573 }
2574 }
2575 }
2576 }
2577
2578 // No delay slot, add a nop to the usage
2579 if (!_unconditional_delay_slot) {
2580 // See if adding an instruction in the delay slot will overflow
2581 // the bundle.
2582 if (!NodeFitsInBundle(_nop)) {
2583 #ifndef PRODUCT
2584 if (_cfg->C->trace_opto_output())
2585 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2586 #endif
2587 step(1);
2588 }
2589
2590 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2591 _next_node = _nop;
2592 ++_bundle_instr_count;
2593 }
2594
2595 // See if the instruction in the delay slot requires a
2596 // step of the bundles
2597 if (!NodeFitsInBundle(n)) {
2598 #ifndef PRODUCT
2599 if (_cfg->C->trace_opto_output())
2600 tty->print("# *** STEP(branch won't fit) ***\n");
2601 #endif
2602 // Update the state information
2603 _bundle_instr_count = 0;
2604 _bundle_cycle_number += 1;
2605 _bundle_use.step(1);
2606 }
2607 }
2608
2609 // Get the number of instructions
2610 uint instruction_count = node_pipeline->instructionCount();
2611 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2612 instruction_count = 0;
2613
2614 // Compute the latency information
2615 uint delay = 0;
2616
2617 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2618 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2619 if (relative_latency < 0)
2620 relative_latency = 0;
2621
2622 delay = _bundle_use.full_latency(relative_latency, node_usage);
2623
2624 // Does not fit in this bundle, start a new one
2625 if (delay > 0) {
2626 step(delay);
2627
2628 #ifndef PRODUCT
2629 if (_cfg->C->trace_opto_output())
2630 tty->print("# *** STEP(%d) ***\n", delay);
2631 #endif
2632 }
2633 }
2634
2635 // If this was placed in the delay slot, ignore it
2636 if (n != _unconditional_delay_slot) {
2637
2638 if (delay == 0) {
2639 if (node_pipeline->hasMultipleBundles()) {
2640 #ifndef PRODUCT
2641 if (_cfg->C->trace_opto_output())
2642 tty->print("# *** STEP(multiple instructions) ***\n");
2643 #endif
2644 step(1);
2645 }
2646
2647 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2648 #ifndef PRODUCT
2649 if (_cfg->C->trace_opto_output())
2650 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2651 instruction_count + _bundle_instr_count,
2652 Pipeline::_max_instrs_per_cycle);
2653 #endif
2654 step(1);
2655 }
2656 }
2657
2658 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2659 _bundle_instr_count++;
2660
2661 // Set the node's latency
2662 _current_latency[n->_idx] = _bundle_cycle_number;
2663
2664 // Now merge the functional unit information
2665 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2666 _bundle_use.add_usage(node_usage);
2667
2668 // Increment the number of instructions in this bundle
2669 _bundle_instr_count += instruction_count;
2670
2671 // Remember this node for later
2672 if (n->is_Mach())
2673 _next_node = n;
2674 }
2675
2676 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2677 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2678 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2679 // into the block. All other scheduled nodes get put in the schedule here.
2680 int op = n->Opcode();
2681 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2682 (op != Op_Node && // Not an unused antidepedence node and
2683 // not an unallocated boxlock
2684 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2685
2686 // Push any trailing projections
2687 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2688 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2689 Node *foi = n->fast_out(i);
2690 if( foi->is_Proj() )
2691 _scheduled.push(foi);
2692 }
2693 }
2694
2695 // Put the instruction in the schedule list
2696 _scheduled.push(n);
2697 }
2698
2699 #ifndef PRODUCT
2700 if (_cfg->C->trace_opto_output())
2701 dump_available();
2702 #endif
2703
2704 // Walk all the definitions, decrementing use counts, and
2705 // if a definition has a 0 use count, place it in the available list.
2706 DecrementUseCounts(n,bb);
2707 }
2708
2709 // This method sets the use count within a basic block. We will ignore all
2710 // uses outside the current basic block. As we are doing a backwards walk,
2711 // any node we reach that has a use count of 0 may be scheduled. This also
2712 // avoids the problem of cyclic references from phi nodes, as long as phi
2713 // nodes are at the front of the basic block. This method also initializes
2714 // the available list to the set of instructions that have no uses within this
2715 // basic block.
2716 void Scheduling::ComputeUseCount(const Block *bb) {
2717 #ifndef PRODUCT
2718 if (_cfg->C->trace_opto_output())
2719 tty->print("# -> ComputeUseCount\n");
2720 #endif
2721
2722 // Clear the list of available and scheduled instructions, just in case
2723 _available.clear();
2724 _scheduled.clear();
2725
2726 // No delay slot specified
2727 _unconditional_delay_slot = nullptr;
2728
2729 #ifdef ASSERT
2730 for( uint i=0; i < bb->number_of_nodes(); i++ )
2731 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2732 #endif
2733
2734 // Force the _uses count to never go to zero for unscheduable pieces
2735 // of the block
2736 for( uint k = 0; k < _bb_start; k++ )
2737 _uses[bb->get_node(k)->_idx] = 1;
2738 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2739 _uses[bb->get_node(l)->_idx] = 1;
2740
2741 // Iterate backwards over the instructions in the block. Don't count the
2742 // branch projections at end or the block header instructions.
2743 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2744 Node *n = bb->get_node(j);
2745 if( n->is_Proj() ) continue; // Projections handled another way
2746
2747 // Account for all uses
2748 for ( uint k = 0; k < n->len(); k++ ) {
2749 Node *inp = n->in(k);
2750 if (!inp) continue;
2751 assert(inp != n, "no cycles allowed" );
2752 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2753 if (inp->is_Proj()) { // Skip through Proj's
2754 inp = inp->in(0);
2755 }
2756 ++_uses[inp->_idx]; // Count 1 block-local use
2757 }
2758 }
2759
2760 // If this instruction has a 0 use count, then it is available
2761 if (!_uses[n->_idx]) {
2762 _current_latency[n->_idx] = _bundle_cycle_number;
2763 AddNodeToAvailableList(n);
2764 }
2765
2766 #ifndef PRODUCT
2767 if (_cfg->C->trace_opto_output()) {
2768 tty->print("# uses: %3d: ", _uses[n->_idx]);
2769 n->dump();
2770 }
2771 #endif
2772 }
2773
2774 #ifndef PRODUCT
2775 if (_cfg->C->trace_opto_output())
2776 tty->print("# <- ComputeUseCount\n");
2777 #endif
2778 }
2779
2780 // This routine performs scheduling on each basic block in reverse order,
2781 // using instruction latencies and taking into account function unit
2782 // availability.
2783 void Scheduling::DoScheduling() {
2784 #ifndef PRODUCT
2785 if (_cfg->C->trace_opto_output())
2786 tty->print("# -> DoScheduling\n");
2787 #endif
2788
2789 Block *succ_bb = nullptr;
2790 Block *bb;
2791 Compile* C = Compile::current();
2792
2793 // Walk over all the basic blocks in reverse order
2794 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2795 bb = _cfg->get_block(i);
2796
2797 #ifndef PRODUCT
2798 if (_cfg->C->trace_opto_output()) {
2799 tty->print("# Schedule BB#%03d (initial)\n", i);
2800 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2801 bb->get_node(j)->dump();
2802 }
2803 }
2804 #endif
2805
2806 // On the head node, skip processing
2807 if (bb == _cfg->get_root_block()) {
2808 continue;
2809 }
2810
2811 // Skip empty, connector blocks
2812 if (bb->is_connector())
2813 continue;
2814
2815 // If the following block is not the sole successor of
2816 // this one, then reset the pipeline information
2817 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2818 #ifndef PRODUCT
2819 if (_cfg->C->trace_opto_output()) {
2820 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2821 _next_node->_idx, _bundle_instr_count);
2822 }
2823 #endif
2824 step_and_clear();
2825 }
2826
2827 // Leave untouched the starting instruction, any Phis, a CreateEx node
2828 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2829 _bb_end = bb->number_of_nodes()-1;
2830 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2831 Node *n = bb->get_node(_bb_start);
2832 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2833 // Also, MachIdealNodes do not get scheduled
2834 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2835 MachNode *mach = n->as_Mach();
2836 int iop = mach->ideal_Opcode();
2837 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2838 if( iop == Op_Con ) continue; // Do not schedule Top
2839 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2840 mach->pipeline() == MachNode::pipeline_class() &&
2841 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2842 continue;
2843 break; // Funny loop structure to be sure...
2844 }
2845 // Compute last "interesting" instruction in block - last instruction we
2846 // might schedule. _bb_end points just after last schedulable inst. We
2847 // normally schedule conditional branches (despite them being forced last
2848 // in the block), because they have delay slots we can fill. Calls all
2849 // have their delay slots filled in the template expansions, so we don't
2850 // bother scheduling them.
2851 Node *last = bb->get_node(_bb_end);
2852 // Ignore trailing NOPs.
2853 while (_bb_end > 0 && last->is_Mach() &&
2854 last->as_Mach()->ideal_Opcode() == Op_Con) {
2855 last = bb->get_node(--_bb_end);
2856 }
2857 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2858 if( last->is_Catch() ||
2859 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2860 // There might be a prior call. Skip it.
2861 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2862 } else if( last->is_MachNullCheck() ) {
2863 // Backup so the last null-checked memory instruction is
2864 // outside the schedulable range. Skip over the nullcheck,
2865 // projection, and the memory nodes.
2866 Node *mem = last->in(1);
2867 do {
2868 _bb_end--;
2869 } while (mem != bb->get_node(_bb_end));
2870 } else {
2871 // Set _bb_end to point after last schedulable inst.
2872 _bb_end++;
2873 }
2874
2875 assert( _bb_start <= _bb_end, "inverted block ends" );
2876
2877 // Compute the register antidependencies for the basic block
2878 ComputeRegisterAntidependencies(bb);
2879 if (C->failing()) return; // too many D-U pinch points
2880
2881 // Compute the usage within the block, and set the list of all nodes
2882 // in the block that have no uses within the block.
2883 ComputeUseCount(bb);
2884
2885 // Schedule the remaining instructions in the block
2886 while ( _available.size() > 0 ) {
2887 Node *n = ChooseNodeToBundle();
2888 guarantee(n != nullptr, "no nodes available");
2889 AddNodeToBundle(n,bb);
2890 }
2891
2892 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2893 #ifdef ASSERT
2894 for( uint l = _bb_start; l < _bb_end; l++ ) {
2895 Node *n = bb->get_node(l);
2896 uint m;
2897 for( m = 0; m < _bb_end-_bb_start; m++ )
2898 if( _scheduled[m] == n )
2899 break;
2900 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2901 }
2902 #endif
2903
2904 // Now copy the instructions (in reverse order) back to the block
2905 for ( uint k = _bb_start; k < _bb_end; k++ )
2906 bb->map_node(_scheduled[_bb_end-k-1], k);
2907
2908 #ifndef PRODUCT
2909 if (_cfg->C->trace_opto_output()) {
2910 tty->print("# Schedule BB#%03d (final)\n", i);
2911 uint current = 0;
2912 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2913 Node *n = bb->get_node(j);
2914 if( valid_bundle_info(n) ) {
2915 Bundle *bundle = node_bundling(n);
2916 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2917 tty->print("*** Bundle: ");
2918 bundle->dump();
2919 }
2920 n->dump();
2921 }
2922 }
2923 }
2924 #endif
2925 #ifdef ASSERT
2926 verify_good_schedule(bb,"after block local scheduling");
2927 #endif
2928 }
2929
2930 #ifndef PRODUCT
2931 if (_cfg->C->trace_opto_output())
2932 tty->print("# <- DoScheduling\n");
2933 #endif
2934
2935 // Record final node-bundling array location
2936 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2937
2938 } // end DoScheduling
2939
2940 // Verify that no live-range used in the block is killed in the block by a
2941 // wrong DEF. This doesn't verify live-ranges that span blocks.
2942
2943 // Check for edge existence. Used to avoid adding redundant precedence edges.
2944 static bool edge_from_to( Node *from, Node *to ) {
2945 for( uint i=0; i<from->len(); i++ )
2946 if( from->in(i) == to )
2947 return true;
2948 return false;
2949 }
2950
2951 #ifdef ASSERT
2952 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2953 // Check for bad kills
2954 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2955 Node *prior_use = _reg_node[def];
2956 if( prior_use && !edge_from_to(prior_use,n) ) {
2957 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2958 n->dump();
2959 tty->print_cr("...");
2960 prior_use->dump();
2961 assert(edge_from_to(prior_use,n), "%s", msg);
2962 }
2963 _reg_node.map(def,nullptr); // Kill live USEs
2964 }
2965 }
2966
2967 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2968
2969 // Zap to something reasonable for the verify code
2970 _reg_node.clear();
2971
2972 // Walk over the block backwards. Check to make sure each DEF doesn't
2973 // kill a live value (other than the one it's supposed to). Add each
2974 // USE to the live set.
2975 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2976 Node *n = b->get_node(i);
2977 int n_op = n->Opcode();
2978 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2979 // Fat-proj kills a slew of registers
2980 RegMaskIterator rmi(n->out_RegMask());
2981 while (rmi.has_next()) {
2982 OptoReg::Name kill = rmi.next();
2983 verify_do_def(n, kill, msg);
2984 }
2985 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2986 // Get DEF'd registers the normal way
2987 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2988 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2989 }
2990
2991 // Now make all USEs live
2992 for( uint i=1; i<n->req(); i++ ) {
2993 Node *def = n->in(i);
2994 assert(def != nullptr, "input edge required");
2995 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2996 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2997 if( OptoReg::is_valid(reg_lo) ) {
2998 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2999 _reg_node.map(reg_lo,n);
3000 }
3001 if( OptoReg::is_valid(reg_hi) ) {
3002 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
3003 _reg_node.map(reg_hi,n);
3004 }
3005 }
3006
3007 }
3008
3009 // Zap to something reasonable for the Antidependence code
3010 _reg_node.clear();
3011 }
3012 #endif
3013
3014 // Conditionally add precedence edges. Avoid putting edges on Projs.
3015 static void add_prec_edge_from_to( Node *from, Node *to ) {
3016 if( from->is_Proj() ) { // Put precedence edge on Proj's input
3017 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
3018 from = from->in(0);
3019 }
3020 if( from != to && // No cycles (for things like LD L0,[L0+4] )
3021 !edge_from_to( from, to ) ) // Avoid duplicate edge
3022 from->add_prec(to);
3023 }
3024
3025 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
3026 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
3027 return;
3028
3029 if (OptoReg::is_reg(def_reg)) {
3030 VMReg vmreg = OptoReg::as_VMReg(def_reg);
3031 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
3032 // This is one of the high slots of a vector register.
3033 // ScheduleAndBundle already checked there are no live wide
3034 // vectors in this method so it can be safely ignored.
3035 return;
3036 }
3037 }
3038
3039 Node *pinch = _reg_node[def_reg]; // Get pinch point
3040 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
3041 is_def ) { // Check for a true def (not a kill)
3042 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
3043 return;
3044 }
3045
3046 Node *kill = def; // Rename 'def' to more descriptive 'kill'
3047 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
3048
3049 // After some number of kills there _may_ be a later def
3050 Node *later_def = nullptr;
3051
3052 Compile* C = Compile::current();
3053
3054 // Finding a kill requires a real pinch-point.
3055 // Check for not already having a pinch-point.
3056 // Pinch points are Op_Node's.
3057 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
3058 later_def = pinch; // Must be def/kill as optimistic pinch-point
3059 if ( _pinch_free_list.size() > 0) {
3060 pinch = _pinch_free_list.pop();
3061 } else {
3062 pinch = new Node(1); // Pinch point to-be
3063 }
3064 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
3065 DEBUG_ONLY( pinch->dump(); );
3066 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
3067 _cfg->C->record_method_not_compilable("too many D-U pinch points");
3068 return;
3069 }
3070 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3071 _reg_node.map(def_reg,pinch); // Record pinch-point
3072 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3073 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3074 pinch->init_req(0, C->top()); // set not null for the next call
3075 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3076 later_def = nullptr; // and no later def
3077 }
3078 pinch->set_req(0,later_def); // Hook later def so we can find it
3079 } else { // Else have valid pinch point
3080 if( pinch->in(0) ) // If there is a later-def
3081 later_def = pinch->in(0); // Get it
3082 }
3083
3084 // Add output-dependence edge from later def to kill
3085 if( later_def ) // If there is some original def
3086 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3087
3088 // See if current kill is also a use, and so is forced to be the pinch-point.
3089 if( pinch->Opcode() == Op_Node ) {
3090 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3091 for( uint i=1; i<uses->req(); i++ ) {
3092 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3093 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3094 // Yes, found a use/kill pinch-point
3095 pinch->set_req(0,nullptr); //
3096 pinch->replace_by(kill); // Move anti-dep edges up
3097 pinch = kill;
3098 _reg_node.map(def_reg,pinch);
3099 return;
3100 }
3101 }
3102 }
3103
3104 // Add edge from kill to pinch-point
3105 add_prec_edge_from_to(kill,pinch);
3106 }
3107
3108 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3109 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3110 return;
3111 Node *pinch = _reg_node[use_reg]; // Get pinch point
3112 // Check for no later def_reg/kill in block
3113 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3114 // Use has to be block-local as well
3115 _cfg->get_block_for_node(use) == b) {
3116 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3117 pinch->req() == 1 ) { // pinch not yet in block?
3118 pinch->del_req(0); // yank pointer to later-def, also set flag
3119 // Insert the pinch-point in the block just after the last use
3120 b->insert_node(pinch, b->find_node(use) + 1);
3121 _bb_end++; // Increase size scheduled region in block
3122 }
3123
3124 add_prec_edge_from_to(pinch,use);
3125 }
3126 }
3127
3128 // We insert antidependences between the reads and following write of
3129 // allocated registers to prevent illegal code motion. Hopefully, the
3130 // number of added references should be fairly small, especially as we
3131 // are only adding references within the current basic block.
3132 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3133
3134 #ifdef ASSERT
3135 verify_good_schedule(b,"before block local scheduling");
3136 #endif
3137
3138 // A valid schedule, for each register independently, is an endless cycle
3139 // of: a def, then some uses (connected to the def by true dependencies),
3140 // then some kills (defs with no uses), finally the cycle repeats with a new
3141 // def. The uses are allowed to float relative to each other, as are the
3142 // kills. No use is allowed to slide past a kill (or def). This requires
3143 // antidependencies between all uses of a single def and all kills that
3144 // follow, up to the next def. More edges are redundant, because later defs
3145 // & kills are already serialized with true or antidependencies. To keep
3146 // the edge count down, we add a 'pinch point' node if there's more than
3147 // one use or more than one kill/def.
3148
3149 // We add dependencies in one bottom-up pass.
3150
3151 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3152
3153 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3154 // register. If not, we record the DEF/KILL in _reg_node, the
3155 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3156 // "pinch point", a new Node that's in the graph but not in the block.
3157 // We put edges from the prior and current DEF/KILLs to the pinch point.
3158 // We put the pinch point in _reg_node. If there's already a pinch point
3159 // we merely add an edge from the current DEF/KILL to the pinch point.
3160
3161 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3162 // put an edge from the pinch point to the USE.
3163
3164 // To be expedient, the _reg_node array is pre-allocated for the whole
3165 // compilation. _reg_node is lazily initialized; it either contains a null,
3166 // or a valid def/kill/pinch-point, or a leftover node from some prior
3167 // block. Leftover node from some prior block is treated like a null (no
3168 // prior def, so no anti-dependence needed). Valid def is distinguished by
3169 // it being in the current block.
3170 bool fat_proj_seen = false;
3171 uint last_safept = _bb_end-1;
3172 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3173 Node* last_safept_node = end_node;
3174 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3175 Node *n = b->get_node(i);
3176 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3177 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3178 // Fat-proj kills a slew of registers
3179 // This can add edges to 'n' and obscure whether or not it was a def,
3180 // hence the is_def flag.
3181 fat_proj_seen = true;
3182 RegMaskIterator rmi(n->out_RegMask());
3183 while (rmi.has_next()) {
3184 OptoReg::Name kill = rmi.next();
3185 anti_do_def(b, n, kill, is_def);
3186 }
3187 } else {
3188 // Get DEF'd registers the normal way
3189 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3190 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3191 }
3192
3193 // Kill projections on a branch should appear to occur on the
3194 // branch, not afterwards, so grab the masks from the projections
3195 // and process them.
3196 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3197 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3198 Node* use = n->fast_out(i);
3199 if (use->is_Proj()) {
3200 RegMaskIterator rmi(use->out_RegMask());
3201 while (rmi.has_next()) {
3202 OptoReg::Name kill = rmi.next();
3203 anti_do_def(b, n, kill, false);
3204 }
3205 }
3206 }
3207 }
3208
3209 // Check each register used by this instruction for a following DEF/KILL
3210 // that must occur afterward and requires an anti-dependence edge.
3211 for( uint j=0; j<n->req(); j++ ) {
3212 Node *def = n->in(j);
3213 if( def ) {
3214 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3215 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3216 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3217 }
3218 }
3219 // Do not allow defs of new derived values to float above GC
3220 // points unless the base is definitely available at the GC point.
3221
3222 Node *m = b->get_node(i);
3223
3224 // Add precedence edge from following safepoint to use of derived pointer
3225 if( last_safept_node != end_node &&
3226 m != last_safept_node) {
3227 for (uint k = 1; k < m->req(); k++) {
3228 const Type *t = m->in(k)->bottom_type();
3229 if( t->isa_oop_ptr() &&
3230 t->is_ptr()->offset() != 0 ) {
3231 last_safept_node->add_prec( m );
3232 break;
3233 }
3234 }
3235
3236 // Do not allow a CheckCastPP node whose input is a raw pointer to
3237 // float past a safepoint. This can occur when a buffered inline
3238 // type is allocated in a loop and the CheckCastPP from that
3239 // allocation is reused outside the loop. If the use inside the
3240 // loop is scalarized the CheckCastPP will no longer be connected
3241 // to the loop safepoint. See JDK-8264340.
3242 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3243 Node *def = m->in(1);
3244 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3245 last_safept_node->add_prec(m);
3246 }
3247 }
3248 }
3249
3250 if( n->jvms() ) { // Precedence edge from derived to safept
3251 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3252 if( b->get_node(last_safept) != last_safept_node ) {
3253 last_safept = b->find_node(last_safept_node);
3254 }
3255 for( uint j=last_safept; j > i; j-- ) {
3256 Node *mach = b->get_node(j);
3257 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3258 mach->add_prec( n );
3259 }
3260 last_safept = i;
3261 last_safept_node = m;
3262 }
3263 }
3264
3265 if (fat_proj_seen) {
3266 // Garbage collect pinch nodes that were not consumed.
3267 // They are usually created by a fat kill MachProj for a call.
3268 garbage_collect_pinch_nodes();
3269 }
3270 }
3271
3272 // Garbage collect pinch nodes for reuse by other blocks.
3273 //
3274 // The block scheduler's insertion of anti-dependence
3275 // edges creates many pinch nodes when the block contains
3276 // 2 or more Calls. A pinch node is used to prevent a
3277 // combinatorial explosion of edges. If a set of kills for a
3278 // register is anti-dependent on a set of uses (or defs), rather
3279 // than adding an edge in the graph between each pair of kill
3280 // and use (or def), a pinch is inserted between them:
3281 //
3282 // use1 use2 use3
3283 // \ | /
3284 // \ | /
3285 // pinch
3286 // / | \
3287 // / | \
3288 // kill1 kill2 kill3
3289 //
3290 // One pinch node is created per register killed when
3291 // the second call is encountered during a backwards pass
3292 // over the block. Most of these pinch nodes are never
3293 // wired into the graph because the register is never
3294 // used or def'ed in the block.
3295 //
3296 void Scheduling::garbage_collect_pinch_nodes() {
3297 #ifndef PRODUCT
3298 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3299 #endif
3300 int trace_cnt = 0;
3301 for (uint k = 0; k < _reg_node.max(); k++) {
3302 Node* pinch = _reg_node[k];
3303 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3304 // no predecence input edges
3305 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3306 cleanup_pinch(pinch);
3307 _pinch_free_list.push(pinch);
3308 _reg_node.map(k, nullptr);
3309 #ifndef PRODUCT
3310 if (_cfg->C->trace_opto_output()) {
3311 trace_cnt++;
3312 if (trace_cnt > 40) {
3313 tty->print("\n");
3314 trace_cnt = 0;
3315 }
3316 tty->print(" %d", pinch->_idx);
3317 }
3318 #endif
3319 }
3320 }
3321 #ifndef PRODUCT
3322 if (_cfg->C->trace_opto_output()) tty->print("\n");
3323 #endif
3324 }
3325
3326 // Clean up a pinch node for reuse.
3327 void Scheduling::cleanup_pinch( Node *pinch ) {
3328 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3329
3330 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3331 Node* use = pinch->last_out(i);
3332 uint uses_found = 0;
3333 for (uint j = use->req(); j < use->len(); j++) {
3334 if (use->in(j) == pinch) {
3335 use->rm_prec(j);
3336 uses_found++;
3337 }
3338 }
3339 assert(uses_found > 0, "must be a precedence edge");
3340 i -= uses_found; // we deleted 1 or more copies of this edge
3341 }
3342 // May have a later_def entry
3343 pinch->set_req(0, nullptr);
3344 }
3345
3346 #ifndef PRODUCT
3347
3348 void Scheduling::dump_available() const {
3349 tty->print("#Availist ");
3350 for (uint i = 0; i < _available.size(); i++)
3351 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3352 tty->cr();
3353 }
3354
3355 // Print Scheduling Statistics
3356 void Scheduling::print_statistics() {
3357 // Print the size added by nops for bundling
3358 tty->print("Nops added %d bytes to total of %d bytes",
3359 _total_nop_size, _total_method_size);
3360 if (_total_method_size > 0)
3361 tty->print(", for %.2f%%",
3362 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3363 tty->print("\n");
3364
3365 // Print the number of branch shadows filled
3366 if (Pipeline::_branch_has_delay_slot) {
3367 tty->print("Of %d branches, %d had unconditional delay slots filled",
3368 _total_branches, _total_unconditional_delays);
3369 if (_total_branches > 0)
3370 tty->print(", for %.2f%%",
3371 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3372 tty->print("\n");
3373 }
3374
3375 uint total_instructions = 0, total_bundles = 0;
3376
3377 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3378 uint bundle_count = _total_instructions_per_bundle[i];
3379 total_instructions += bundle_count * i;
3380 total_bundles += bundle_count;
3381 }
3382
3383 if (total_bundles > 0)
3384 tty->print("Average ILP (excluding nops) is %.2f\n",
3385 ((double)total_instructions) / ((double)total_bundles));
3386 }
3387 #endif
3388
3389 //-----------------------init_scratch_buffer_blob------------------------------
3390 // Construct a temporary BufferBlob and cache it for this compile.
3391 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3392 // If there is already a scratch buffer blob allocated and the
3393 // constant section is big enough, use it. Otherwise free the
3394 // current and allocate a new one.
3395 BufferBlob* blob = scratch_buffer_blob();
3396 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3397 // Use the current blob.
3398 } else {
3399 if (blob != nullptr) {
3400 BufferBlob::free(blob);
3401 }
3402
3403 ResourceMark rm;
3404 _scratch_const_size = const_size;
3405 int size = C2Compiler::initial_code_buffer_size(const_size);
3406 if (C->has_scalarized_args()) {
3407 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3408 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3409 ciMethod* method = C->method();
3410 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3411 int arg_num = 0;
3412 if (!method->is_static()) {
3413 if (method->is_scalarized_arg(arg_num)) {
3414 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3415 }
3416 arg_num++;
3417 }
3418 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3419 if (method->is_scalarized_arg(arg_num)) {
3420 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3421 }
3422 arg_num++;
3423 }
3424 }
3425 blob = BufferBlob::create("Compile::scratch_buffer", size);
3426 // Record the buffer blob for next time.
3427 set_scratch_buffer_blob(blob);
3428 // Have we run out of code space?
3429 if (scratch_buffer_blob() == nullptr) {
3430 // Let CompilerBroker disable further compilations.
3431 C->record_failure("Not enough space for scratch buffer in CodeCache");
3432 return;
3433 }
3434 }
3435
3436 // Initialize the relocation buffers
3437 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3438 set_scratch_locs_memory(locs_buf);
3439 }
3440
3441
3442 //-----------------------scratch_emit_size-------------------------------------
3443 // Helper function that computes size by emitting code
3444 uint PhaseOutput::scratch_emit_size(const Node* n) {
3445 // Start scratch_emit_size section.
3446 set_in_scratch_emit_size(true);
3447
3448 // Emit into a trash buffer and count bytes emitted.
3449 // This is a pretty expensive way to compute a size,
3450 // but it works well enough if seldom used.
3451 // All common fixed-size instructions are given a size
3452 // method by the AD file.
3453 // Note that the scratch buffer blob and locs memory are
3454 // allocated at the beginning of the compile task, and
3455 // may be shared by several calls to scratch_emit_size.
3456 // The allocation of the scratch buffer blob is particularly
3457 // expensive, since it has to grab the code cache lock.
3458 BufferBlob* blob = this->scratch_buffer_blob();
3459 assert(blob != nullptr, "Initialize BufferBlob at start");
3460 assert(blob->size() > MAX_inst_size, "sanity");
3461 relocInfo* locs_buf = scratch_locs_memory();
3462 address blob_begin = blob->content_begin();
3463 address blob_end = (address)locs_buf;
3464 assert(blob->contains(blob_end), "sanity");
3465 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3466 buf.initialize_consts_size(_scratch_const_size);
3467 buf.initialize_stubs_size(MAX_stubs_size);
3468 assert(locs_buf != nullptr, "sanity");
3469 int lsize = MAX_locs_size / 3;
3470 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3471 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3472 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3473 // Mark as scratch buffer.
3474 buf.consts()->set_scratch_emit();
3475 buf.insts()->set_scratch_emit();
3476 buf.stubs()->set_scratch_emit();
3477
3478 // Do the emission.
3479
3480 Label fakeL; // Fake label for branch instructions.
3481 Label* saveL = nullptr;
3482 uint save_bnum = 0;
3483 bool is_branch = n->is_MachBranch();
3484 C2_MacroAssembler masm(&buf);
3485 masm.bind(fakeL);
3486 if (is_branch) {
3487 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3488 n->as_MachBranch()->label_set(&fakeL, 0);
3489 }
3490 n->emit(&masm, C->regalloc());
3491
3492 // Emitting into the scratch buffer should not fail
3493 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3494
3495 // Restore label.
3496 if (is_branch) {
3497 n->as_MachBranch()->label_set(saveL, save_bnum);
3498 }
3499
3500 // End scratch_emit_size section.
3501 set_in_scratch_emit_size(false);
3502
3503 return buf.insts_size();
3504 }
3505
3506 void PhaseOutput::install() {
3507 if (!C->should_install_code()) {
3508 return;
3509 } else if (C->stub_function() != nullptr) {
3510 install_stub(C->stub_name());
3511 } else {
3512 install_code(C->method(),
3513 C->entry_bci(),
3514 CompileBroker::compiler2(),
3515 C->has_unsafe_access(),
3516 SharedRuntime::is_wide_vector(C->max_vector_size()));
3517 }
3518 }
3519
3520 void PhaseOutput::install_code(ciMethod* target,
3521 int entry_bci,
3522 AbstractCompiler* compiler,
3523 bool has_unsafe_access,
3524 bool has_wide_vectors) {
3525 // Check if we want to skip execution of all compiled code.
3526 {
3527 #ifndef PRODUCT
3528 if (OptoNoExecute) {
3529 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3530 return;
3531 }
3532 #endif
3533 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3534
3535 if (C->is_osr_compilation()) {
3536 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3537 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3538 } else {
3539 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3540 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3541 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3542 }
3543 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3544 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3545 }
3546 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3547 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3548 }
3549 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3550 }
3551
3552 C->env()->register_method(target,
3553 entry_bci,
3554 &_code_offsets,
3555 _orig_pc_slot_offset_in_bytes,
3556 code_buffer(),
3557 frame_size_in_words(),
3558 _oop_map_set,
3559 &_handler_table,
3560 inc_table(),
3561 compiler,
3562 has_unsafe_access,
3563 SharedRuntime::is_wide_vector(C->max_vector_size()),
3564 C->has_monitors(),
3565 C->has_scoped_access(),
3566 0);
3567
3568 if (C->log() != nullptr) { // Print code cache state into compiler log
3569 C->log()->code_cache_state();
3570 }
3571 }
3572 }
3573 void PhaseOutput::install_stub(const char* stub_name) {
3574 // Entry point will be accessed using stub_entry_point();
3575 if (code_buffer() == nullptr) {
3576 Matcher::soft_match_failure();
3577 } else {
3578 if (PrintAssembly && (WizardMode || Verbose))
3579 tty->print_cr("### Stub::%s", stub_name);
3580
3581 if (!C->failing()) {
3582 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3583
3584 // Make the NMethod
3585 // For now we mark the frame as never safe for profile stackwalking
3586 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3587 code_buffer(),
3588 CodeOffsets::frame_never_safe,
3589 // _code_offsets.value(CodeOffsets::Frame_Complete),
3590 frame_size_in_words(),
3591 oop_map_set(),
3592 false);
3593 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3594
3595 C->set_stub_entry_point(rs->entry_point());
3596 }
3597 }
3598 }
3599
3600 // Support for bundling info
3601 Bundle* PhaseOutput::node_bundling(const Node *n) {
3602 assert(valid_bundle_info(n), "oob");
3603 return &_node_bundling_base[n->_idx];
3604 }
3605
3606 bool PhaseOutput::valid_bundle_info(const Node *n) {
3607 return (_node_bundling_limit > n->_idx);
3608 }
3609
3610 //------------------------------frame_size_in_words-----------------------------
3611 // frame_slots in units of words
3612 int PhaseOutput::frame_size_in_words() const {
3613 // shift is 0 in LP32 and 1 in LP64
3614 const int shift = (LogBytesPerWord - LogBytesPerInt);
3615 int words = _frame_slots >> shift;
3616 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3617 return words;
3618 }
3619
3620 // To bang the stack of this compiled method we use the stack size
3621 // that the interpreter would need in case of a deoptimization. This
3622 // removes the need to bang the stack in the deoptimization blob which
3623 // in turn simplifies stack overflow handling.
3624 int PhaseOutput::bang_size_in_bytes() const {
3625 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3626 }
3627
3628 //------------------------------dump_asm---------------------------------------
3629 // Dump formatted assembly
3630 #if defined(SUPPORT_OPTO_ASSEMBLY)
3631 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3632
3633 int pc_digits = 3; // #chars required for pc
3634 int sb_chars = 3; // #chars for "start bundle" indicator
3635 int tab_size = 8;
3636 if (pcs != nullptr) {
3637 int max_pc = 0;
3638 for (uint i = 0; i < pc_limit; i++) {
3639 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3640 }
3641 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3642 }
3643 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3644
3645 bool cut_short = false;
3646 st->print_cr("#");
3647 st->print("# "); C->tf()->dump_on(st); st->cr();
3648 st->print_cr("#");
3649
3650 // For all blocks
3651 int pc = 0x0; // Program counter
3652 char starts_bundle = ' ';
3653 C->regalloc()->dump_frame();
3654
3655 Node *n = nullptr;
3656 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3657 if (VMThread::should_terminate()) {
3658 cut_short = true;
3659 break;
3660 }
3661 Block* block = C->cfg()->get_block(i);
3662 if (block->is_connector() && !Verbose) {
3663 continue;
3664 }
3665 n = block->head();
3666 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3667 pc = pcs[n->_idx];
3668 st->print("%*.*x", pc_digits, pc_digits, pc);
3669 }
3670 st->fill_to(prefix_len);
3671 block->dump_head(C->cfg(), st);
3672 if (block->is_connector()) {
3673 st->fill_to(prefix_len);
3674 st->print_cr("# Empty connector block");
3675 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3676 st->fill_to(prefix_len);
3677 st->print_cr("# Block is sole successor of call");
3678 }
3679
3680 // For all instructions
3681 Node *delay = nullptr;
3682 for (uint j = 0; j < block->number_of_nodes(); j++) {
3683 if (VMThread::should_terminate()) {
3684 cut_short = true;
3685 break;
3686 }
3687 n = block->get_node(j);
3688 if (valid_bundle_info(n)) {
3689 Bundle* bundle = node_bundling(n);
3690 if (bundle->used_in_unconditional_delay()) {
3691 delay = n;
3692 continue;
3693 }
3694 if (bundle->starts_bundle()) {
3695 starts_bundle = '+';
3696 }
3697 }
3698
3699 if (WizardMode) {
3700 n->dump();
3701 }
3702
3703 if( !n->is_Region() && // Dont print in the Assembly
3704 !n->is_Phi() && // a few noisely useless nodes
3705 !n->is_Proj() &&
3706 !n->is_MachTemp() &&
3707 !n->is_SafePointScalarObject() &&
3708 !n->is_Catch() && // Would be nice to print exception table targets
3709 !n->is_MergeMem() && // Not very interesting
3710 !n->is_top() && // Debug info table constants
3711 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3712 ) {
3713 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3714 pc = pcs[n->_idx];
3715 st->print("%*.*x", pc_digits, pc_digits, pc);
3716 } else {
3717 st->fill_to(pc_digits);
3718 }
3719 st->print(" %c ", starts_bundle);
3720 starts_bundle = ' ';
3721 st->fill_to(prefix_len);
3722 n->format(C->regalloc(), st);
3723 st->cr();
3724 }
3725
3726 // If we have an instruction with a delay slot, and have seen a delay,
3727 // then back up and print it
3728 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3729 // Coverity finding - Explicit null dereferenced.
3730 guarantee(delay != nullptr, "no unconditional delay instruction");
3731 if (WizardMode) delay->dump();
3732
3733 if (node_bundling(delay)->starts_bundle())
3734 starts_bundle = '+';
3735 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3736 pc = pcs[n->_idx];
3737 st->print("%*.*x", pc_digits, pc_digits, pc);
3738 } else {
3739 st->fill_to(pc_digits);
3740 }
3741 st->print(" %c ", starts_bundle);
3742 starts_bundle = ' ';
3743 st->fill_to(prefix_len);
3744 delay->format(C->regalloc(), st);
3745 st->cr();
3746 delay = nullptr;
3747 }
3748
3749 // Dump the exception table as well
3750 if( n->is_Catch() && (Verbose || WizardMode) ) {
3751 // Print the exception table for this offset
3752 _handler_table.print_subtable_for(pc);
3753 }
3754 st->bol(); // Make sure we start on a new line
3755 }
3756 st->cr(); // one empty line between blocks
3757 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3758 } // End of per-block dump
3759
3760 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3761 }
3762 #endif
3763
3764 #ifndef PRODUCT
3765 void PhaseOutput::print_statistics() {
3766 Scheduling::print_statistics();
3767 }
3768 #endif