1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/assembler.inline.hpp"
26 #include "code/aotCodeCache.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compilerDirectives.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "gc/shared/barrierSet.hpp"
35 #include "gc/shared/gc_globals.hpp"
36 #include "gc/shared/c2/barrierSetC2.hpp"
37 #include "memory/allocation.hpp"
38 #include "opto/ad.hpp"
39 #include "opto/block.hpp"
40 #include "opto/c2compiler.hpp"
41 #include "opto/c2_MacroAssembler.hpp"
42 #include "opto/callnode.hpp"
43 #include "opto/cfgnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/node.hpp"
47 #include "opto/optoreg.hpp"
48 #include "opto/output.hpp"
49 #include "opto/regalloc.hpp"
50 #include "opto/type.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/macros.hpp"
53 #include "utilities/powerOfTwo.hpp"
54 #include "utilities/xmlstream.hpp"
55
56 #ifndef PRODUCT
57 #define DEBUG_ARG(x) , x
58 #else
59 #define DEBUG_ARG(x)
60 #endif
61
62 //------------------------------Scheduling----------------------------------
63 // This class contains all the information necessary to implement instruction
64 // scheduling and bundling.
65 class Scheduling {
66
67 private:
68 // Arena to use
69 Arena *_arena;
70
71 // Control-Flow Graph info
72 PhaseCFG *_cfg;
73
74 // Register Allocation info
75 PhaseRegAlloc *_regalloc;
76
77 // Number of nodes in the method
78 uint _node_bundling_limit;
79
80 // List of scheduled nodes. Generated in reverse order
81 Node_List _scheduled;
82
83 // List of nodes currently available for choosing for scheduling
84 Node_List _available;
85
86 // For each instruction beginning a bundle, the number of following
87 // nodes to be bundled with it.
88 Bundle *_node_bundling_base;
89
90 // Mapping from register to Node
91 Node_List _reg_node;
92
93 // Free list for pinch nodes.
94 Node_List _pinch_free_list;
95
96 // Number of uses of this node within the containing basic block.
97 short *_uses;
98
99 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
100 // front, branch+proj at end. Also skips Catch/CProj (same as
101 // branch-at-end), plus just-prior exception-throwing call.
102 uint _bb_start, _bb_end;
103
104 // Latency from the end of the basic block as scheduled
105 unsigned short *_current_latency;
106
107 // Remember the next node
108 Node *_next_node;
109
110 // Use this for an unconditional branch delay slot
111 Node *_unconditional_delay_slot;
112
113 // Pointer to a Nop
114 MachNopNode *_nop;
115
116 // Length of the current bundle, in instructions
117 uint _bundle_instr_count;
118
119 // Current Cycle number, for computing latencies and bundling
120 uint _bundle_cycle_number;
121
122 // Bundle information
123 Pipeline_Use_Element _bundle_use_elements[resource_count];
124 Pipeline_Use _bundle_use;
125
126 // Dump the available list
127 void dump_available() const;
128
129 public:
130 Scheduling(Arena *arena, Compile &compile);
131
132 // Destructor
133 NOT_PRODUCT( ~Scheduling(); )
134
135 // Step ahead "i" cycles
136 void step(uint i);
137
138 // Step ahead 1 cycle, and clear the bundle state (for example,
139 // at a branch target)
140 void step_and_clear();
141
142 Bundle* node_bundling(const Node *n) {
143 assert(valid_bundle_info(n), "oob");
144 return (&_node_bundling_base[n->_idx]);
145 }
146
147 bool valid_bundle_info(const Node *n) const {
148 return (_node_bundling_limit > n->_idx);
149 }
150
151 bool starts_bundle(const Node *n) const {
152 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
153 }
154
155 // Do the scheduling
156 void DoScheduling();
157
158 // Compute the register antidependencies within a basic block
159 void ComputeRegisterAntidependencies(Block *bb);
160 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
161 void verify_good_schedule( Block *b, const char *msg );
162 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
163 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
164
165 // Add a node to the current bundle
166 void AddNodeToBundle(Node *n, const Block *bb);
167
168 // Return an integer less than, equal to, or greater than zero
169 // if the stack offset of the first argument is respectively
170 // less than, equal to, or greater than the second.
171 int compare_two_spill_nodes(Node* first, Node* second);
172
173 // Add a node to the list of available nodes
174 void AddNodeToAvailableList(Node *n);
175
176 // Compute the local use count for the nodes in a block, and compute
177 // the list of instructions with no uses in the block as available
178 void ComputeUseCount(const Block *bb);
179
180 // Choose an instruction from the available list to add to the bundle
181 Node * ChooseNodeToBundle();
182
183 // See if this Node fits into the currently accumulating bundle
184 bool NodeFitsInBundle(Node *n);
185
186 // Decrement the use count for a node
187 void DecrementUseCounts(Node *n, const Block *bb);
188
189 // Garbage collect pinch nodes for reuse by other blocks.
190 void garbage_collect_pinch_nodes();
191 // Clean up a pinch node for reuse (helper for above).
192 void cleanup_pinch( Node *pinch );
193
194 // Information for statistics gathering
195 #ifndef PRODUCT
196 private:
197 // Gather information on size of nops relative to total
198 uint _branches, _unconditional_delays;
199
200 static uint _total_nop_size, _total_method_size;
201 static uint _total_branches, _total_unconditional_delays;
202 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
203
204 public:
205 static void print_statistics();
206
207 static void increment_instructions_per_bundle(uint i) {
208 _total_instructions_per_bundle[i]++;
209 }
210
211 static void increment_nop_size(uint s) {
212 _total_nop_size += s;
213 }
214
215 static void increment_method_size(uint s) {
216 _total_method_size += s;
217 }
218 #endif
219
220 };
221
222 PhaseOutput::PhaseOutput()
223 : Phase(Phase::Output),
224 _code_buffer("Compile::Fill_buffer"),
225 _first_block_size(0),
226 _handler_table(),
227 _inc_table(),
228 _stub_list(),
229 _oop_map_set(nullptr),
230 _scratch_buffer_blob(nullptr),
231 _scratch_locs_memory(nullptr),
232 _scratch_const_size(-1),
233 _in_scratch_emit_size(false),
234 _frame_slots(0),
235 _code_offsets(),
236 _node_bundling_limit(0),
237 _node_bundling_base(nullptr),
238 _orig_pc_slot(0),
239 _orig_pc_slot_offset_in_bytes(0),
240 _buf_sizes(),
241 _block(nullptr),
242 _index(0) {
243 C->set_output(this);
244 if (C->stub_name() == nullptr) {
245 int fixed_slots = C->fixed_slots();
246 if (C->needs_stack_repair()) {
247 fixed_slots -= 2;
248 }
249 // TODO 8284443 Only reserve extra slot if needed
250 if (InlineTypeReturnedAsFields) {
251 fixed_slots -= 2;
252 }
253 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
254 }
255 }
256
257 PhaseOutput::~PhaseOutput() {
258 C->set_output(nullptr);
259 if (_scratch_buffer_blob != nullptr) {
260 BufferBlob::free(_scratch_buffer_blob);
261 }
262 }
263
264 void PhaseOutput::perform_mach_node_analysis() {
265 // Late barrier analysis must be done after schedule and bundle
266 // Otherwise liveness based spilling will fail
267 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
268 bs->late_barrier_analysis();
269
270 pd_perform_mach_node_analysis();
271
272 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
273 }
274
275 // Convert Nodes to instruction bits and pass off to the VM
276 void PhaseOutput::Output() {
277 // RootNode goes
278 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
279
280 // The number of new nodes (mostly MachNop) is proportional to
281 // the number of java calls and inner loops which are aligned.
282 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
283 C->inner_loops()*(OptoLoopAlignment-1)),
284 "out of nodes before code generation" ) ) {
285 return;
286 }
287 // Make sure I can find the Start Node
288 Block *entry = C->cfg()->get_block(1);
289 Block *broot = C->cfg()->get_root_block();
290
291 const StartNode *start = entry->head()->as_Start();
292
293 // Replace StartNode with prolog
294 Label verified_entry;
295 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
296 entry->map_node(prolog, 0);
297 C->cfg()->map_node_to_block(prolog, entry);
298 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
299
300 // Virtual methods need an unverified entry point
301 if (C->is_osr_compilation()) {
302 if (PoisonOSREntry) {
303 // TODO: Should use a ShouldNotReachHereNode...
304 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
305 }
306 } else {
307 if (C->method()) {
308 if (C->method()->has_scalarized_args()) {
309 // Add entry point to unpack all inline type arguments
310 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
311 if (!C->method()->is_static()) {
312 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
313 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
314 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
315 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
316 }
317 } else if (!C->method()->is_static()) {
318 // Insert unvalidated entry point
319 C->cfg()->insert(broot, 0, new MachUEPNode());
320 }
321 }
322 }
323
324 // Break before main entry point
325 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
326 (OptoBreakpoint && C->is_method_compilation()) ||
327 (OptoBreakpointOSR && C->is_osr_compilation()) ||
328 (OptoBreakpointC2R && !C->method()) ) {
329 // checking for C->method() means that OptoBreakpoint does not apply to
330 // runtime stubs or frame converters
331 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
332 }
333
334 // Insert epilogs before every return
335 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
336 Block* block = C->cfg()->get_block(i);
337 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
338 Node* m = block->end();
339 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
340 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
341 block->add_inst(epilog);
342 C->cfg()->map_node_to_block(epilog, block);
343 }
344 }
345 }
346
347 // Keeper of sizing aspects
348 _buf_sizes = BufferSizingData();
349
350 // Initialize code buffer
351 estimate_buffer_size(_buf_sizes._const);
352 if (C->failing()) return;
353
354 // Pre-compute the length of blocks and replace
355 // long branches with short if machine supports it.
356 // Must be done before ScheduleAndBundle due to SPARC delay slots
357 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
358 blk_starts[0] = 0;
359 shorten_branches(blk_starts);
360
361 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
362 // Compute the offsets of the entry points required by the inline type calling convention
363 if (!C->method()->is_static()) {
364 // We have entries at the beginning of the method, implemented by the first 4 nodes.
365 // Entry (unverified) @ offset 0
366 // Verified_Inline_Entry_RO
367 // Inline_Entry (unverified)
368 // Verified_Inline_Entry
369 uint offset = 0;
370 _code_offsets.set_value(CodeOffsets::Entry, offset);
371
372 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
373 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
374
375 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
376 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
377
378 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
379 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
380 } else {
381 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
382 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
383 }
384 }
385
386 ScheduleAndBundle();
387 if (C->failing()) {
388 return;
389 }
390
391 perform_mach_node_analysis();
392
393 // Complete sizing of codebuffer
394 CodeBuffer* cb = init_buffer();
395 if (cb == nullptr || C->failing()) {
396 return;
397 }
398
399 BuildOopMaps();
400
401 if (C->failing()) {
402 return;
403 }
404
405 C2_MacroAssembler masm(cb);
406 fill_buffer(&masm, blk_starts);
407 }
408
409 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
410 // Determine if we need to generate a stack overflow check.
411 // Do it if the method is not a stub function and
412 // has java calls or has frame size > vm_page_size/8.
413 // The debug VM checks that deoptimization doesn't trigger an
414 // unexpected stack overflow (compiled method stack banging should
415 // guarantee it doesn't happen) so we always need the stack bang in
416 // a debug VM.
417 return (C->stub_function() == nullptr &&
418 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
419 DEBUG_ONLY(|| true)));
420 }
421
422 bool PhaseOutput::need_register_stack_bang() const {
423 // Determine if we need to generate a register stack overflow check.
424 // This is only used on architectures which have split register
425 // and memory stacks.
426 // Bang if the method is not a stub function and has java calls
427 return (C->stub_function() == nullptr && C->has_java_calls());
428 }
429
430
431 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
432 // of a loop. When aligning a loop we need to provide enough instructions
433 // in cpu's fetch buffer to feed decoders. The loop alignment could be
434 // avoided if we have enough instructions in fetch buffer at the head of a loop.
435 // By default, the size is set to 999999 by Block's constructor so that
436 // a loop will be aligned if the size is not reset here.
437 //
438 // Note: Mach instructions could contain several HW instructions
439 // so the size is estimated only.
440 //
441 void PhaseOutput::compute_loop_first_inst_sizes() {
442 // The next condition is used to gate the loop alignment optimization.
443 // Don't aligned a loop if there are enough instructions at the head of a loop
444 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
445 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
446 // equal to 11 bytes which is the largest address NOP instruction.
447 if (MaxLoopPad < OptoLoopAlignment - 1) {
448 uint last_block = C->cfg()->number_of_blocks() - 1;
449 for (uint i = 1; i <= last_block; i++) {
450 Block* block = C->cfg()->get_block(i);
451 // Check the first loop's block which requires an alignment.
452 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
453 uint sum_size = 0;
454 uint inst_cnt = NumberOfLoopInstrToAlign;
455 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
456
457 // Check subsequent fallthrough blocks if the loop's first
458 // block(s) does not have enough instructions.
459 Block *nb = block;
460 while(inst_cnt > 0 &&
461 i < last_block &&
462 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
463 !nb->has_successor(block)) {
464 i++;
465 nb = C->cfg()->get_block(i);
466 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
467 } // while( inst_cnt > 0 && i < last_block )
468
469 block->set_first_inst_size(sum_size);
470 } // f( b->head()->is_Loop() )
471 } // for( i <= last_block )
472 } // if( MaxLoopPad < OptoLoopAlignment-1 )
473 }
474
475 // The architecture description provides short branch variants for some long
476 // branch instructions. Replace eligible long branches with short branches.
477 void PhaseOutput::shorten_branches(uint* blk_starts) {
478
479 Compile::TracePhase tp(_t_shortenBranches);
480
481 // Compute size of each block, method size, and relocation information size
482 uint nblocks = C->cfg()->number_of_blocks();
483
484 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
485 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
486 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
487
488 // Collect worst case block paddings
489 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
490 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
491
492 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
493 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
494
495 bool has_short_branch_candidate = false;
496
497 // Initialize the sizes to 0
498 int code_size = 0; // Size in bytes of generated code
499 int stub_size = 0; // Size in bytes of all stub entries
500 // Size in bytes of all relocation entries, including those in local stubs.
501 // Start with 2-bytes of reloc info for the unvalidated entry point
502 int reloc_size = 1; // Number of relocation entries
503
504 // Make three passes. The first computes pessimistic blk_starts,
505 // relative jmp_offset and reloc_size information. The second performs
506 // short branch substitution using the pessimistic sizing. The
507 // third inserts nops where needed.
508
509 // Step one, perform a pessimistic sizing pass.
510 uint last_call_adr = max_juint;
511 uint last_avoid_back_to_back_adr = max_juint;
512 uint nop_size = (new MachNopNode())->size(C->regalloc());
513 for (uint i = 0; i < nblocks; i++) { // For all blocks
514 Block* block = C->cfg()->get_block(i);
515 _block = block;
516
517 // During short branch replacement, we store the relative (to blk_starts)
518 // offset of jump in jmp_offset, rather than the absolute offset of jump.
519 // This is so that we do not need to recompute sizes of all nodes when
520 // we compute correct blk_starts in our next sizing pass.
521 jmp_offset[i] = 0;
522 jmp_size[i] = 0;
523 jmp_nidx[i] = -1;
524 DEBUG_ONLY( jmp_target[i] = 0; )
525 DEBUG_ONLY( jmp_rule[i] = 0; )
526
527 // Sum all instruction sizes to compute block size
528 uint last_inst = block->number_of_nodes();
529 uint blk_size = 0;
530 for (uint j = 0; j < last_inst; j++) {
531 _index = j;
532 Node* nj = block->get_node(_index);
533 // Handle machine instruction nodes
534 if (nj->is_Mach()) {
535 MachNode* mach = nj->as_Mach();
536 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
537 reloc_size += mach->reloc();
538 if (mach->is_MachCall()) {
539 // add size information for trampoline stub
540 // class CallStubImpl is platform-specific and defined in the *.ad files.
541 stub_size += CallStubImpl::size_call_trampoline();
542 reloc_size += CallStubImpl::reloc_call_trampoline();
543
544 MachCallNode *mcall = mach->as_MachCall();
545 // This destination address is NOT PC-relative
546
547 if (mcall->entry_point() != nullptr) {
548 mcall->method_set((intptr_t)mcall->entry_point());
549 }
550
551 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
552 stub_size += CompiledDirectCall::to_interp_stub_size();
553 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
554 }
555 } else if (mach->is_MachSafePoint()) {
556 // If call/safepoint are adjacent, account for possible
557 // nop to disambiguate the two safepoints.
558 // ScheduleAndBundle() can rearrange nodes in a block,
559 // check for all offsets inside this block.
560 if (last_call_adr >= blk_starts[i]) {
561 blk_size += nop_size;
562 }
563 }
564 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
565 // Nop is inserted between "avoid back to back" instructions.
566 // ScheduleAndBundle() can rearrange nodes in a block,
567 // check for all offsets inside this block.
568 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
569 blk_size += nop_size;
570 }
571 }
572 if (mach->may_be_short_branch()) {
573 if (!nj->is_MachBranch()) {
574 #ifndef PRODUCT
575 nj->dump(3);
576 #endif
577 Unimplemented();
578 }
579 assert(jmp_nidx[i] == -1, "block should have only one branch");
580 jmp_offset[i] = blk_size;
581 jmp_size[i] = nj->size(C->regalloc());
582 jmp_nidx[i] = j;
583 has_short_branch_candidate = true;
584 }
585 }
586 blk_size += nj->size(C->regalloc());
587 // Remember end of call offset
588 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
589 last_call_adr = blk_starts[i]+blk_size;
590 }
591 // Remember end of avoid_back_to_back offset
592 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
593 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
594 }
595 }
596
597 // When the next block starts a loop, we may insert pad NOP
598 // instructions. Since we cannot know our future alignment,
599 // assume the worst.
600 if (i < nblocks - 1) {
601 Block* nb = C->cfg()->get_block(i + 1);
602 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
603 if (max_loop_pad > 0) {
604 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
605 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
606 // If either is the last instruction in this block, bump by
607 // max_loop_pad in lock-step with blk_size, so sizing
608 // calculations in subsequent blocks still can conservatively
609 // detect that it may the last instruction in this block.
610 if (last_call_adr == blk_starts[i]+blk_size) {
611 last_call_adr += max_loop_pad;
612 }
613 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
614 last_avoid_back_to_back_adr += max_loop_pad;
615 }
616 blk_size += max_loop_pad;
617 block_worst_case_pad[i + 1] = max_loop_pad;
618 }
619 }
620
621 // Save block size; update total method size
622 blk_starts[i+1] = blk_starts[i]+blk_size;
623 }
624
625 // Step two, replace eligible long jumps.
626 bool progress = true;
627 uint last_may_be_short_branch_adr = max_juint;
628 while (has_short_branch_candidate && progress) {
629 progress = false;
630 has_short_branch_candidate = false;
631 int adjust_block_start = 0;
632 for (uint i = 0; i < nblocks; i++) {
633 Block* block = C->cfg()->get_block(i);
634 int idx = jmp_nidx[i];
635 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
636 if (mach != nullptr && mach->may_be_short_branch()) {
637 #ifdef ASSERT
638 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
639 int j;
640 // Find the branch; ignore trailing NOPs.
641 for (j = block->number_of_nodes()-1; j>=0; j--) {
642 Node* n = block->get_node(j);
643 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
644 break;
645 }
646 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
647 #endif
648 int br_size = jmp_size[i];
649 int br_offs = blk_starts[i] + jmp_offset[i];
650
651 // This requires the TRUE branch target be in succs[0]
652 uint bnum = block->non_connector_successor(0)->_pre_order;
653 int offset = blk_starts[bnum] - br_offs;
654 if (bnum > i) { // adjust following block's offset
655 offset -= adjust_block_start;
656 }
657
658 // This block can be a loop header, account for the padding
659 // in the previous block.
660 int block_padding = block_worst_case_pad[i];
661 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
662 // In the following code a nop could be inserted before
663 // the branch which will increase the backward distance.
664 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
665 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
666
667 if (needs_padding && offset <= 0)
668 offset -= nop_size;
669
670 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
671 // We've got a winner. Replace this branch.
672 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
673
674 // Update the jmp_size.
675 int new_size = replacement->size(C->regalloc());
676 int diff = br_size - new_size;
677 assert(diff >= (int)nop_size, "short_branch size should be smaller");
678 // Conservatively take into account padding between
679 // avoid_back_to_back branches. Previous branch could be
680 // converted into avoid_back_to_back branch during next
681 // rounds.
682 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
683 jmp_offset[i] += nop_size;
684 diff -= nop_size;
685 }
686 adjust_block_start += diff;
687 block->map_node(replacement, idx);
688 mach->subsume_by(replacement, C);
689 mach = replacement;
690 progress = true;
691
692 jmp_size[i] = new_size;
693 DEBUG_ONLY( jmp_target[i] = bnum; );
694 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
695 } else {
696 // The jump distance is not short, try again during next iteration.
697 has_short_branch_candidate = true;
698 }
699 } // (mach->may_be_short_branch())
700 if (mach != nullptr && (mach->may_be_short_branch() ||
701 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
702 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
703 }
704 blk_starts[i+1] -= adjust_block_start;
705 }
706 }
707
708 #ifdef ASSERT
709 for (uint i = 0; i < nblocks; i++) { // For all blocks
710 if (jmp_target[i] != 0) {
711 int br_size = jmp_size[i];
712 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
713 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
714 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]);
715 }
716 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
717 }
718 }
719 #endif
720
721 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
722 // after ScheduleAndBundle().
723
724 // ------------------
725 // Compute size for code buffer
726 code_size = blk_starts[nblocks];
727
728 // Relocation records
729 reloc_size += 1; // Relo entry for exception handler
730
731 // Adjust reloc_size to number of record of relocation info
732 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
733 // a relocation index.
734 // The CodeBuffer will expand the locs array if this estimate is too low.
735 reloc_size *= 10 / sizeof(relocInfo);
736
737 _buf_sizes._reloc = reloc_size;
738 _buf_sizes._code = code_size;
739 _buf_sizes._stub = stub_size;
740 }
741
742 //------------------------------FillLocArray-----------------------------------
743 // Create a bit of debug info and append it to the array. The mapping is from
744 // Java local or expression stack to constant, register or stack-slot. For
745 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
746 // entry has been taken care of and caller should skip it).
747 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
748 // This should never have accepted Bad before
749 assert(OptoReg::is_valid(regnum), "location must be valid");
750 return (OptoReg::is_reg(regnum))
751 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
752 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
753 }
754
755
756 ObjectValue*
757 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
758 for (int i = 0; i < objs->length(); i++) {
759 assert(objs->at(i)->is_object(), "corrupt object cache");
760 ObjectValue* sv = objs->at(i)->as_ObjectValue();
761 if (sv->id() == id) {
762 return sv;
763 }
764 }
765 // Otherwise..
766 return nullptr;
767 }
768
769 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
770 ObjectValue* sv ) {
771 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
772 objs->append(sv);
773 }
774
775
776 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
777 GrowableArray<ScopeValue*> *array,
778 GrowableArray<ScopeValue*> *objs ) {
779 assert( local, "use _top instead of null" );
780 if (array->length() != idx) {
781 assert(array->length() == idx + 1, "Unexpected array count");
782 // Old functionality:
783 // return
784 // New functionality:
785 // Assert if the local is not top. In product mode let the new node
786 // override the old entry.
787 assert(local == C->top(), "LocArray collision");
788 if (local == C->top()) {
789 return;
790 }
791 array->pop();
792 }
793 const Type *t = local->bottom_type();
794
795 // Is it a safepoint scalar object node?
796 if (local->is_SafePointScalarObject()) {
797 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
798
799 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
800 if (sv == nullptr) {
801 ciKlass* cik = t->is_oopptr()->exact_klass();
802 assert(cik->is_instance_klass() ||
803 cik->is_array_klass(), "Not supported allocation.");
804 uint first_ind = spobj->first_index(sfpt->jvms());
805 // Nullable, scalarized inline types have a null_marker input
806 // that needs to be checked before using the field values.
807 ScopeValue* null_marker = nullptr;
808 if (cik->is_inlinetype()) {
809 Node* null_marker_node = sfpt->in(first_ind++);
810 assert(null_marker_node != nullptr, "null_marker node not found");
811 if (!null_marker_node->is_top()) {
812 const TypeInt* null_marker_type = null_marker_node->bottom_type()->is_int();
813 if (null_marker_node->is_Con()) {
814 null_marker = new ConstantIntValue(null_marker_type->get_con());
815 } else {
816 OptoReg::Name null_marker_reg = C->regalloc()->get_reg_first(null_marker_node);
817 null_marker = new_loc_value(C->regalloc(), null_marker_reg, Location::normal);
818 }
819 }
820 }
821 sv = new ObjectValue(spobj->_idx,
822 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, null_marker);
823 set_sv_for_object_node(objs, sv);
824
825 for (uint i = 0; i < spobj->n_fields(); i++) {
826 Node* fld_node = sfpt->in(first_ind+i);
827 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
828 }
829 }
830 array->append(sv);
831 return;
832 } else if (local->is_SafePointScalarMerge()) {
833 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
834 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
835
836 if (mv == nullptr) {
837 GrowableArray<ScopeValue*> deps;
838
839 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
840 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
841 assert(deps.length() == 1, "missing value");
842
843 int selector_idx = smerge->selector_idx(sfpt->jvms());
844 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
845 assert(deps.length() == 2, "missing value");
846
847 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
848 set_sv_for_object_node(objs, mv);
849
850 for (uint i = 1; i < smerge->req(); i++) {
851 Node* obj_node = smerge->in(i);
852 int idx = mv->possible_objects()->length();
853 (void)FillLocArray(idx, sfpt, obj_node, mv->possible_objects(), objs);
854
855 // By default ObjectValues that are in 'possible_objects' are not root objects.
856 // They will be marked as root later if they are directly referenced in a JVMS.
857 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
858 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
859 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
860 }
861 }
862 array->append(mv);
863 return;
864 }
865
866 // Grab the register number for the local
867 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
868 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
869 // Record the double as two float registers.
870 // The register mask for such a value always specifies two adjacent
871 // float registers, with the lower register number even.
872 // Normally, the allocation of high and low words to these registers
873 // is irrelevant, because nearly all operations on register pairs
874 // (e.g., StoreD) treat them as a single unit.
875 // Here, we assume in addition that the words in these two registers
876 // stored "naturally" (by operations like StoreD and double stores
877 // within the interpreter) such that the lower-numbered register
878 // is written to the lower memory address. This may seem like
879 // a machine dependency, but it is not--it is a requirement on
880 // the author of the <arch>.ad file to ensure that, for every
881 // even/odd double-register pair to which a double may be allocated,
882 // the word in the even single-register is stored to the first
883 // memory word. (Note that register numbers are completely
884 // arbitrary, and are not tied to any machine-level encodings.)
885 #ifdef _LP64
886 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
887 array->append(new ConstantIntValue((jint)0));
888 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
889 } else if ( t->base() == Type::Long ) {
890 array->append(new ConstantIntValue((jint)0));
891 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
892 } else if ( t->base() == Type::RawPtr ) {
893 // jsr/ret return address which must be restored into the full
894 // width 64-bit stack slot.
895 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
896 }
897 #else //_LP64
898 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
899 // Repack the double/long as two jints.
900 // The convention the interpreter uses is that the second local
901 // holds the first raw word of the native double representation.
902 // This is actually reasonable, since locals and stack arrays
903 // grow downwards in all implementations.
904 // (If, on some machine, the interpreter's Java locals or stack
905 // were to grow upwards, the embedded doubles would be word-swapped.)
906 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
907 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
908 }
909 #endif //_LP64
910 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
911 OptoReg::is_reg(regnum) ) {
912 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
913 ? Location::float_in_dbl : Location::normal ));
914 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
915 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
916 ? Location::int_in_long : Location::normal ));
917 } else if( t->base() == Type::NarrowOop ) {
918 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
919 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
920 t->base() == Type::VectorD || t->base() == Type::VectorX ||
921 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
922 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
923 } else if (C->regalloc()->is_oop(local)) {
924 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
925 t->base() == Type::AryPtr,
926 "Unexpected type: %s", t->msg());
927 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
928 } else {
929 assert(t->base() == Type::Int || t->base() == Type::Half ||
930 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
931 "Unexpected type: %s", t->msg());
932 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
933 }
934 return;
935 }
936
937 // No register. It must be constant data.
938 switch (t->base()) {
939 case Type::Half: // Second half of a double
940 ShouldNotReachHere(); // Caller should skip 2nd halves
941 break;
942 case Type::AnyPtr:
943 array->append(new ConstantOopWriteValue(nullptr));
944 break;
945 case Type::AryPtr:
946 case Type::InstPtr: // fall through
947 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
948 break;
949 case Type::NarrowOop:
950 if (t == TypeNarrowOop::NULL_PTR) {
951 array->append(new ConstantOopWriteValue(nullptr));
952 } else {
953 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
954 }
955 break;
956 case Type::Int:
957 array->append(new ConstantIntValue(t->is_int()->get_con()));
958 break;
959 case Type::RawPtr:
960 // A return address (T_ADDRESS).
961 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
962 #ifdef _LP64
963 // Must be restored to the full-width 64-bit stack slot.
964 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
965 #else
966 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
967 #endif
968 break;
969 case Type::FloatCon: {
970 float f = t->is_float_constant()->getf();
971 array->append(new ConstantIntValue(jint_cast(f)));
972 break;
973 }
974 case Type::DoubleCon: {
975 jdouble d = t->is_double_constant()->getd();
976 #ifdef _LP64
977 array->append(new ConstantIntValue((jint)0));
978 array->append(new ConstantDoubleValue(d));
979 #else
980 // Repack the double as two jints.
981 // The convention the interpreter uses is that the second local
982 // holds the first raw word of the native double representation.
983 // This is actually reasonable, since locals and stack arrays
984 // grow downwards in all implementations.
985 // (If, on some machine, the interpreter's Java locals or stack
986 // were to grow upwards, the embedded doubles would be word-swapped.)
987 jlong_accessor acc;
988 acc.long_value = jlong_cast(d);
989 array->append(new ConstantIntValue(acc.words[1]));
990 array->append(new ConstantIntValue(acc.words[0]));
991 #endif
992 break;
993 }
994 case Type::Long: {
995 jlong d = t->is_long()->get_con();
996 #ifdef _LP64
997 array->append(new ConstantIntValue((jint)0));
998 array->append(new ConstantLongValue(d));
999 #else
1000 // Repack the long as two jints.
1001 // The convention the interpreter uses is that the second local
1002 // holds the first raw word of the native double representation.
1003 // This is actually reasonable, since locals and stack arrays
1004 // grow downwards in all implementations.
1005 // (If, on some machine, the interpreter's Java locals or stack
1006 // were to grow upwards, the embedded doubles would be word-swapped.)
1007 jlong_accessor acc;
1008 acc.long_value = d;
1009 array->append(new ConstantIntValue(acc.words[1]));
1010 array->append(new ConstantIntValue(acc.words[0]));
1011 #endif
1012 break;
1013 }
1014 case Type::Top: // Add an illegal value here
1015 array->append(new LocationValue(Location()));
1016 break;
1017 default:
1018 ShouldNotReachHere();
1019 break;
1020 }
1021 }
1022
1023 // Determine if this node starts a bundle
1024 bool PhaseOutput::starts_bundle(const Node *n) const {
1025 return (_node_bundling_limit > n->_idx &&
1026 _node_bundling_base[n->_idx].starts_bundle());
1027 }
1028
1029 // Determine if there is a monitor that has 'ov' as its owner.
1030 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
1031 for (int k = 0; k < monarray->length(); k++) {
1032 MonitorValue* mv = monarray->at(k);
1033 if (mv->owner() == ov) {
1034 return true;
1035 }
1036 }
1037
1038 return false;
1039 }
1040
1041 // Determine if there is a scalar replaced object description represented by 'ov'.
1042 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
1043 GrowableArray<ScopeValue*>* objs,
1044 ObjectValue* ov) const {
1045 for (int i = 0; i < jvms->scl_size(); i++) {
1046 Node* n = sfn->scalarized_obj(jvms, i);
1047 // Other kinds of nodes that we may encounter here, for instance constants
1048 // representing values of fields of objects scalarized, aren't relevant for
1049 // us, since they don't map to ObjectValue.
1050 if (!n->is_SafePointScalarObject()) {
1051 continue;
1052 }
1053
1054 ObjectValue* other = sv_for_node_id(objs, n->_idx);
1055 if (ov == other) {
1056 return true;
1057 }
1058 }
1059 return false;
1060 }
1061
1062 //--------------------------Process_OopMap_Node--------------------------------
1063 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1064 // Handle special safepoint nodes for synchronization
1065 MachSafePointNode *sfn = mach->as_MachSafePoint();
1066 MachCallNode *mcall;
1067
1068 int safepoint_pc_offset = current_offset;
1069 bool is_method_handle_invoke = false;
1070 bool return_oop = false;
1071 bool return_scalarized = false;
1072 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1073 bool arg_escape = false;
1074
1075 // Add the safepoint in the DebugInfoRecorder
1076 if( !mach->is_MachCall() ) {
1077 mcall = nullptr;
1078 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1079 } else {
1080 mcall = mach->as_MachCall();
1081
1082 // Is the call a MethodHandle call?
1083 if (mcall->is_MachCallJava()) {
1084 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1085 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1086 is_method_handle_invoke = true;
1087 }
1088 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1089 }
1090
1091 // Check if a call returns an object.
1092 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1093 return_oop = true;
1094 }
1095 if (mcall->returns_scalarized()) {
1096 return_scalarized = true;
1097 }
1098 safepoint_pc_offset += mcall->ret_addr_offset();
1099 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1100 }
1101
1102 // Loop over the JVMState list to add scope information
1103 // Do not skip safepoints with a null method, they need monitor info
1104 JVMState* youngest_jvms = sfn->jvms();
1105 int max_depth = youngest_jvms->depth();
1106
1107 // Allocate the object pool for scalar-replaced objects -- the map from
1108 // small-integer keys (which can be recorded in the local and ostack
1109 // arrays) to descriptions of the object state.
1110 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1111
1112 // Visit scopes from oldest to youngest.
1113 for (int depth = 1; depth <= max_depth; depth++) {
1114 JVMState* jvms = youngest_jvms->of_depth(depth);
1115 int idx;
1116 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1117 // Safepoints that do not have method() set only provide oop-map and monitor info
1118 // to support GC; these do not support deoptimization.
1119 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1120 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1121 int num_mon = jvms->nof_monitors();
1122 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1123 "JVMS local count must match that of the method");
1124
1125 // Add Local and Expression Stack Information
1126
1127 // Insert locals into the locarray
1128 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1129 for( idx = 0; idx < num_locs; idx++ ) {
1130 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1131 }
1132
1133 // Insert expression stack entries into the exparray
1134 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1135 for( idx = 0; idx < num_exps; idx++ ) {
1136 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1137 }
1138
1139 // Add in mappings of the monitors
1140 assert( !method ||
1141 !method->is_synchronized() ||
1142 method->is_native() ||
1143 num_mon > 0 ||
1144 !GenerateSynchronizationCode,
1145 "monitors must always exist for synchronized methods");
1146
1147 // Build the growable array of ScopeValues for exp stack
1148 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1149
1150 // Loop over monitors and insert into array
1151 for (idx = 0; idx < num_mon; idx++) {
1152 // Grab the node that defines this monitor
1153 Node* box_node = sfn->monitor_box(jvms, idx);
1154 Node* obj_node = sfn->monitor_obj(jvms, idx);
1155
1156 // Create ScopeValue for object
1157 ScopeValue *scval = nullptr;
1158
1159 if (obj_node->is_SafePointScalarObject()) {
1160 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1161 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1162 if (scval == nullptr) {
1163 const Type *t = spobj->bottom_type();
1164 ciKlass* cik = t->is_oopptr()->exact_klass();
1165 assert(cik->is_instance_klass() ||
1166 cik->is_array_klass(), "Not supported allocation.");
1167 ObjectValue* sv = new ObjectValue(spobj->_idx,
1168 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1169 PhaseOutput::set_sv_for_object_node(objs, sv);
1170
1171 uint first_ind = spobj->first_index(youngest_jvms);
1172 for (uint i = 0; i < spobj->n_fields(); i++) {
1173 Node* fld_node = sfn->in(first_ind+i);
1174 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1175 }
1176 scval = sv;
1177 }
1178 } else if (obj_node->is_SafePointScalarMerge()) {
1179 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1180 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1181
1182 if (mv == nullptr) {
1183 GrowableArray<ScopeValue*> deps;
1184
1185 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1186 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1187 assert(deps.length() == 1, "missing value");
1188
1189 int selector_idx = smerge->selector_idx(youngest_jvms);
1190 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1191 assert(deps.length() == 2, "missing value");
1192
1193 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1194 set_sv_for_object_node(objs, mv);
1195
1196 for (uint i = 1; i < smerge->req(); i++) {
1197 Node* obj_node = smerge->in(i);
1198 int idx = mv->possible_objects()->length();
1199 (void)FillLocArray(idx, sfn, obj_node, mv->possible_objects(), objs);
1200
1201 // By default ObjectValues that are in 'possible_objects' are not root objects.
1202 // They will be marked as root later if they are directly referenced in a JVMS.
1203 assert(mv->possible_objects()->length() > idx, "Didn't add entry to possible_objects?!");
1204 assert(mv->possible_objects()->at(idx)->is_object(), "Entries in possible_objects should be ObjectValue.");
1205 mv->possible_objects()->at(idx)->as_ObjectValue()->set_root(false);
1206 }
1207 }
1208 scval = mv;
1209 } else if (!obj_node->is_Con()) {
1210 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1211 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1212 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1213 } else {
1214 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1215 }
1216 } else {
1217 const TypePtr *tp = obj_node->get_ptr_type();
1218 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1219 }
1220
1221 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1222 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1223 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1224 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1225 }
1226
1227 // Mark ObjectValue nodes as root nodes if they are directly
1228 // referenced in the JVMS.
1229 for (int i = 0; i < objs->length(); i++) {
1230 ScopeValue* sv = objs->at(i);
1231 if (sv->is_object_merge()) {
1232 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1233
1234 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1235 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1236 if (ov->is_root()) {
1237 // Already flagged as 'root' by something else. We shouldn't change it
1238 // to non-root in a younger JVMS because it may need to be alive in
1239 // a younger JVMS.
1240 } else {
1241 bool is_root = locarray->contains(ov) ||
1242 exparray->contains(ov) ||
1243 contains_as_owner(monarray, ov) ||
1244 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1245 ov->set_root(is_root);
1246 }
1247 }
1248 }
1249 }
1250
1251 // We dump the object pool first, since deoptimization reads it in first.
1252 C->debug_info()->dump_object_pool(objs);
1253
1254 // Build first class objects to pass to scope
1255 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1256 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1257 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1258
1259 // Make method available for all Safepoints
1260 ciMethod* scope_method = method ? method : C->method();
1261 // Describe the scope here
1262 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1263 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1264 // Now we can describe the scope.
1265 methodHandle null_mh;
1266 bool rethrow_exception = false;
1267 C->debug_info()->describe_scope(
1268 safepoint_pc_offset,
1269 null_mh,
1270 scope_method,
1271 jvms->bci(),
1272 jvms->should_reexecute(),
1273 rethrow_exception,
1274 is_method_handle_invoke,
1275 return_oop,
1276 return_scalarized,
1277 has_ea_local_in_scope,
1278 arg_escape,
1279 locvals,
1280 expvals,
1281 monvals
1282 );
1283 } // End jvms loop
1284
1285 // Mark the end of the scope set.
1286 C->debug_info()->end_safepoint(safepoint_pc_offset);
1287 }
1288
1289
1290
1291 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1292 class NonSafepointEmitter {
1293 Compile* C;
1294 JVMState* _pending_jvms;
1295 int _pending_offset;
1296
1297 void emit_non_safepoint();
1298
1299 public:
1300 NonSafepointEmitter(Compile* compile) {
1301 this->C = compile;
1302 _pending_jvms = nullptr;
1303 _pending_offset = 0;
1304 }
1305
1306 void observe_instruction(Node* n, int pc_offset) {
1307 if (!C->debug_info()->recording_non_safepoints()) return;
1308
1309 Node_Notes* nn = C->node_notes_at(n->_idx);
1310 if (nn == nullptr || nn->jvms() == nullptr) return;
1311 if (_pending_jvms != nullptr &&
1312 _pending_jvms->same_calls_as(nn->jvms())) {
1313 // Repeated JVMS? Stretch it up here.
1314 _pending_offset = pc_offset;
1315 } else {
1316 if (_pending_jvms != nullptr &&
1317 _pending_offset < pc_offset) {
1318 emit_non_safepoint();
1319 }
1320 _pending_jvms = nullptr;
1321 if (pc_offset > C->debug_info()->last_pc_offset()) {
1322 // This is the only way _pending_jvms can become non-null:
1323 _pending_jvms = nn->jvms();
1324 _pending_offset = pc_offset;
1325 }
1326 }
1327 }
1328
1329 // Stay out of the way of real safepoints:
1330 void observe_safepoint(JVMState* jvms, int pc_offset) {
1331 if (_pending_jvms != nullptr &&
1332 !_pending_jvms->same_calls_as(jvms) &&
1333 _pending_offset < pc_offset) {
1334 emit_non_safepoint();
1335 }
1336 _pending_jvms = nullptr;
1337 }
1338
1339 void flush_at_end() {
1340 if (_pending_jvms != nullptr) {
1341 emit_non_safepoint();
1342 }
1343 _pending_jvms = nullptr;
1344 }
1345 };
1346
1347 void NonSafepointEmitter::emit_non_safepoint() {
1348 JVMState* youngest_jvms = _pending_jvms;
1349 int pc_offset = _pending_offset;
1350
1351 // Clear it now:
1352 _pending_jvms = nullptr;
1353
1354 DebugInformationRecorder* debug_info = C->debug_info();
1355 assert(debug_info->recording_non_safepoints(), "sanity");
1356
1357 debug_info->add_non_safepoint(pc_offset);
1358 int max_depth = youngest_jvms->depth();
1359
1360 // Visit scopes from oldest to youngest.
1361 for (int depth = 1; depth <= max_depth; depth++) {
1362 JVMState* jvms = youngest_jvms->of_depth(depth);
1363 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1364 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1365 methodHandle null_mh;
1366 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1367 }
1368
1369 // Mark the end of the scope set.
1370 debug_info->end_non_safepoint(pc_offset);
1371 }
1372
1373 //------------------------------init_buffer------------------------------------
1374 void PhaseOutput::estimate_buffer_size(int& const_req) {
1375
1376 // Set the initially allocated size
1377 const_req = initial_const_capacity;
1378
1379 // The extra spacing after the code is necessary on some platforms.
1380 // Sometimes we need to patch in a jump after the last instruction,
1381 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1382
1383 // Compute the byte offset where we can store the deopt pc.
1384 if (C->fixed_slots() != 0) {
1385 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1386 }
1387
1388 // Compute prolog code size
1389 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1390 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1391
1392 if (C->has_mach_constant_base_node()) {
1393 uint add_size = 0;
1394 // Fill the constant table.
1395 // Note: This must happen before shorten_branches.
1396 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1397 Block* b = C->cfg()->get_block(i);
1398
1399 for (uint j = 0; j < b->number_of_nodes(); j++) {
1400 Node* n = b->get_node(j);
1401
1402 // If the node is a MachConstantNode evaluate the constant
1403 // value section.
1404 if (n->is_MachConstant()) {
1405 MachConstantNode* machcon = n->as_MachConstant();
1406 machcon->eval_constant(C);
1407 } else if (n->is_Mach()) {
1408 // On Power there are more nodes that issue constants.
1409 add_size += (n->as_Mach()->ins_num_consts() * 8);
1410 }
1411 }
1412 }
1413
1414 // Calculate the offsets of the constants and the size of the
1415 // constant table (including the padding to the next section).
1416 constant_table().calculate_offsets_and_size();
1417 const_req = constant_table().alignment() + constant_table().size() + add_size;
1418 }
1419
1420 // Initialize the space for the BufferBlob used to find and verify
1421 // instruction size in MachNode::emit_size()
1422 init_scratch_buffer_blob(const_req);
1423 }
1424
1425 CodeBuffer* PhaseOutput::init_buffer() {
1426 int stub_req = _buf_sizes._stub;
1427 int code_req = _buf_sizes._code;
1428 int const_req = _buf_sizes._const;
1429
1430 int pad_req = NativeCall::byte_size();
1431
1432 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1433 stub_req += bs->estimate_stub_size();
1434
1435 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1436 // class HandlerImpl is platform-specific and defined in the *.ad files.
1437 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1438 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1439 stub_req += MAX_stubs_size; // ensure per-stub margin
1440 code_req += MAX_inst_size; // ensure per-instruction margin
1441
1442 if (StressCodeBuffers)
1443 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1444
1445 int total_req =
1446 const_req +
1447 code_req +
1448 pad_req +
1449 stub_req +
1450 exception_handler_req +
1451 deopt_handler_req; // deopt handler
1452
1453 if (C->has_method_handle_invokes())
1454 total_req += deopt_handler_req; // deopt MH handler
1455
1456 CodeBuffer* cb = code_buffer();
1457 cb->set_const_section_alignment(constant_table().alignment());
1458 cb->initialize(total_req, _buf_sizes._reloc);
1459
1460 // Have we run out of code space?
1461 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1462 C->record_failure("CodeCache is full");
1463 return nullptr;
1464 }
1465 // Configure the code buffer.
1466 cb->initialize_consts_size(const_req);
1467 cb->initialize_stubs_size(stub_req);
1468 cb->initialize_oop_recorder(C->env()->oop_recorder());
1469
1470 // fill in the nop array for bundling computations
1471 MachNode *_nop_list[Bundle::_nop_count];
1472 Bundle::initialize_nops(_nop_list);
1473
1474 return cb;
1475 }
1476
1477 //------------------------------fill_buffer------------------------------------
1478 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1479 // blk_starts[] contains offsets calculated during short branches processing,
1480 // offsets should not be increased during following steps.
1481
1482 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1483 // of a loop. It is used to determine the padding for loop alignment.
1484 Compile::TracePhase tp(_t_fillBuffer);
1485
1486 compute_loop_first_inst_sizes();
1487
1488 // Create oopmap set.
1489 _oop_map_set = new OopMapSet();
1490
1491 // !!!!! This preserves old handling of oopmaps for now
1492 C->debug_info()->set_oopmaps(_oop_map_set);
1493
1494 uint nblocks = C->cfg()->number_of_blocks();
1495 // Count and start of implicit null check instructions
1496 uint inct_cnt = 0;
1497 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1498
1499 // Count and start of calls
1500 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1501
1502 uint return_offset = 0;
1503 int nop_size = (new MachNopNode())->size(C->regalloc());
1504
1505 int previous_offset = 0;
1506 int current_offset = 0;
1507 int last_call_offset = -1;
1508 int last_avoid_back_to_back_offset = -1;
1509 #ifdef ASSERT
1510 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1511 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1512 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1513 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1514 #endif
1515
1516 // Create an array of unused labels, one for each basic block, if printing is enabled
1517 #if defined(SUPPORT_OPTO_ASSEMBLY)
1518 int* node_offsets = nullptr;
1519 uint node_offset_limit = C->unique();
1520
1521 if (C->print_assembly()) {
1522 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1523 }
1524 if (node_offsets != nullptr) {
1525 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1526 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1527 }
1528 #endif
1529
1530 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1531
1532 // Emit the constant table.
1533 if (C->has_mach_constant_base_node()) {
1534 if (!constant_table().emit(masm)) {
1535 C->record_failure("consts section overflow");
1536 return;
1537 }
1538 }
1539
1540 // Create an array of labels, one for each basic block
1541 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1542 for (uint i = 0; i <= nblocks; i++) {
1543 blk_labels[i].init();
1544 }
1545
1546 // Now fill in the code buffer
1547 Node* delay_slot = nullptr;
1548 for (uint i = 0; i < nblocks; i++) {
1549 Block* block = C->cfg()->get_block(i);
1550 _block = block;
1551 Node* head = block->head();
1552
1553 // If this block needs to start aligned (i.e, can be reached other
1554 // than by falling-thru from the previous block), then force the
1555 // start of a new bundle.
1556 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1557 masm->code()->flush_bundle(true);
1558 }
1559
1560 #ifdef ASSERT
1561 if (!block->is_connector()) {
1562 stringStream st;
1563 block->dump_head(C->cfg(), &st);
1564 masm->block_comment(st.freeze());
1565 }
1566 jmp_target[i] = 0;
1567 jmp_offset[i] = 0;
1568 jmp_size[i] = 0;
1569 jmp_rule[i] = 0;
1570 #endif
1571 int blk_offset = current_offset;
1572
1573 // Define the label at the beginning of the basic block
1574 masm->bind(blk_labels[block->_pre_order]);
1575
1576 uint last_inst = block->number_of_nodes();
1577
1578 // Emit block normally, except for last instruction.
1579 // Emit means "dump code bits into code buffer".
1580 for (uint j = 0; j<last_inst; j++) {
1581 _index = j;
1582
1583 // Get the node
1584 Node* n = block->get_node(j);
1585
1586 // See if delay slots are supported
1587 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1588 assert(delay_slot == nullptr, "no use of delay slot node");
1589 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1590
1591 delay_slot = n;
1592 continue;
1593 }
1594
1595 // If this starts a new instruction group, then flush the current one
1596 // (but allow split bundles)
1597 if (Pipeline::requires_bundling() && starts_bundle(n))
1598 masm->code()->flush_bundle(false);
1599
1600 // Special handling for SafePoint/Call Nodes
1601 bool is_mcall = false;
1602 if (n->is_Mach()) {
1603 MachNode *mach = n->as_Mach();
1604 is_mcall = n->is_MachCall();
1605 bool is_sfn = n->is_MachSafePoint();
1606
1607 // If this requires all previous instructions be flushed, then do so
1608 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1609 masm->code()->flush_bundle(true);
1610 current_offset = masm->offset();
1611 }
1612
1613 // A padding may be needed again since a previous instruction
1614 // could be moved to delay slot.
1615
1616 // align the instruction if necessary
1617 int padding = mach->compute_padding(current_offset);
1618 // Make sure safepoint node for polling is distinct from a call's
1619 // return by adding a nop if needed.
1620 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1621 padding = nop_size;
1622 }
1623 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1624 current_offset == last_avoid_back_to_back_offset) {
1625 // Avoid back to back some instructions.
1626 padding = nop_size;
1627 }
1628
1629 if (padding > 0) {
1630 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1631 int nops_cnt = padding / nop_size;
1632 MachNode *nop = new MachNopNode(nops_cnt);
1633 block->insert_node(nop, j++);
1634 last_inst++;
1635 C->cfg()->map_node_to_block(nop, block);
1636 // Ensure enough space.
1637 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1638 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1639 C->record_failure("CodeCache is full");
1640 return;
1641 }
1642 nop->emit(masm, C->regalloc());
1643 masm->code()->flush_bundle(true);
1644 current_offset = masm->offset();
1645 }
1646
1647 bool observe_safepoint = is_sfn;
1648 // Remember the start of the last call in a basic block
1649 if (is_mcall) {
1650 MachCallNode *mcall = mach->as_MachCall();
1651
1652 if (mcall->entry_point() != nullptr) {
1653 // This destination address is NOT PC-relative
1654 mcall->method_set((intptr_t)mcall->entry_point());
1655 }
1656
1657 // Save the return address
1658 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1659
1660 observe_safepoint = mcall->guaranteed_safepoint();
1661 }
1662
1663 // sfn will be valid whenever mcall is valid now because of inheritance
1664 if (observe_safepoint) {
1665 // Handle special safepoint nodes for synchronization
1666 if (!is_mcall) {
1667 MachSafePointNode *sfn = mach->as_MachSafePoint();
1668 // !!!!! Stubs only need an oopmap right now, so bail out
1669 if (sfn->jvms()->method() == nullptr) {
1670 // Write the oopmap directly to the code blob??!!
1671 continue;
1672 }
1673 } // End synchronization
1674
1675 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1676 current_offset);
1677 Process_OopMap_Node(mach, current_offset);
1678 } // End if safepoint
1679
1680 // If this is a null check, then add the start of the previous instruction to the list
1681 else if( mach->is_MachNullCheck() ) {
1682 inct_starts[inct_cnt++] = previous_offset;
1683 }
1684
1685 // If this is a branch, then fill in the label with the target BB's label
1686 else if (mach->is_MachBranch()) {
1687 // This requires the TRUE branch target be in succs[0]
1688 uint block_num = block->non_connector_successor(0)->_pre_order;
1689
1690 // Try to replace long branch if delay slot is not used,
1691 // it is mostly for back branches since forward branch's
1692 // distance is not updated yet.
1693 bool delay_slot_is_used = valid_bundle_info(n) &&
1694 C->output()->node_bundling(n)->use_unconditional_delay();
1695 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1696 assert(delay_slot == nullptr, "not expecting delay slot node");
1697 int br_size = n->size(C->regalloc());
1698 int offset = blk_starts[block_num] - current_offset;
1699 if (block_num >= i) {
1700 // Current and following block's offset are not
1701 // finalized yet, adjust distance by the difference
1702 // between calculated and final offsets of current block.
1703 offset -= (blk_starts[i] - blk_offset);
1704 }
1705 // In the following code a nop could be inserted before
1706 // the branch which will increase the backward distance.
1707 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1708 if (needs_padding && offset <= 0)
1709 offset -= nop_size;
1710
1711 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1712 // We've got a winner. Replace this branch.
1713 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1714
1715 // Update the jmp_size.
1716 int new_size = replacement->size(C->regalloc());
1717 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1718 // Insert padding between avoid_back_to_back branches.
1719 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1720 MachNode *nop = new MachNopNode();
1721 block->insert_node(nop, j++);
1722 C->cfg()->map_node_to_block(nop, block);
1723 last_inst++;
1724 nop->emit(masm, C->regalloc());
1725 masm->code()->flush_bundle(true);
1726 current_offset = masm->offset();
1727 }
1728 #ifdef ASSERT
1729 jmp_target[i] = block_num;
1730 jmp_offset[i] = current_offset - blk_offset;
1731 jmp_size[i] = new_size;
1732 jmp_rule[i] = mach->rule();
1733 #endif
1734 block->map_node(replacement, j);
1735 mach->subsume_by(replacement, C);
1736 n = replacement;
1737 mach = replacement;
1738 }
1739 }
1740 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1741 } else if (mach->ideal_Opcode() == Op_Jump) {
1742 for (uint h = 0; h < block->_num_succs; h++) {
1743 Block* succs_block = block->_succs[h];
1744 for (uint j = 1; j < succs_block->num_preds(); j++) {
1745 Node* jpn = succs_block->pred(j);
1746 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1747 uint block_num = succs_block->non_connector()->_pre_order;
1748 Label *blkLabel = &blk_labels[block_num];
1749 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1750 }
1751 }
1752 }
1753 } else if (!n->is_Proj()) {
1754 // Remember the beginning of the previous instruction, in case
1755 // it's followed by a flag-kill and a null-check. Happens on
1756 // Intel all the time, with add-to-memory kind of opcodes.
1757 previous_offset = current_offset;
1758 }
1759
1760 // Not an else-if!
1761 // If this is a trap based cmp then add its offset to the list.
1762 if (mach->is_TrapBasedCheckNode()) {
1763 inct_starts[inct_cnt++] = current_offset;
1764 }
1765 }
1766
1767 // Verify that there is sufficient space remaining
1768 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1769 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1770 C->record_failure("CodeCache is full");
1771 return;
1772 }
1773
1774 // Save the offset for the listing
1775 #if defined(SUPPORT_OPTO_ASSEMBLY)
1776 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1777 node_offsets[n->_idx] = masm->offset();
1778 }
1779 #endif
1780 assert(!C->failing_internal() || C->failure_is_artificial(), "Should not reach here if failing.");
1781
1782 // "Normal" instruction case
1783 DEBUG_ONLY(uint instr_offset = masm->offset());
1784 n->emit(masm, C->regalloc());
1785 current_offset = masm->offset();
1786
1787 // Above we only verified that there is enough space in the instruction section.
1788 // However, the instruction may emit stubs that cause code buffer expansion.
1789 // Bail out here if expansion failed due to a lack of code cache space.
1790 if (C->failing()) {
1791 return;
1792 }
1793
1794 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1795 "ret_addr_offset() not within emitted code");
1796 #ifdef ASSERT
1797 uint n_size = n->size(C->regalloc());
1798 if (n_size < (current_offset-instr_offset)) {
1799 MachNode* mach = n->as_Mach();
1800 n->dump();
1801 mach->dump_format(C->regalloc(), tty);
1802 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1803 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1804 tty->print_cr(" ------------------- ");
1805 BufferBlob* blob = this->scratch_buffer_blob();
1806 address blob_begin = blob->content_begin();
1807 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1808 assert(false, "wrong size of mach node");
1809 }
1810 #endif
1811 non_safepoints.observe_instruction(n, current_offset);
1812
1813 // mcall is last "call" that can be a safepoint
1814 // record it so we can see if a poll will directly follow it
1815 // in which case we'll need a pad to make the PcDesc sites unique
1816 // see 5010568. This can be slightly inaccurate but conservative
1817 // in the case that return address is not actually at current_offset.
1818 // This is a small price to pay.
1819
1820 if (is_mcall) {
1821 last_call_offset = current_offset;
1822 }
1823
1824 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1825 // Avoid back to back some instructions.
1826 last_avoid_back_to_back_offset = current_offset;
1827 }
1828
1829 // See if this instruction has a delay slot
1830 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1831 guarantee(delay_slot != nullptr, "expecting delay slot node");
1832
1833 // Back up 1 instruction
1834 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1835
1836 // Save the offset for the listing
1837 #if defined(SUPPORT_OPTO_ASSEMBLY)
1838 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1839 node_offsets[delay_slot->_idx] = masm->offset();
1840 }
1841 #endif
1842
1843 // Support a SafePoint in the delay slot
1844 if (delay_slot->is_MachSafePoint()) {
1845 MachNode *mach = delay_slot->as_Mach();
1846 // !!!!! Stubs only need an oopmap right now, so bail out
1847 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1848 // Write the oopmap directly to the code blob??!!
1849 delay_slot = nullptr;
1850 continue;
1851 }
1852
1853 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1854 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1855 adjusted_offset);
1856 // Generate an OopMap entry
1857 Process_OopMap_Node(mach, adjusted_offset);
1858 }
1859
1860 // Insert the delay slot instruction
1861 delay_slot->emit(masm, C->regalloc());
1862
1863 // Don't reuse it
1864 delay_slot = nullptr;
1865 }
1866
1867 } // End for all instructions in block
1868
1869 // If the next block is the top of a loop, pad this block out to align
1870 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1871 if (i < nblocks-1) {
1872 Block *nb = C->cfg()->get_block(i + 1);
1873 int padding = nb->alignment_padding(current_offset);
1874 if( padding > 0 ) {
1875 MachNode *nop = new MachNopNode(padding / nop_size);
1876 block->insert_node(nop, block->number_of_nodes());
1877 C->cfg()->map_node_to_block(nop, block);
1878 nop->emit(masm, C->regalloc());
1879 current_offset = masm->offset();
1880 }
1881 }
1882 // Verify that the distance for generated before forward
1883 // short branches is still valid.
1884 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1885
1886 // Save new block start offset
1887 blk_starts[i] = blk_offset;
1888 } // End of for all blocks
1889 blk_starts[nblocks] = current_offset;
1890
1891 non_safepoints.flush_at_end();
1892
1893 // Offset too large?
1894 if (C->failing()) return;
1895
1896 // Define a pseudo-label at the end of the code
1897 masm->bind( blk_labels[nblocks] );
1898
1899 // Compute the size of the first block
1900 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1901
1902 #ifdef ASSERT
1903 for (uint i = 0; i < nblocks; i++) { // For all blocks
1904 if (jmp_target[i] != 0) {
1905 int br_size = jmp_size[i];
1906 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1907 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1908 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]);
1909 assert(false, "Displacement too large for short jmp");
1910 }
1911 }
1912 }
1913 #endif
1914
1915 if (!masm->code()->finalize_stubs()) {
1916 C->record_failure("CodeCache is full");
1917 return;
1918 }
1919
1920 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1921 bs->emit_stubs(*masm->code());
1922 if (C->failing()) return;
1923
1924 // Fill in stubs.
1925 assert(masm->inst_mark() == nullptr, "should be.");
1926 _stub_list.emit(*masm);
1927 if (C->failing()) return;
1928
1929 #ifndef PRODUCT
1930 // Information on the size of the method, without the extraneous code
1931 Scheduling::increment_method_size(masm->offset());
1932 #endif
1933
1934 // ------------------
1935 // Fill in exception table entries.
1936 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1937
1938 // Only java methods have exception handlers and deopt handlers
1939 // class HandlerImpl is platform-specific and defined in the *.ad files.
1940 if (C->method()) {
1941 // Emit the exception handler code.
1942 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1943 if (C->failing()) {
1944 return; // CodeBuffer::expand failed
1945 }
1946 // Emit the deopt handler code.
1947 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1948
1949 // Emit the MethodHandle deopt handler code (if required).
1950 if (C->has_method_handle_invokes() && !C->failing()) {
1951 // We can use the same code as for the normal deopt handler, we
1952 // just need a different entry point address.
1953 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1954 }
1955 }
1956
1957 // One last check for failed CodeBuffer::expand:
1958 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1959 C->record_failure("CodeCache is full");
1960 return;
1961 }
1962
1963 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1964 if (C->print_assembly()) {
1965 tty->cr();
1966 tty->print_cr("============================= C2-compiled nmethod ==============================");
1967 }
1968 #endif
1969
1970 #if defined(SUPPORT_OPTO_ASSEMBLY)
1971 // Dump the assembly code, including basic-block numbers
1972 if (C->print_assembly()) {
1973 ttyLocker ttyl; // keep the following output all in one block
1974 if (!VMThread::should_terminate()) { // test this under the tty lock
1975 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1976 // We call them first and write to a stringStream, then we retake the lock to
1977 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1978 ResourceMark rm;
1979 stringStream method_metadata_str;
1980 if (C->method() != nullptr) {
1981 C->method()->print_metadata(&method_metadata_str);
1982 }
1983 stringStream dump_asm_str;
1984 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1985
1986 NoSafepointVerifier nsv;
1987 ttyLocker ttyl2;
1988 // This output goes directly to the tty, not the compiler log.
1989 // To enable tools to match it up with the compilation activity,
1990 // be sure to tag this tty output with the compile ID.
1991 if (xtty != nullptr) {
1992 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1993 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1994 }
1995 if (C->method() != nullptr) {
1996 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1997 tty->print_raw(method_metadata_str.freeze());
1998 } else if (C->stub_name() != nullptr) {
1999 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
2000 }
2001 tty->cr();
2002 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
2003 tty->print_raw(dump_asm_str.freeze());
2004 tty->print_cr("--------------------------------------------------------------------------------");
2005 if (xtty != nullptr) {
2006 xtty->tail("opto_assembly");
2007 }
2008 }
2009 }
2010 #endif
2011 }
2012
2013 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
2014 _inc_table.set_size(cnt);
2015
2016 uint inct_cnt = 0;
2017 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2018 Block* block = C->cfg()->get_block(i);
2019 Node *n = nullptr;
2020 int j;
2021
2022 // Find the branch; ignore trailing NOPs.
2023 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
2024 n = block->get_node(j);
2025 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
2026 break;
2027 }
2028 }
2029
2030 // If we didn't find anything, continue
2031 if (j < 0) {
2032 continue;
2033 }
2034
2035 // Compute ExceptionHandlerTable subtable entry and add it
2036 // (skip empty blocks)
2037 if (n->is_Catch()) {
2038
2039 // Get the offset of the return from the call
2040 uint call_return = call_returns[block->_pre_order];
2041 #ifdef ASSERT
2042 assert( call_return > 0, "no call seen for this basic block" );
2043 while (block->get_node(--j)->is_MachProj()) ;
2044 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
2045 #endif
2046 // last instruction is a CatchNode, find it's CatchProjNodes
2047 int nof_succs = block->_num_succs;
2048 // allocate space
2049 GrowableArray<intptr_t> handler_bcis(nof_succs);
2050 GrowableArray<intptr_t> handler_pcos(nof_succs);
2051 // iterate through all successors
2052 for (int j = 0; j < nof_succs; j++) {
2053 Block* s = block->_succs[j];
2054 bool found_p = false;
2055 for (uint k = 1; k < s->num_preds(); k++) {
2056 Node* pk = s->pred(k);
2057 if (pk->is_CatchProj() && pk->in(0) == n) {
2058 const CatchProjNode* p = pk->as_CatchProj();
2059 found_p = true;
2060 // add the corresponding handler bci & pco information
2061 if (p->_con != CatchProjNode::fall_through_index) {
2062 // p leads to an exception handler (and is not fall through)
2063 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2064 // no duplicates, please
2065 if (!handler_bcis.contains(p->handler_bci())) {
2066 uint block_num = s->non_connector()->_pre_order;
2067 handler_bcis.append(p->handler_bci());
2068 handler_pcos.append(blk_labels[block_num].loc_pos());
2069 }
2070 }
2071 }
2072 }
2073 assert(found_p, "no matching predecessor found");
2074 // Note: Due to empty block removal, one block may have
2075 // several CatchProj inputs, from the same Catch.
2076 }
2077
2078 // Set the offset of the return from the call
2079 assert(handler_bcis.find(-1) != -1, "must have default handler");
2080 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2081 continue;
2082 }
2083
2084 // Handle implicit null exception table updates
2085 if (n->is_MachNullCheck()) {
2086 assert(n->in(1)->as_Mach()->barrier_data() == 0,
2087 "Implicit null checks on memory accesses with barriers are not yet supported");
2088 uint block_num = block->non_connector_successor(0)->_pre_order;
2089 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2090 continue;
2091 }
2092 // Handle implicit exception table updates: trap instructions.
2093 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2094 uint block_num = block->non_connector_successor(0)->_pre_order;
2095 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2096 continue;
2097 }
2098 } // End of for all blocks fill in exception table entries
2099 }
2100
2101 // Static Variables
2102 #ifndef PRODUCT
2103 uint Scheduling::_total_nop_size = 0;
2104 uint Scheduling::_total_method_size = 0;
2105 uint Scheduling::_total_branches = 0;
2106 uint Scheduling::_total_unconditional_delays = 0;
2107 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2108 #endif
2109
2110 // Initializer for class Scheduling
2111
2112 Scheduling::Scheduling(Arena *arena, Compile &compile)
2113 : _arena(arena),
2114 _cfg(compile.cfg()),
2115 _regalloc(compile.regalloc()),
2116 _scheduled(arena),
2117 _available(arena),
2118 _reg_node(arena),
2119 _pinch_free_list(arena),
2120 _next_node(nullptr),
2121 _bundle_instr_count(0),
2122 _bundle_cycle_number(0),
2123 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2124 #ifndef PRODUCT
2125 , _branches(0)
2126 , _unconditional_delays(0)
2127 #endif
2128 {
2129 // Create a MachNopNode
2130 _nop = new MachNopNode();
2131
2132 // Now that the nops are in the array, save the count
2133 // (but allow entries for the nops)
2134 _node_bundling_limit = compile.unique();
2135 uint node_max = _regalloc->node_regs_max_index();
2136
2137 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2138
2139 // This one is persistent within the Compile class
2140 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2141
2142 // Allocate space for fixed-size arrays
2143 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2144 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2145
2146 // Clear the arrays
2147 for (uint i = 0; i < node_max; i++) {
2148 ::new (&_node_bundling_base[i]) Bundle();
2149 }
2150 memset(_uses, 0, node_max * sizeof(short));
2151 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2152
2153 // Clear the bundling information
2154 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2155
2156 // Get the last node
2157 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2158
2159 _next_node = block->get_node(block->number_of_nodes() - 1);
2160 }
2161
2162 #ifndef PRODUCT
2163 // Scheduling destructor
2164 Scheduling::~Scheduling() {
2165 _total_branches += _branches;
2166 _total_unconditional_delays += _unconditional_delays;
2167 }
2168 #endif
2169
2170 // Step ahead "i" cycles
2171 void Scheduling::step(uint i) {
2172
2173 Bundle *bundle = node_bundling(_next_node);
2174 bundle->set_starts_bundle();
2175
2176 // Update the bundle record, but leave the flags information alone
2177 if (_bundle_instr_count > 0) {
2178 bundle->set_instr_count(_bundle_instr_count);
2179 bundle->set_resources_used(_bundle_use.resourcesUsed());
2180 }
2181
2182 // Update the state information
2183 _bundle_instr_count = 0;
2184 _bundle_cycle_number += i;
2185 _bundle_use.step(i);
2186 }
2187
2188 void Scheduling::step_and_clear() {
2189 Bundle *bundle = node_bundling(_next_node);
2190 bundle->set_starts_bundle();
2191
2192 // Update the bundle record
2193 if (_bundle_instr_count > 0) {
2194 bundle->set_instr_count(_bundle_instr_count);
2195 bundle->set_resources_used(_bundle_use.resourcesUsed());
2196
2197 _bundle_cycle_number += 1;
2198 }
2199
2200 // Clear the bundling information
2201 _bundle_instr_count = 0;
2202 _bundle_use.reset();
2203
2204 memcpy(_bundle_use_elements,
2205 Pipeline_Use::elaborated_elements,
2206 sizeof(Pipeline_Use::elaborated_elements));
2207 }
2208
2209 // Perform instruction scheduling and bundling over the sequence of
2210 // instructions in backwards order.
2211 void PhaseOutput::ScheduleAndBundle() {
2212
2213 // Don't optimize this if it isn't a method
2214 if (!C->method())
2215 return;
2216
2217 // Don't optimize this if scheduling is disabled
2218 if (!C->do_scheduling())
2219 return;
2220
2221 // Scheduling code works only with pairs (8 bytes) maximum.
2222 // And when the scalable vector register is used, we may spill/unspill
2223 // the whole reg regardless of the max vector size.
2224 if (C->max_vector_size() > 8 ||
2225 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2226 return;
2227 }
2228
2229 Compile::TracePhase tp(_t_instrSched);
2230
2231 // Create a data structure for all the scheduling information
2232 Scheduling scheduling(Thread::current()->resource_area(), *C);
2233
2234 // Walk backwards over each basic block, computing the needed alignment
2235 // Walk over all the basic blocks
2236 scheduling.DoScheduling();
2237
2238 #ifndef PRODUCT
2239 if (C->trace_opto_output()) {
2240 // Buffer and print all at once
2241 ResourceMark rm;
2242 stringStream ss;
2243 ss.print("\n---- After ScheduleAndBundle ----\n");
2244 print_scheduling(&ss);
2245 tty->print("%s", ss.as_string());
2246 }
2247 #endif
2248 }
2249
2250 #ifndef PRODUCT
2251 // Separated out so that it can be called directly from debugger
2252 void PhaseOutput::print_scheduling() {
2253 print_scheduling(tty);
2254 }
2255
2256 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2257 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2258 output_stream->print("\nBB#%03d:\n", i);
2259 Block* block = C->cfg()->get_block(i);
2260 for (uint j = 0; j < block->number_of_nodes(); j++) {
2261 Node* n = block->get_node(j);
2262 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2263 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2264 n->dump("\n", false, output_stream);
2265 }
2266 }
2267 }
2268 #endif
2269
2270 // See if this node fits into the present instruction bundle
2271 bool Scheduling::NodeFitsInBundle(Node *n) {
2272 uint n_idx = n->_idx;
2273
2274 // If this is the unconditional delay instruction, then it fits
2275 if (n == _unconditional_delay_slot) {
2276 #ifndef PRODUCT
2277 if (_cfg->C->trace_opto_output())
2278 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2279 #endif
2280 return (true);
2281 }
2282
2283 // If the node cannot be scheduled this cycle, skip it
2284 if (_current_latency[n_idx] > _bundle_cycle_number) {
2285 #ifndef PRODUCT
2286 if (_cfg->C->trace_opto_output())
2287 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2288 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2289 #endif
2290 return (false);
2291 }
2292
2293 const Pipeline *node_pipeline = n->pipeline();
2294
2295 uint instruction_count = node_pipeline->instructionCount();
2296 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2297 instruction_count = 0;
2298 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2299 instruction_count++;
2300
2301 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2302 #ifndef PRODUCT
2303 if (_cfg->C->trace_opto_output())
2304 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2305 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2306 #endif
2307 return (false);
2308 }
2309
2310 // Don't allow non-machine nodes to be handled this way
2311 if (!n->is_Mach() && instruction_count == 0)
2312 return (false);
2313
2314 // See if there is any overlap
2315 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2316
2317 if (delay > 0) {
2318 #ifndef PRODUCT
2319 if (_cfg->C->trace_opto_output())
2320 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2321 #endif
2322 return false;
2323 }
2324
2325 #ifndef PRODUCT
2326 if (_cfg->C->trace_opto_output())
2327 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2328 #endif
2329
2330 return true;
2331 }
2332
2333 Node * Scheduling::ChooseNodeToBundle() {
2334 uint siz = _available.size();
2335
2336 if (siz == 0) {
2337
2338 #ifndef PRODUCT
2339 if (_cfg->C->trace_opto_output())
2340 tty->print("# ChooseNodeToBundle: null\n");
2341 #endif
2342 return (nullptr);
2343 }
2344
2345 // Fast path, if only 1 instruction in the bundle
2346 if (siz == 1) {
2347 #ifndef PRODUCT
2348 if (_cfg->C->trace_opto_output()) {
2349 tty->print("# ChooseNodeToBundle (only 1): ");
2350 _available[0]->dump();
2351 }
2352 #endif
2353 return (_available[0]);
2354 }
2355
2356 // Don't bother, if the bundle is already full
2357 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2358 for ( uint i = 0; i < siz; i++ ) {
2359 Node *n = _available[i];
2360
2361 // Skip projections, we'll handle them another way
2362 if (n->is_Proj())
2363 continue;
2364
2365 // This presupposed that instructions are inserted into the
2366 // available list in a legality order; i.e. instructions that
2367 // must be inserted first are at the head of the list
2368 if (NodeFitsInBundle(n)) {
2369 #ifndef PRODUCT
2370 if (_cfg->C->trace_opto_output()) {
2371 tty->print("# ChooseNodeToBundle: ");
2372 n->dump();
2373 }
2374 #endif
2375 return (n);
2376 }
2377 }
2378 }
2379
2380 // Nothing fits in this bundle, choose the highest priority
2381 #ifndef PRODUCT
2382 if (_cfg->C->trace_opto_output()) {
2383 tty->print("# ChooseNodeToBundle: ");
2384 _available[0]->dump();
2385 }
2386 #endif
2387
2388 return _available[0];
2389 }
2390
2391 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2392 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2393
2394 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2395 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2396 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2397 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2398
2399 // Comparison between stack -> reg and stack -> reg
2400 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2401 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2402 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2403 }
2404
2405 // Comparison between reg -> stack and reg -> stack
2406 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2407 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2408 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2409 }
2410
2411 return 0; // Not comparable
2412 }
2413
2414 void Scheduling::AddNodeToAvailableList(Node *n) {
2415 assert( !n->is_Proj(), "projections never directly made available" );
2416 #ifndef PRODUCT
2417 if (_cfg->C->trace_opto_output()) {
2418 tty->print("# AddNodeToAvailableList: ");
2419 n->dump();
2420 }
2421 #endif
2422
2423 int latency = _current_latency[n->_idx];
2424
2425 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2426 // for stack spilling or unspilling have the same latency, we sort
2427 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2428 // have more opportunities to do ld/st merging
2429 uint i;
2430 for (i = 0; i < _available.size(); i++) {
2431 if (_current_latency[_available[i]->_idx] > latency) {
2432 break;
2433 } else if (_current_latency[_available[i]->_idx] == latency &&
2434 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2435 compare_two_spill_nodes(n, _available[i]) > 0) {
2436 break;
2437 }
2438 }
2439
2440 // Special Check for compares following branches
2441 if( n->is_Mach() && _scheduled.size() > 0 ) {
2442 int op = n->as_Mach()->ideal_Opcode();
2443 Node *last = _scheduled[0];
2444 if( last->is_MachIf() && last->in(1) == n &&
2445 ( op == Op_CmpI ||
2446 op == Op_CmpU ||
2447 op == Op_CmpUL ||
2448 op == Op_CmpP ||
2449 op == Op_CmpF ||
2450 op == Op_CmpD ||
2451 op == Op_CmpL ) ) {
2452
2453 // Recalculate position, moving to front of same latency
2454 for ( i=0 ; i < _available.size(); i++ )
2455 if (_current_latency[_available[i]->_idx] >= latency)
2456 break;
2457 }
2458 }
2459
2460 // Insert the node in the available list
2461 _available.insert(i, n);
2462
2463 #ifndef PRODUCT
2464 if (_cfg->C->trace_opto_output())
2465 dump_available();
2466 #endif
2467 }
2468
2469 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2470 for ( uint i=0; i < n->len(); i++ ) {
2471 Node *def = n->in(i);
2472 if (!def) continue;
2473 if( def->is_Proj() ) // If this is a machine projection, then
2474 def = def->in(0); // propagate usage thru to the base instruction
2475
2476 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2477 continue;
2478 }
2479
2480 // Compute the latency
2481 uint l = _bundle_cycle_number + n->latency(i);
2482 if (_current_latency[def->_idx] < l)
2483 _current_latency[def->_idx] = l;
2484
2485 // If this does not have uses then schedule it
2486 if ((--_uses[def->_idx]) == 0)
2487 AddNodeToAvailableList(def);
2488 }
2489 }
2490
2491 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2492 #ifndef PRODUCT
2493 if (_cfg->C->trace_opto_output()) {
2494 tty->print("# AddNodeToBundle: ");
2495 n->dump();
2496 }
2497 #endif
2498
2499 // Remove this from the available list
2500 uint i;
2501 for (i = 0; i < _available.size(); i++)
2502 if (_available[i] == n)
2503 break;
2504 assert(i < _available.size(), "entry in _available list not found");
2505 _available.remove(i);
2506
2507 // See if this fits in the current bundle
2508 const Pipeline *node_pipeline = n->pipeline();
2509 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2510
2511 // Check for instructions to be placed in the delay slot. We
2512 // do this before we actually schedule the current instruction,
2513 // because the delay slot follows the current instruction.
2514 if (Pipeline::_branch_has_delay_slot &&
2515 node_pipeline->hasBranchDelay() &&
2516 !_unconditional_delay_slot) {
2517
2518 uint siz = _available.size();
2519
2520 // Conditional branches can support an instruction that
2521 // is unconditionally executed and not dependent by the
2522 // branch, OR a conditionally executed instruction if
2523 // the branch is taken. In practice, this means that
2524 // the first instruction at the branch target is
2525 // copied to the delay slot, and the branch goes to
2526 // the instruction after that at the branch target
2527 if ( n->is_MachBranch() ) {
2528
2529 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2530 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2531
2532 #ifndef PRODUCT
2533 _branches++;
2534 #endif
2535
2536 // At least 1 instruction is on the available list
2537 // that is not dependent on the branch
2538 for (uint i = 0; i < siz; i++) {
2539 Node *d = _available[i];
2540 const Pipeline *avail_pipeline = d->pipeline();
2541
2542 // Don't allow safepoints in the branch shadow, that will
2543 // cause a number of difficulties
2544 if ( avail_pipeline->instructionCount() == 1 &&
2545 !avail_pipeline->hasMultipleBundles() &&
2546 !avail_pipeline->hasBranchDelay() &&
2547 Pipeline::instr_has_unit_size() &&
2548 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2549 NodeFitsInBundle(d) &&
2550 !node_bundling(d)->used_in_delay()) {
2551
2552 if (d->is_Mach() && !d->is_MachSafePoint()) {
2553 // A node that fits in the delay slot was found, so we need to
2554 // set the appropriate bits in the bundle pipeline information so
2555 // that it correctly indicates resource usage. Later, when we
2556 // attempt to add this instruction to the bundle, we will skip
2557 // setting the resource usage.
2558 _unconditional_delay_slot = d;
2559 node_bundling(n)->set_use_unconditional_delay();
2560 node_bundling(d)->set_used_in_unconditional_delay();
2561 _bundle_use.add_usage(avail_pipeline->resourceUse());
2562 _current_latency[d->_idx] = _bundle_cycle_number;
2563 _next_node = d;
2564 ++_bundle_instr_count;
2565 #ifndef PRODUCT
2566 _unconditional_delays++;
2567 #endif
2568 break;
2569 }
2570 }
2571 }
2572 }
2573
2574 // No delay slot, add a nop to the usage
2575 if (!_unconditional_delay_slot) {
2576 // See if adding an instruction in the delay slot will overflow
2577 // the bundle.
2578 if (!NodeFitsInBundle(_nop)) {
2579 #ifndef PRODUCT
2580 if (_cfg->C->trace_opto_output())
2581 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2582 #endif
2583 step(1);
2584 }
2585
2586 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2587 _next_node = _nop;
2588 ++_bundle_instr_count;
2589 }
2590
2591 // See if the instruction in the delay slot requires a
2592 // step of the bundles
2593 if (!NodeFitsInBundle(n)) {
2594 #ifndef PRODUCT
2595 if (_cfg->C->trace_opto_output())
2596 tty->print("# *** STEP(branch won't fit) ***\n");
2597 #endif
2598 // Update the state information
2599 _bundle_instr_count = 0;
2600 _bundle_cycle_number += 1;
2601 _bundle_use.step(1);
2602 }
2603 }
2604
2605 // Get the number of instructions
2606 uint instruction_count = node_pipeline->instructionCount();
2607 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2608 instruction_count = 0;
2609
2610 // Compute the latency information
2611 uint delay = 0;
2612
2613 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2614 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2615 if (relative_latency < 0)
2616 relative_latency = 0;
2617
2618 delay = _bundle_use.full_latency(relative_latency, node_usage);
2619
2620 // Does not fit in this bundle, start a new one
2621 if (delay > 0) {
2622 step(delay);
2623
2624 #ifndef PRODUCT
2625 if (_cfg->C->trace_opto_output())
2626 tty->print("# *** STEP(%d) ***\n", delay);
2627 #endif
2628 }
2629 }
2630
2631 // If this was placed in the delay slot, ignore it
2632 if (n != _unconditional_delay_slot) {
2633
2634 if (delay == 0) {
2635 if (node_pipeline->hasMultipleBundles()) {
2636 #ifndef PRODUCT
2637 if (_cfg->C->trace_opto_output())
2638 tty->print("# *** STEP(multiple instructions) ***\n");
2639 #endif
2640 step(1);
2641 }
2642
2643 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2644 #ifndef PRODUCT
2645 if (_cfg->C->trace_opto_output())
2646 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2647 instruction_count + _bundle_instr_count,
2648 Pipeline::_max_instrs_per_cycle);
2649 #endif
2650 step(1);
2651 }
2652 }
2653
2654 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2655 _bundle_instr_count++;
2656
2657 // Set the node's latency
2658 _current_latency[n->_idx] = _bundle_cycle_number;
2659
2660 // Now merge the functional unit information
2661 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2662 _bundle_use.add_usage(node_usage);
2663
2664 // Increment the number of instructions in this bundle
2665 _bundle_instr_count += instruction_count;
2666
2667 // Remember this node for later
2668 if (n->is_Mach())
2669 _next_node = n;
2670 }
2671
2672 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2673 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2674 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2675 // into the block. All other scheduled nodes get put in the schedule here.
2676 int op = n->Opcode();
2677 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2678 (op != Op_Node && // Not an unused antidepedence node and
2679 // not an unallocated boxlock
2680 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2681
2682 // Push any trailing projections
2683 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2684 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2685 Node *foi = n->fast_out(i);
2686 if( foi->is_Proj() )
2687 _scheduled.push(foi);
2688 }
2689 }
2690
2691 // Put the instruction in the schedule list
2692 _scheduled.push(n);
2693 }
2694
2695 #ifndef PRODUCT
2696 if (_cfg->C->trace_opto_output())
2697 dump_available();
2698 #endif
2699
2700 // Walk all the definitions, decrementing use counts, and
2701 // if a definition has a 0 use count, place it in the available list.
2702 DecrementUseCounts(n,bb);
2703 }
2704
2705 // This method sets the use count within a basic block. We will ignore all
2706 // uses outside the current basic block. As we are doing a backwards walk,
2707 // any node we reach that has a use count of 0 may be scheduled. This also
2708 // avoids the problem of cyclic references from phi nodes, as long as phi
2709 // nodes are at the front of the basic block. This method also initializes
2710 // the available list to the set of instructions that have no uses within this
2711 // basic block.
2712 void Scheduling::ComputeUseCount(const Block *bb) {
2713 #ifndef PRODUCT
2714 if (_cfg->C->trace_opto_output())
2715 tty->print("# -> ComputeUseCount\n");
2716 #endif
2717
2718 // Clear the list of available and scheduled instructions, just in case
2719 _available.clear();
2720 _scheduled.clear();
2721
2722 // No delay slot specified
2723 _unconditional_delay_slot = nullptr;
2724
2725 #ifdef ASSERT
2726 for( uint i=0; i < bb->number_of_nodes(); i++ )
2727 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2728 #endif
2729
2730 // Force the _uses count to never go to zero for unscheduable pieces
2731 // of the block
2732 for( uint k = 0; k < _bb_start; k++ )
2733 _uses[bb->get_node(k)->_idx] = 1;
2734 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2735 _uses[bb->get_node(l)->_idx] = 1;
2736
2737 // Iterate backwards over the instructions in the block. Don't count the
2738 // branch projections at end or the block header instructions.
2739 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2740 Node *n = bb->get_node(j);
2741 if( n->is_Proj() ) continue; // Projections handled another way
2742
2743 // Account for all uses
2744 for ( uint k = 0; k < n->len(); k++ ) {
2745 Node *inp = n->in(k);
2746 if (!inp) continue;
2747 assert(inp != n, "no cycles allowed" );
2748 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2749 if (inp->is_Proj()) { // Skip through Proj's
2750 inp = inp->in(0);
2751 }
2752 ++_uses[inp->_idx]; // Count 1 block-local use
2753 }
2754 }
2755
2756 // If this instruction has a 0 use count, then it is available
2757 if (!_uses[n->_idx]) {
2758 _current_latency[n->_idx] = _bundle_cycle_number;
2759 AddNodeToAvailableList(n);
2760 }
2761
2762 #ifndef PRODUCT
2763 if (_cfg->C->trace_opto_output()) {
2764 tty->print("# uses: %3d: ", _uses[n->_idx]);
2765 n->dump();
2766 }
2767 #endif
2768 }
2769
2770 #ifndef PRODUCT
2771 if (_cfg->C->trace_opto_output())
2772 tty->print("# <- ComputeUseCount\n");
2773 #endif
2774 }
2775
2776 // This routine performs scheduling on each basic block in reverse order,
2777 // using instruction latencies and taking into account function unit
2778 // availability.
2779 void Scheduling::DoScheduling() {
2780 #ifndef PRODUCT
2781 if (_cfg->C->trace_opto_output())
2782 tty->print("# -> DoScheduling\n");
2783 #endif
2784
2785 Block *succ_bb = nullptr;
2786 Block *bb;
2787 Compile* C = Compile::current();
2788
2789 // Walk over all the basic blocks in reverse order
2790 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2791 bb = _cfg->get_block(i);
2792
2793 #ifndef PRODUCT
2794 if (_cfg->C->trace_opto_output()) {
2795 tty->print("# Schedule BB#%03d (initial)\n", i);
2796 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2797 bb->get_node(j)->dump();
2798 }
2799 }
2800 #endif
2801
2802 // On the head node, skip processing
2803 if (bb == _cfg->get_root_block()) {
2804 continue;
2805 }
2806
2807 // Skip empty, connector blocks
2808 if (bb->is_connector())
2809 continue;
2810
2811 // If the following block is not the sole successor of
2812 // this one, then reset the pipeline information
2813 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2814 #ifndef PRODUCT
2815 if (_cfg->C->trace_opto_output()) {
2816 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2817 _next_node->_idx, _bundle_instr_count);
2818 }
2819 #endif
2820 step_and_clear();
2821 }
2822
2823 // Leave untouched the starting instruction, any Phis, a CreateEx node
2824 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2825 _bb_end = bb->number_of_nodes()-1;
2826 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2827 Node *n = bb->get_node(_bb_start);
2828 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2829 // Also, MachIdealNodes do not get scheduled
2830 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2831 MachNode *mach = n->as_Mach();
2832 int iop = mach->ideal_Opcode();
2833 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2834 if( iop == Op_Con ) continue; // Do not schedule Top
2835 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2836 mach->pipeline() == MachNode::pipeline_class() &&
2837 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2838 continue;
2839 break; // Funny loop structure to be sure...
2840 }
2841 // Compute last "interesting" instruction in block - last instruction we
2842 // might schedule. _bb_end points just after last schedulable inst. We
2843 // normally schedule conditional branches (despite them being forced last
2844 // in the block), because they have delay slots we can fill. Calls all
2845 // have their delay slots filled in the template expansions, so we don't
2846 // bother scheduling them.
2847 Node *last = bb->get_node(_bb_end);
2848 // Ignore trailing NOPs.
2849 while (_bb_end > 0 && last->is_Mach() &&
2850 last->as_Mach()->ideal_Opcode() == Op_Con) {
2851 last = bb->get_node(--_bb_end);
2852 }
2853 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2854 if( last->is_Catch() ||
2855 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2856 // There might be a prior call. Skip it.
2857 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2858 } else if( last->is_MachNullCheck() ) {
2859 // Backup so the last null-checked memory instruction is
2860 // outside the schedulable range. Skip over the nullcheck,
2861 // projection, and the memory nodes.
2862 Node *mem = last->in(1);
2863 do {
2864 _bb_end--;
2865 } while (mem != bb->get_node(_bb_end));
2866 } else {
2867 // Set _bb_end to point after last schedulable inst.
2868 _bb_end++;
2869 }
2870
2871 assert( _bb_start <= _bb_end, "inverted block ends" );
2872
2873 // Compute the register antidependencies for the basic block
2874 ComputeRegisterAntidependencies(bb);
2875 if (C->failing()) return; // too many D-U pinch points
2876
2877 // Compute the usage within the block, and set the list of all nodes
2878 // in the block that have no uses within the block.
2879 ComputeUseCount(bb);
2880
2881 // Schedule the remaining instructions in the block
2882 while ( _available.size() > 0 ) {
2883 Node *n = ChooseNodeToBundle();
2884 guarantee(n != nullptr, "no nodes available");
2885 AddNodeToBundle(n,bb);
2886 }
2887
2888 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2889 #ifdef ASSERT
2890 for( uint l = _bb_start; l < _bb_end; l++ ) {
2891 Node *n = bb->get_node(l);
2892 uint m;
2893 for( m = 0; m < _bb_end-_bb_start; m++ )
2894 if( _scheduled[m] == n )
2895 break;
2896 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2897 }
2898 #endif
2899
2900 // Now copy the instructions (in reverse order) back to the block
2901 for ( uint k = _bb_start; k < _bb_end; k++ )
2902 bb->map_node(_scheduled[_bb_end-k-1], k);
2903
2904 #ifndef PRODUCT
2905 if (_cfg->C->trace_opto_output()) {
2906 tty->print("# Schedule BB#%03d (final)\n", i);
2907 uint current = 0;
2908 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2909 Node *n = bb->get_node(j);
2910 if( valid_bundle_info(n) ) {
2911 Bundle *bundle = node_bundling(n);
2912 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2913 tty->print("*** Bundle: ");
2914 bundle->dump();
2915 }
2916 n->dump();
2917 }
2918 }
2919 }
2920 #endif
2921 #ifdef ASSERT
2922 verify_good_schedule(bb,"after block local scheduling");
2923 #endif
2924 }
2925
2926 #ifndef PRODUCT
2927 if (_cfg->C->trace_opto_output())
2928 tty->print("# <- DoScheduling\n");
2929 #endif
2930
2931 // Record final node-bundling array location
2932 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2933
2934 } // end DoScheduling
2935
2936 // Verify that no live-range used in the block is killed in the block by a
2937 // wrong DEF. This doesn't verify live-ranges that span blocks.
2938
2939 // Check for edge existence. Used to avoid adding redundant precedence edges.
2940 static bool edge_from_to( Node *from, Node *to ) {
2941 for( uint i=0; i<from->len(); i++ )
2942 if( from->in(i) == to )
2943 return true;
2944 return false;
2945 }
2946
2947 #ifdef ASSERT
2948 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2949 // Check for bad kills
2950 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2951 Node *prior_use = _reg_node[def];
2952 if( prior_use && !edge_from_to(prior_use,n) ) {
2953 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2954 n->dump();
2955 tty->print_cr("...");
2956 prior_use->dump();
2957 assert(edge_from_to(prior_use,n), "%s", msg);
2958 }
2959 _reg_node.map(def,nullptr); // Kill live USEs
2960 }
2961 }
2962
2963 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2964
2965 // Zap to something reasonable for the verify code
2966 _reg_node.clear();
2967
2968 // Walk over the block backwards. Check to make sure each DEF doesn't
2969 // kill a live value (other than the one it's supposed to). Add each
2970 // USE to the live set.
2971 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2972 Node *n = b->get_node(i);
2973 int n_op = n->Opcode();
2974 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2975 // Fat-proj kills a slew of registers
2976 RegMaskIterator rmi(n->out_RegMask());
2977 while (rmi.has_next()) {
2978 OptoReg::Name kill = rmi.next();
2979 verify_do_def(n, kill, msg);
2980 }
2981 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2982 // Get DEF'd registers the normal way
2983 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2984 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2985 }
2986
2987 // Now make all USEs live
2988 for( uint i=1; i<n->req(); i++ ) {
2989 Node *def = n->in(i);
2990 assert(def != nullptr, "input edge required");
2991 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2992 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2993 if( OptoReg::is_valid(reg_lo) ) {
2994 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2995 _reg_node.map(reg_lo,n);
2996 }
2997 if( OptoReg::is_valid(reg_hi) ) {
2998 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2999 _reg_node.map(reg_hi,n);
3000 }
3001 }
3002
3003 }
3004
3005 // Zap to something reasonable for the Antidependence code
3006 _reg_node.clear();
3007 }
3008 #endif
3009
3010 // Conditionally add precedence edges. Avoid putting edges on Projs.
3011 static void add_prec_edge_from_to( Node *from, Node *to ) {
3012 if( from->is_Proj() ) { // Put precedence edge on Proj's input
3013 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
3014 from = from->in(0);
3015 }
3016 if( from != to && // No cycles (for things like LD L0,[L0+4] )
3017 !edge_from_to( from, to ) ) // Avoid duplicate edge
3018 from->add_prec(to);
3019 }
3020
3021 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
3022 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
3023 return;
3024
3025 if (OptoReg::is_reg(def_reg)) {
3026 VMReg vmreg = OptoReg::as_VMReg(def_reg);
3027 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
3028 // This is one of the high slots of a vector register.
3029 // ScheduleAndBundle already checked there are no live wide
3030 // vectors in this method so it can be safely ignored.
3031 return;
3032 }
3033 }
3034
3035 Node *pinch = _reg_node[def_reg]; // Get pinch point
3036 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
3037 is_def ) { // Check for a true def (not a kill)
3038 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
3039 return;
3040 }
3041
3042 Node *kill = def; // Rename 'def' to more descriptive 'kill'
3043 DEBUG_ONLY( def = (Node*)((intptr_t)0xdeadbeef); )
3044
3045 // After some number of kills there _may_ be a later def
3046 Node *later_def = nullptr;
3047
3048 Compile* C = Compile::current();
3049
3050 // Finding a kill requires a real pinch-point.
3051 // Check for not already having a pinch-point.
3052 // Pinch points are Op_Node's.
3053 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
3054 later_def = pinch; // Must be def/kill as optimistic pinch-point
3055 if ( _pinch_free_list.size() > 0) {
3056 pinch = _pinch_free_list.pop();
3057 } else {
3058 pinch = new Node(1); // Pinch point to-be
3059 }
3060 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
3061 DEBUG_ONLY( pinch->dump(); );
3062 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
3063 _cfg->C->record_method_not_compilable("too many D-U pinch points");
3064 return;
3065 }
3066 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3067 _reg_node.map(def_reg,pinch); // Record pinch-point
3068 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3069 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3070 pinch->init_req(0, C->top()); // set not null for the next call
3071 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3072 later_def = nullptr; // and no later def
3073 }
3074 pinch->set_req(0,later_def); // Hook later def so we can find it
3075 } else { // Else have valid pinch point
3076 if( pinch->in(0) ) // If there is a later-def
3077 later_def = pinch->in(0); // Get it
3078 }
3079
3080 // Add output-dependence edge from later def to kill
3081 if( later_def ) // If there is some original def
3082 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3083
3084 // See if current kill is also a use, and so is forced to be the pinch-point.
3085 if( pinch->Opcode() == Op_Node ) {
3086 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3087 for( uint i=1; i<uses->req(); i++ ) {
3088 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3089 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3090 // Yes, found a use/kill pinch-point
3091 pinch->set_req(0,nullptr); //
3092 pinch->replace_by(kill); // Move anti-dep edges up
3093 pinch = kill;
3094 _reg_node.map(def_reg,pinch);
3095 return;
3096 }
3097 }
3098 }
3099
3100 // Add edge from kill to pinch-point
3101 add_prec_edge_from_to(kill,pinch);
3102 }
3103
3104 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3105 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3106 return;
3107 Node *pinch = _reg_node[use_reg]; // Get pinch point
3108 // Check for no later def_reg/kill in block
3109 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3110 // Use has to be block-local as well
3111 _cfg->get_block_for_node(use) == b) {
3112 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3113 pinch->req() == 1 ) { // pinch not yet in block?
3114 pinch->del_req(0); // yank pointer to later-def, also set flag
3115 // Insert the pinch-point in the block just after the last use
3116 b->insert_node(pinch, b->find_node(use) + 1);
3117 _bb_end++; // Increase size scheduled region in block
3118 }
3119
3120 add_prec_edge_from_to(pinch,use);
3121 }
3122 }
3123
3124 // We insert antidependences between the reads and following write of
3125 // allocated registers to prevent illegal code motion. Hopefully, the
3126 // number of added references should be fairly small, especially as we
3127 // are only adding references within the current basic block.
3128 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3129
3130 #ifdef ASSERT
3131 verify_good_schedule(b,"before block local scheduling");
3132 #endif
3133
3134 // A valid schedule, for each register independently, is an endless cycle
3135 // of: a def, then some uses (connected to the def by true dependencies),
3136 // then some kills (defs with no uses), finally the cycle repeats with a new
3137 // def. The uses are allowed to float relative to each other, as are the
3138 // kills. No use is allowed to slide past a kill (or def). This requires
3139 // antidependencies between all uses of a single def and all kills that
3140 // follow, up to the next def. More edges are redundant, because later defs
3141 // & kills are already serialized with true or antidependencies. To keep
3142 // the edge count down, we add a 'pinch point' node if there's more than
3143 // one use or more than one kill/def.
3144
3145 // We add dependencies in one bottom-up pass.
3146
3147 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3148
3149 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3150 // register. If not, we record the DEF/KILL in _reg_node, the
3151 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3152 // "pinch point", a new Node that's in the graph but not in the block.
3153 // We put edges from the prior and current DEF/KILLs to the pinch point.
3154 // We put the pinch point in _reg_node. If there's already a pinch point
3155 // we merely add an edge from the current DEF/KILL to the pinch point.
3156
3157 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3158 // put an edge from the pinch point to the USE.
3159
3160 // To be expedient, the _reg_node array is pre-allocated for the whole
3161 // compilation. _reg_node is lazily initialized; it either contains a null,
3162 // or a valid def/kill/pinch-point, or a leftover node from some prior
3163 // block. Leftover node from some prior block is treated like a null (no
3164 // prior def, so no anti-dependence needed). Valid def is distinguished by
3165 // it being in the current block.
3166 bool fat_proj_seen = false;
3167 uint last_safept = _bb_end-1;
3168 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3169 Node* last_safept_node = end_node;
3170 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3171 Node *n = b->get_node(i);
3172 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3173 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3174 // Fat-proj kills a slew of registers
3175 // This can add edges to 'n' and obscure whether or not it was a def,
3176 // hence the is_def flag.
3177 fat_proj_seen = true;
3178 RegMaskIterator rmi(n->out_RegMask());
3179 while (rmi.has_next()) {
3180 OptoReg::Name kill = rmi.next();
3181 anti_do_def(b, n, kill, is_def);
3182 }
3183 } else {
3184 // Get DEF'd registers the normal way
3185 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3186 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3187 }
3188
3189 // Kill projections on a branch should appear to occur on the
3190 // branch, not afterwards, so grab the masks from the projections
3191 // and process them.
3192 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3193 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3194 Node* use = n->fast_out(i);
3195 if (use->is_Proj()) {
3196 RegMaskIterator rmi(use->out_RegMask());
3197 while (rmi.has_next()) {
3198 OptoReg::Name kill = rmi.next();
3199 anti_do_def(b, n, kill, false);
3200 }
3201 }
3202 }
3203 }
3204
3205 // Check each register used by this instruction for a following DEF/KILL
3206 // that must occur afterward and requires an anti-dependence edge.
3207 for( uint j=0; j<n->req(); j++ ) {
3208 Node *def = n->in(j);
3209 if( def ) {
3210 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3211 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3212 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3213 }
3214 }
3215 // Do not allow defs of new derived values to float above GC
3216 // points unless the base is definitely available at the GC point.
3217
3218 Node *m = b->get_node(i);
3219
3220 // Add precedence edge from following safepoint to use of derived pointer
3221 if( last_safept_node != end_node &&
3222 m != last_safept_node) {
3223 for (uint k = 1; k < m->req(); k++) {
3224 const Type *t = m->in(k)->bottom_type();
3225 if( t->isa_oop_ptr() &&
3226 t->is_ptr()->offset() != 0 ) {
3227 last_safept_node->add_prec( m );
3228 break;
3229 }
3230 }
3231
3232 // Do not allow a CheckCastPP node whose input is a raw pointer to
3233 // float past a safepoint. This can occur when a buffered inline
3234 // type is allocated in a loop and the CheckCastPP from that
3235 // allocation is reused outside the loop. If the use inside the
3236 // loop is scalarized the CheckCastPP will no longer be connected
3237 // to the loop safepoint. See JDK-8264340.
3238 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3239 Node *def = m->in(1);
3240 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3241 last_safept_node->add_prec(m);
3242 }
3243 }
3244 }
3245
3246 if( n->jvms() ) { // Precedence edge from derived to safept
3247 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3248 if( b->get_node(last_safept) != last_safept_node ) {
3249 last_safept = b->find_node(last_safept_node);
3250 }
3251 for( uint j=last_safept; j > i; j-- ) {
3252 Node *mach = b->get_node(j);
3253 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3254 mach->add_prec( n );
3255 }
3256 last_safept = i;
3257 last_safept_node = m;
3258 }
3259 }
3260
3261 if (fat_proj_seen) {
3262 // Garbage collect pinch nodes that were not consumed.
3263 // They are usually created by a fat kill MachProj for a call.
3264 garbage_collect_pinch_nodes();
3265 }
3266 }
3267
3268 // Garbage collect pinch nodes for reuse by other blocks.
3269 //
3270 // The block scheduler's insertion of anti-dependence
3271 // edges creates many pinch nodes when the block contains
3272 // 2 or more Calls. A pinch node is used to prevent a
3273 // combinatorial explosion of edges. If a set of kills for a
3274 // register is anti-dependent on a set of uses (or defs), rather
3275 // than adding an edge in the graph between each pair of kill
3276 // and use (or def), a pinch is inserted between them:
3277 //
3278 // use1 use2 use3
3279 // \ | /
3280 // \ | /
3281 // pinch
3282 // / | \
3283 // / | \
3284 // kill1 kill2 kill3
3285 //
3286 // One pinch node is created per register killed when
3287 // the second call is encountered during a backwards pass
3288 // over the block. Most of these pinch nodes are never
3289 // wired into the graph because the register is never
3290 // used or def'ed in the block.
3291 //
3292 void Scheduling::garbage_collect_pinch_nodes() {
3293 #ifndef PRODUCT
3294 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3295 #endif
3296 int trace_cnt = 0;
3297 for (uint k = 0; k < _reg_node.max(); k++) {
3298 Node* pinch = _reg_node[k];
3299 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3300 // no predecence input edges
3301 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3302 cleanup_pinch(pinch);
3303 _pinch_free_list.push(pinch);
3304 _reg_node.map(k, nullptr);
3305 #ifndef PRODUCT
3306 if (_cfg->C->trace_opto_output()) {
3307 trace_cnt++;
3308 if (trace_cnt > 40) {
3309 tty->print("\n");
3310 trace_cnt = 0;
3311 }
3312 tty->print(" %d", pinch->_idx);
3313 }
3314 #endif
3315 }
3316 }
3317 #ifndef PRODUCT
3318 if (_cfg->C->trace_opto_output()) tty->print("\n");
3319 #endif
3320 }
3321
3322 // Clean up a pinch node for reuse.
3323 void Scheduling::cleanup_pinch( Node *pinch ) {
3324 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3325
3326 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3327 Node* use = pinch->last_out(i);
3328 uint uses_found = 0;
3329 for (uint j = use->req(); j < use->len(); j++) {
3330 if (use->in(j) == pinch) {
3331 use->rm_prec(j);
3332 uses_found++;
3333 }
3334 }
3335 assert(uses_found > 0, "must be a precedence edge");
3336 i -= uses_found; // we deleted 1 or more copies of this edge
3337 }
3338 // May have a later_def entry
3339 pinch->set_req(0, nullptr);
3340 }
3341
3342 #ifndef PRODUCT
3343
3344 void Scheduling::dump_available() const {
3345 tty->print("#Availist ");
3346 for (uint i = 0; i < _available.size(); i++)
3347 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3348 tty->cr();
3349 }
3350
3351 // Print Scheduling Statistics
3352 void Scheduling::print_statistics() {
3353 // Print the size added by nops for bundling
3354 tty->print("Nops added %d bytes to total of %d bytes",
3355 _total_nop_size, _total_method_size);
3356 if (_total_method_size > 0)
3357 tty->print(", for %.2f%%",
3358 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3359 tty->print("\n");
3360
3361 // Print the number of branch shadows filled
3362 if (Pipeline::_branch_has_delay_slot) {
3363 tty->print("Of %d branches, %d had unconditional delay slots filled",
3364 _total_branches, _total_unconditional_delays);
3365 if (_total_branches > 0)
3366 tty->print(", for %.2f%%",
3367 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3368 tty->print("\n");
3369 }
3370
3371 uint total_instructions = 0, total_bundles = 0;
3372
3373 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3374 uint bundle_count = _total_instructions_per_bundle[i];
3375 total_instructions += bundle_count * i;
3376 total_bundles += bundle_count;
3377 }
3378
3379 if (total_bundles > 0)
3380 tty->print("Average ILP (excluding nops) is %.2f\n",
3381 ((double)total_instructions) / ((double)total_bundles));
3382 }
3383 #endif
3384
3385 //-----------------------init_scratch_buffer_blob------------------------------
3386 // Construct a temporary BufferBlob and cache it for this compile.
3387 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3388 // If there is already a scratch buffer blob allocated and the
3389 // constant section is big enough, use it. Otherwise free the
3390 // current and allocate a new one.
3391 BufferBlob* blob = scratch_buffer_blob();
3392 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3393 // Use the current blob.
3394 } else {
3395 if (blob != nullptr) {
3396 BufferBlob::free(blob);
3397 }
3398
3399 ResourceMark rm;
3400 _scratch_const_size = const_size;
3401 int size = C2Compiler::initial_code_buffer_size(const_size);
3402 if (C->has_scalarized_args()) {
3403 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3404 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3405 ciMethod* method = C->method();
3406 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3407 int arg_num = 0;
3408 if (!method->is_static()) {
3409 if (method->is_scalarized_arg(arg_num)) {
3410 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3411 }
3412 arg_num++;
3413 }
3414 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3415 if (method->is_scalarized_arg(arg_num)) {
3416 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3417 }
3418 arg_num++;
3419 }
3420 }
3421 blob = BufferBlob::create("Compile::scratch_buffer", size);
3422 // Record the buffer blob for next time.
3423 set_scratch_buffer_blob(blob);
3424 // Have we run out of code space?
3425 if (scratch_buffer_blob() == nullptr) {
3426 // Let CompilerBroker disable further compilations.
3427 C->record_failure("Not enough space for scratch buffer in CodeCache");
3428 return;
3429 }
3430 }
3431
3432 // Initialize the relocation buffers
3433 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3434 set_scratch_locs_memory(locs_buf);
3435 }
3436
3437
3438 //-----------------------scratch_emit_size-------------------------------------
3439 // Helper function that computes size by emitting code
3440 uint PhaseOutput::scratch_emit_size(const Node* n) {
3441 // Start scratch_emit_size section.
3442 set_in_scratch_emit_size(true);
3443
3444 // Emit into a trash buffer and count bytes emitted.
3445 // This is a pretty expensive way to compute a size,
3446 // but it works well enough if seldom used.
3447 // All common fixed-size instructions are given a size
3448 // method by the AD file.
3449 // Note that the scratch buffer blob and locs memory are
3450 // allocated at the beginning of the compile task, and
3451 // may be shared by several calls to scratch_emit_size.
3452 // The allocation of the scratch buffer blob is particularly
3453 // expensive, since it has to grab the code cache lock.
3454 BufferBlob* blob = this->scratch_buffer_blob();
3455 assert(blob != nullptr, "Initialize BufferBlob at start");
3456 assert(blob->size() > MAX_inst_size, "sanity");
3457 relocInfo* locs_buf = scratch_locs_memory();
3458 address blob_begin = blob->content_begin();
3459 address blob_end = (address)locs_buf;
3460 assert(blob->contains(blob_end), "sanity");
3461 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3462 buf.initialize_consts_size(_scratch_const_size);
3463 buf.initialize_stubs_size(MAX_stubs_size);
3464 assert(locs_buf != nullptr, "sanity");
3465 int lsize = MAX_locs_size / 3;
3466 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3467 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3468 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3469 // Mark as scratch buffer.
3470 buf.consts()->set_scratch_emit();
3471 buf.insts()->set_scratch_emit();
3472 buf.stubs()->set_scratch_emit();
3473
3474 // Do the emission.
3475
3476 Label fakeL; // Fake label for branch instructions.
3477 Label* saveL = nullptr;
3478 uint save_bnum = 0;
3479 bool is_branch = n->is_MachBranch();
3480 C2_MacroAssembler masm(&buf);
3481 masm.bind(fakeL);
3482 if (is_branch) {
3483 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3484 n->as_MachBranch()->label_set(&fakeL, 0);
3485 }
3486 n->emit(&masm, C->regalloc());
3487
3488 // Emitting into the scratch buffer should not fail
3489 assert(!C->failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3490
3491 // Restore label.
3492 if (is_branch) {
3493 n->as_MachBranch()->label_set(saveL, save_bnum);
3494 }
3495
3496 // End scratch_emit_size section.
3497 set_in_scratch_emit_size(false);
3498
3499 return buf.insts_size();
3500 }
3501
3502 void PhaseOutput::install() {
3503 if (!C->should_install_code()) {
3504 return;
3505 } else if (C->stub_function() != nullptr) {
3506 install_stub(C->stub_name());
3507 } else {
3508 install_code(C->method(),
3509 C->entry_bci(),
3510 CompileBroker::compiler2(),
3511 C->has_unsafe_access(),
3512 SharedRuntime::is_wide_vector(C->max_vector_size()));
3513 }
3514 }
3515
3516 void PhaseOutput::install_code(ciMethod* target,
3517 int entry_bci,
3518 AbstractCompiler* compiler,
3519 bool has_unsafe_access,
3520 bool has_wide_vectors) {
3521 // Check if we want to skip execution of all compiled code.
3522 {
3523 #ifndef PRODUCT
3524 if (OptoNoExecute) {
3525 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3526 return;
3527 }
3528 #endif
3529 Compile::TracePhase tp(_t_registerMethod);
3530
3531 if (C->is_osr_compilation()) {
3532 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3533 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3534 } else {
3535 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3536 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3537 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3538 }
3539 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3540 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3541 }
3542 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3543 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3544 }
3545 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3546 }
3547
3548 C->env()->register_method(target,
3549 entry_bci,
3550 &_code_offsets,
3551 _orig_pc_slot_offset_in_bytes,
3552 code_buffer(),
3553 frame_size_in_words(),
3554 _oop_map_set,
3555 &_handler_table,
3556 inc_table(),
3557 compiler,
3558 has_unsafe_access,
3559 SharedRuntime::is_wide_vector(C->max_vector_size()),
3560 C->has_monitors(),
3561 C->has_scoped_access(),
3562 0);
3563
3564 if (C->log() != nullptr) { // Print code cache state into compiler log
3565 C->log()->code_cache_state();
3566 }
3567 }
3568 }
3569 void PhaseOutput::install_stub(const char* stub_name) {
3570 // Entry point will be accessed using stub_entry_point();
3571 if (code_buffer() == nullptr) {
3572 Matcher::soft_match_failure();
3573 } else {
3574 if (PrintAssembly && (WizardMode || Verbose))
3575 tty->print_cr("### Stub::%s", stub_name);
3576
3577 if (!C->failing()) {
3578 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3579
3580 // Make the NMethod
3581 // For now we mark the frame as never safe for profile stackwalking
3582 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3583 code_buffer(),
3584 CodeOffsets::frame_never_safe,
3585 // _code_offsets.value(CodeOffsets::Frame_Complete),
3586 frame_size_in_words(),
3587 oop_map_set(),
3588 false,
3589 false);
3590
3591 if (rs == nullptr) {
3592 C->record_failure("CodeCache is full");
3593 } else {
3594 assert(rs->is_runtime_stub(), "sanity check");
3595 C->set_stub_entry_point(rs->entry_point());
3596 AOTCodeCache::store_code_blob(*rs, AOTCodeEntry::C2Blob, C->stub_id(), stub_name);
3597 }
3598 }
3599 }
3600 }
3601
3602 // Support for bundling info
3603 Bundle* PhaseOutput::node_bundling(const Node *n) {
3604 assert(valid_bundle_info(n), "oob");
3605 return &_node_bundling_base[n->_idx];
3606 }
3607
3608 bool PhaseOutput::valid_bundle_info(const Node *n) {
3609 return (_node_bundling_limit > n->_idx);
3610 }
3611
3612 //------------------------------frame_size_in_words-----------------------------
3613 // frame_slots in units of words
3614 int PhaseOutput::frame_size_in_words() const {
3615 // shift is 0 in LP32 and 1 in LP64
3616 const int shift = (LogBytesPerWord - LogBytesPerInt);
3617 int words = _frame_slots >> shift;
3618 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3619 return words;
3620 }
3621
3622 // To bang the stack of this compiled method we use the stack size
3623 // that the interpreter would need in case of a deoptimization. This
3624 // removes the need to bang the stack in the deoptimization blob which
3625 // in turn simplifies stack overflow handling.
3626 int PhaseOutput::bang_size_in_bytes() const {
3627 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3628 }
3629
3630 //------------------------------dump_asm---------------------------------------
3631 // Dump formatted assembly
3632 #if defined(SUPPORT_OPTO_ASSEMBLY)
3633 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3634
3635 int pc_digits = 3; // #chars required for pc
3636 int sb_chars = 3; // #chars for "start bundle" indicator
3637 int tab_size = 8;
3638 if (pcs != nullptr) {
3639 int max_pc = 0;
3640 for (uint i = 0; i < pc_limit; i++) {
3641 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3642 }
3643 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3644 }
3645 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3646
3647 bool cut_short = false;
3648 st->print_cr("#");
3649 st->print("# "); C->tf()->dump_on(st); st->cr();
3650 st->print_cr("#");
3651
3652 // For all blocks
3653 int pc = 0x0; // Program counter
3654 char starts_bundle = ' ';
3655 C->regalloc()->dump_frame();
3656
3657 Node *n = nullptr;
3658 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3659 if (VMThread::should_terminate()) {
3660 cut_short = true;
3661 break;
3662 }
3663 Block* block = C->cfg()->get_block(i);
3664 if (block->is_connector() && !Verbose) {
3665 continue;
3666 }
3667 n = block->head();
3668 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3669 pc = pcs[n->_idx];
3670 st->print("%*.*x", pc_digits, pc_digits, pc);
3671 }
3672 st->fill_to(prefix_len);
3673 block->dump_head(C->cfg(), st);
3674 if (block->is_connector()) {
3675 st->fill_to(prefix_len);
3676 st->print_cr("# Empty connector block");
3677 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3678 st->fill_to(prefix_len);
3679 st->print_cr("# Block is sole successor of call");
3680 }
3681
3682 // For all instructions
3683 Node *delay = nullptr;
3684 for (uint j = 0; j < block->number_of_nodes(); j++) {
3685 if (VMThread::should_terminate()) {
3686 cut_short = true;
3687 break;
3688 }
3689 n = block->get_node(j);
3690 if (valid_bundle_info(n)) {
3691 Bundle* bundle = node_bundling(n);
3692 if (bundle->used_in_unconditional_delay()) {
3693 delay = n;
3694 continue;
3695 }
3696 if (bundle->starts_bundle()) {
3697 starts_bundle = '+';
3698 }
3699 }
3700
3701 if (WizardMode) {
3702 n->dump();
3703 }
3704
3705 if( !n->is_Region() && // Dont print in the Assembly
3706 !n->is_Phi() && // a few noisely useless nodes
3707 !n->is_Proj() &&
3708 !n->is_MachTemp() &&
3709 !n->is_SafePointScalarObject() &&
3710 !n->is_Catch() && // Would be nice to print exception table targets
3711 !n->is_MergeMem() && // Not very interesting
3712 !n->is_top() && // Debug info table constants
3713 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3714 ) {
3715 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3716 pc = pcs[n->_idx];
3717 st->print("%*.*x", pc_digits, pc_digits, pc);
3718 } else {
3719 st->fill_to(pc_digits);
3720 }
3721 st->print(" %c ", starts_bundle);
3722 starts_bundle = ' ';
3723 st->fill_to(prefix_len);
3724 n->format(C->regalloc(), st);
3725 st->cr();
3726 }
3727
3728 // If we have an instruction with a delay slot, and have seen a delay,
3729 // then back up and print it
3730 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3731 // Coverity finding - Explicit null dereferenced.
3732 guarantee(delay != nullptr, "no unconditional delay instruction");
3733 if (WizardMode) delay->dump();
3734
3735 if (node_bundling(delay)->starts_bundle())
3736 starts_bundle = '+';
3737 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3738 pc = pcs[n->_idx];
3739 st->print("%*.*x", pc_digits, pc_digits, pc);
3740 } else {
3741 st->fill_to(pc_digits);
3742 }
3743 st->print(" %c ", starts_bundle);
3744 starts_bundle = ' ';
3745 st->fill_to(prefix_len);
3746 delay->format(C->regalloc(), st);
3747 st->cr();
3748 delay = nullptr;
3749 }
3750
3751 // Dump the exception table as well
3752 if( n->is_Catch() && (Verbose || WizardMode) ) {
3753 // Print the exception table for this offset
3754 _handler_table.print_subtable_for(pc);
3755 }
3756 st->bol(); // Make sure we start on a new line
3757 }
3758 st->cr(); // one empty line between blocks
3759 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3760 } // End of per-block dump
3761
3762 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3763 }
3764 #endif
3765
3766 #ifndef PRODUCT
3767 void PhaseOutput::print_statistics() {
3768 Scheduling::print_statistics();
3769 }
3770 #endif