1 /*
2 * Copyright (c) 1998, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/debugInfo.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/compilerDirectives.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "gc/shared/barrierSet.hpp"
36 #include "gc/shared/gc_globals.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/allocation.inline.hpp"
39 #include "memory/allocation.hpp"
40 #include "opto/ad.hpp"
41 #include "opto/block.hpp"
42 #include "opto/c2compiler.hpp"
43 #include "opto/c2_MacroAssembler.hpp"
44 #include "opto/callnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/node.hpp"
49 #include "opto/optoreg.hpp"
50 #include "opto/output.hpp"
51 #include "opto/regalloc.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/subnode.hpp"
54 #include "opto/type.hpp"
55 #include "runtime/handles.inline.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "utilities/macros.hpp"
58 #include "utilities/powerOfTwo.hpp"
59 #include "utilities/xmlstream.hpp"
60
61 #ifndef PRODUCT
62 #define DEBUG_ARG(x) , x
63 #else
64 #define DEBUG_ARG(x)
65 #endif
66
67 //------------------------------Scheduling----------------------------------
68 // This class contains all the information necessary to implement instruction
69 // scheduling and bundling.
70 class Scheduling {
71
72 private:
73 // Arena to use
74 Arena *_arena;
75
76 // Control-Flow Graph info
77 PhaseCFG *_cfg;
78
79 // Register Allocation info
80 PhaseRegAlloc *_regalloc;
81
82 // Number of nodes in the method
83 uint _node_bundling_limit;
84
85 // List of scheduled nodes. Generated in reverse order
86 Node_List _scheduled;
87
88 // List of nodes currently available for choosing for scheduling
89 Node_List _available;
90
91 // For each instruction beginning a bundle, the number of following
92 // nodes to be bundled with it.
93 Bundle *_node_bundling_base;
94
95 // Mapping from register to Node
96 Node_List _reg_node;
97
98 // Free list for pinch nodes.
99 Node_List _pinch_free_list;
100
101 // Number of uses of this node within the containing basic block.
102 short *_uses;
103
104 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
105 // front, branch+proj at end. Also skips Catch/CProj (same as
106 // branch-at-end), plus just-prior exception-throwing call.
107 uint _bb_start, _bb_end;
108
109 // Latency from the end of the basic block as scheduled
110 unsigned short *_current_latency;
111
112 // Remember the next node
113 Node *_next_node;
114
115 // Use this for an unconditional branch delay slot
116 Node *_unconditional_delay_slot;
117
118 // Pointer to a Nop
119 MachNopNode *_nop;
120
121 // Length of the current bundle, in instructions
122 uint _bundle_instr_count;
123
124 // Current Cycle number, for computing latencies and bundling
125 uint _bundle_cycle_number;
126
127 // Bundle information
128 Pipeline_Use_Element _bundle_use_elements[resource_count];
129 Pipeline_Use _bundle_use;
130
131 // Dump the available list
132 void dump_available() const;
133
134 public:
135 Scheduling(Arena *arena, Compile &compile);
136
137 // Destructor
138 NOT_PRODUCT( ~Scheduling(); )
139
140 // Step ahead "i" cycles
141 void step(uint i);
142
143 // Step ahead 1 cycle, and clear the bundle state (for example,
144 // at a branch target)
145 void step_and_clear();
146
147 Bundle* node_bundling(const Node *n) {
148 assert(valid_bundle_info(n), "oob");
149 return (&_node_bundling_base[n->_idx]);
150 }
151
152 bool valid_bundle_info(const Node *n) const {
153 return (_node_bundling_limit > n->_idx);
154 }
155
156 bool starts_bundle(const Node *n) const {
157 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
158 }
159
160 // Do the scheduling
161 void DoScheduling();
162
163 // Compute the register antidependencies within a basic block
164 void ComputeRegisterAntidependencies(Block *bb);
165 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
166 void verify_good_schedule( Block *b, const char *msg );
167 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
168 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
169
170 // Add a node to the current bundle
171 void AddNodeToBundle(Node *n, const Block *bb);
172
173 // Return an integer less than, equal to, or greater than zero
174 // if the stack offset of the first argument is respectively
175 // less than, equal to, or greater than the second.
176 int compare_two_spill_nodes(Node* first, Node* second);
177
178 // Add a node to the list of available nodes
179 void AddNodeToAvailableList(Node *n);
180
181 // Compute the local use count for the nodes in a block, and compute
182 // the list of instructions with no uses in the block as available
183 void ComputeUseCount(const Block *bb);
184
185 // Choose an instruction from the available list to add to the bundle
186 Node * ChooseNodeToBundle();
187
188 // See if this Node fits into the currently accumulating bundle
189 bool NodeFitsInBundle(Node *n);
190
191 // Decrement the use count for a node
192 void DecrementUseCounts(Node *n, const Block *bb);
193
194 // Garbage collect pinch nodes for reuse by other blocks.
195 void garbage_collect_pinch_nodes();
196 // Clean up a pinch node for reuse (helper for above).
197 void cleanup_pinch( Node *pinch );
198
199 // Information for statistics gathering
200 #ifndef PRODUCT
201 private:
202 // Gather information on size of nops relative to total
203 uint _branches, _unconditional_delays;
204
205 static uint _total_nop_size, _total_method_size;
206 static uint _total_branches, _total_unconditional_delays;
207 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
208
209 public:
210 static void print_statistics();
211
212 static void increment_instructions_per_bundle(uint i) {
213 _total_instructions_per_bundle[i]++;
214 }
215
216 static void increment_nop_size(uint s) {
217 _total_nop_size += s;
218 }
219
220 static void increment_method_size(uint s) {
221 _total_method_size += s;
222 }
223 #endif
224
225 };
226
227 PhaseOutput::PhaseOutput()
228 : Phase(Phase::Output),
229 _code_buffer("Compile::Fill_buffer"),
230 _first_block_size(0),
231 _handler_table(),
232 _inc_table(),
233 _stub_list(),
234 _oop_map_set(nullptr),
235 _scratch_buffer_blob(nullptr),
236 _scratch_locs_memory(nullptr),
237 _scratch_const_size(-1),
238 _in_scratch_emit_size(false),
239 _frame_slots(0),
240 _code_offsets(),
241 _node_bundling_limit(0),
242 _node_bundling_base(nullptr),
243 _orig_pc_slot(0),
244 _orig_pc_slot_offset_in_bytes(0),
245 _buf_sizes(),
246 _block(nullptr),
247 _index(0) {
248 C->set_output(this);
249 if (C->stub_name() == nullptr) {
250 int fixed_slots = C->fixed_slots();
251 if (C->needs_stack_repair()) {
252 fixed_slots -= 2;
253 }
254 // TODO 8284443 Only reserve extra slot if needed
255 if (InlineTypeReturnedAsFields) {
256 fixed_slots -= 2;
257 }
258 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
259 }
260 }
261
262 PhaseOutput::~PhaseOutput() {
263 C->set_output(nullptr);
264 if (_scratch_buffer_blob != nullptr) {
265 BufferBlob::free(_scratch_buffer_blob);
266 }
267 }
268
269 void PhaseOutput::perform_mach_node_analysis() {
270 // Late barrier analysis must be done after schedule and bundle
271 // Otherwise liveness based spilling will fail
272 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
273 bs->late_barrier_analysis();
274
275 pd_perform_mach_node_analysis();
276
277 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
278 }
279
280 // Convert Nodes to instruction bits and pass off to the VM
281 void PhaseOutput::Output() {
282 // RootNode goes
283 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
284
285 // The number of new nodes (mostly MachNop) is proportional to
286 // the number of java calls and inner loops which are aligned.
287 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
288 C->inner_loops()*(OptoLoopAlignment-1)),
289 "out of nodes before code generation" ) ) {
290 return;
291 }
292 // Make sure I can find the Start Node
293 Block *entry = C->cfg()->get_block(1);
294 Block *broot = C->cfg()->get_root_block();
295
296 const StartNode *start = entry->head()->as_Start();
297
298 // Replace StartNode with prolog
299 Label verified_entry;
300 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
301 entry->map_node(prolog, 0);
302 C->cfg()->map_node_to_block(prolog, entry);
303 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
304
305 // Virtual methods need an unverified entry point
306 if (C->is_osr_compilation()) {
307 if (PoisonOSREntry) {
308 // TODO: Should use a ShouldNotReachHereNode...
309 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
310 }
311 } else {
312 if (C->method()) {
313 if (C->method()->has_scalarized_args()) {
314 // Add entry point to unpack all inline type arguments
315 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
316 if (!C->method()->is_static()) {
317 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
318 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
319 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
320 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
321 }
322 } else if (!C->method()->is_static()) {
323 // Insert unvalidated entry point
324 C->cfg()->insert(broot, 0, new MachUEPNode());
325 }
326 }
327 }
328
329 // Break before main entry point
330 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
331 (OptoBreakpoint && C->is_method_compilation()) ||
332 (OptoBreakpointOSR && C->is_osr_compilation()) ||
333 (OptoBreakpointC2R && !C->method()) ) {
334 // checking for C->method() means that OptoBreakpoint does not apply to
335 // runtime stubs or frame converters
336 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
337 }
338
339 // Insert epilogs before every return
340 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
341 Block* block = C->cfg()->get_block(i);
342 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
343 Node* m = block->end();
344 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
345 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
346 block->add_inst(epilog);
347 C->cfg()->map_node_to_block(epilog, block);
348 }
349 }
350 }
351
352 // Keeper of sizing aspects
353 _buf_sizes = BufferSizingData();
354
355 // Initialize code buffer
356 estimate_buffer_size(_buf_sizes._const);
357 if (C->failing()) return;
358
359 // Pre-compute the length of blocks and replace
360 // long branches with short if machine supports it.
361 // Must be done before ScheduleAndBundle due to SPARC delay slots
362 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
363 blk_starts[0] = 0;
364 shorten_branches(blk_starts);
365
366 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
367 // Compute the offsets of the entry points required by the inline type calling convention
368 if (!C->method()->is_static()) {
369 // We have entries at the beginning of the method, implemented by the first 4 nodes.
370 // Entry (unverified) @ offset 0
371 // Verified_Inline_Entry_RO
372 // Inline_Entry (unverified)
373 // Verified_Inline_Entry
374 uint offset = 0;
375 _code_offsets.set_value(CodeOffsets::Entry, offset);
376
377 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
378 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
379
380 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
381 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
382
383 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
384 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
385 } else {
386 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
387 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
388 }
389 }
390
391 ScheduleAndBundle();
392 if (C->failing()) {
393 return;
394 }
395
396 perform_mach_node_analysis();
397
398 // Complete sizing of codebuffer
399 CodeBuffer* cb = init_buffer();
400 if (cb == nullptr || C->failing()) {
401 return;
402 }
403
404 BuildOopMaps();
405
406 if (C->failing()) {
407 return;
408 }
409
410 fill_buffer(cb, blk_starts);
411 }
412
413 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
414 // Determine if we need to generate a stack overflow check.
415 // Do it if the method is not a stub function and
416 // has java calls or has frame size > vm_page_size/8.
417 // The debug VM checks that deoptimization doesn't trigger an
418 // unexpected stack overflow (compiled method stack banging should
419 // guarantee it doesn't happen) so we always need the stack bang in
420 // a debug VM.
421 return (C->stub_function() == nullptr &&
422 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
423 DEBUG_ONLY(|| true)));
424 }
425
426 bool PhaseOutput::need_register_stack_bang() const {
427 // Determine if we need to generate a register stack overflow check.
428 // This is only used on architectures which have split register
429 // and memory stacks (ie. IA64).
430 // Bang if the method is not a stub function and has java calls
431 return (C->stub_function() == nullptr && C->has_java_calls());
432 }
433
434
435 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
436 // of a loop. When aligning a loop we need to provide enough instructions
437 // in cpu's fetch buffer to feed decoders. The loop alignment could be
438 // avoided if we have enough instructions in fetch buffer at the head of a loop.
439 // By default, the size is set to 999999 by Block's constructor so that
440 // a loop will be aligned if the size is not reset here.
441 //
442 // Note: Mach instructions could contain several HW instructions
443 // so the size is estimated only.
444 //
445 void PhaseOutput::compute_loop_first_inst_sizes() {
446 // The next condition is used to gate the loop alignment optimization.
447 // Don't aligned a loop if there are enough instructions at the head of a loop
448 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
449 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
450 // equal to 11 bytes which is the largest address NOP instruction.
451 if (MaxLoopPad < OptoLoopAlignment - 1) {
452 uint last_block = C->cfg()->number_of_blocks() - 1;
453 for (uint i = 1; i <= last_block; i++) {
454 Block* block = C->cfg()->get_block(i);
455 // Check the first loop's block which requires an alignment.
456 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
457 uint sum_size = 0;
458 uint inst_cnt = NumberOfLoopInstrToAlign;
459 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
460
461 // Check subsequent fallthrough blocks if the loop's first
462 // block(s) does not have enough instructions.
463 Block *nb = block;
464 while(inst_cnt > 0 &&
465 i < last_block &&
466 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
467 !nb->has_successor(block)) {
468 i++;
469 nb = C->cfg()->get_block(i);
470 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
471 } // while( inst_cnt > 0 && i < last_block )
472
473 block->set_first_inst_size(sum_size);
474 } // f( b->head()->is_Loop() )
475 } // for( i <= last_block )
476 } // if( MaxLoopPad < OptoLoopAlignment-1 )
477 }
478
479 // The architecture description provides short branch variants for some long
480 // branch instructions. Replace eligible long branches with short branches.
481 void PhaseOutput::shorten_branches(uint* blk_starts) {
482
483 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
484
485 // Compute size of each block, method size, and relocation information size
486 uint nblocks = C->cfg()->number_of_blocks();
487
488 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
489 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
490 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
491
492 // Collect worst case block paddings
493 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
494 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
495
496 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
497 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
498
499 bool has_short_branch_candidate = false;
500
501 // Initialize the sizes to 0
502 int code_size = 0; // Size in bytes of generated code
503 int stub_size = 0; // Size in bytes of all stub entries
504 // Size in bytes of all relocation entries, including those in local stubs.
505 // Start with 2-bytes of reloc info for the unvalidated entry point
506 int reloc_size = 1; // Number of relocation entries
507
508 // Make three passes. The first computes pessimistic blk_starts,
509 // relative jmp_offset and reloc_size information. The second performs
510 // short branch substitution using the pessimistic sizing. The
511 // third inserts nops where needed.
512
513 // Step one, perform a pessimistic sizing pass.
514 uint last_call_adr = max_juint;
515 uint last_avoid_back_to_back_adr = max_juint;
516 uint nop_size = (new MachNopNode())->size(C->regalloc());
517 for (uint i = 0; i < nblocks; i++) { // For all blocks
518 Block* block = C->cfg()->get_block(i);
519 _block = block;
520
521 // During short branch replacement, we store the relative (to blk_starts)
522 // offset of jump in jmp_offset, rather than the absolute offset of jump.
523 // This is so that we do not need to recompute sizes of all nodes when
524 // we compute correct blk_starts in our next sizing pass.
525 jmp_offset[i] = 0;
526 jmp_size[i] = 0;
527 jmp_nidx[i] = -1;
528 DEBUG_ONLY( jmp_target[i] = 0; )
529 DEBUG_ONLY( jmp_rule[i] = 0; )
530
531 // Sum all instruction sizes to compute block size
532 uint last_inst = block->number_of_nodes();
533 uint blk_size = 0;
534 for (uint j = 0; j < last_inst; j++) {
535 _index = j;
536 Node* nj = block->get_node(_index);
537 // Handle machine instruction nodes
538 if (nj->is_Mach()) {
539 MachNode* mach = nj->as_Mach();
540 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
541 reloc_size += mach->reloc();
542 if (mach->is_MachCall()) {
543 // add size information for trampoline stub
544 // class CallStubImpl is platform-specific and defined in the *.ad files.
545 stub_size += CallStubImpl::size_call_trampoline();
546 reloc_size += CallStubImpl::reloc_call_trampoline();
547
548 MachCallNode *mcall = mach->as_MachCall();
549 // This destination address is NOT PC-relative
550
551 if (mcall->entry_point() != nullptr) {
552 mcall->method_set((intptr_t)mcall->entry_point());
553 }
554
555 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
556 stub_size += CompiledStaticCall::to_interp_stub_size();
557 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
558 }
559 } else if (mach->is_MachSafePoint()) {
560 // If call/safepoint are adjacent, account for possible
561 // nop to disambiguate the two safepoints.
562 // ScheduleAndBundle() can rearrange nodes in a block,
563 // check for all offsets inside this block.
564 if (last_call_adr >= blk_starts[i]) {
565 blk_size += nop_size;
566 }
567 }
568 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
569 // Nop is inserted between "avoid back to back" instructions.
570 // ScheduleAndBundle() can rearrange nodes in a block,
571 // check for all offsets inside this block.
572 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
573 blk_size += nop_size;
574 }
575 }
576 if (mach->may_be_short_branch()) {
577 if (!nj->is_MachBranch()) {
578 #ifndef PRODUCT
579 nj->dump(3);
580 #endif
581 Unimplemented();
582 }
583 assert(jmp_nidx[i] == -1, "block should have only one branch");
584 jmp_offset[i] = blk_size;
585 jmp_size[i] = nj->size(C->regalloc());
586 jmp_nidx[i] = j;
587 has_short_branch_candidate = true;
588 }
589 }
590 blk_size += nj->size(C->regalloc());
591 // Remember end of call offset
592 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
593 last_call_adr = blk_starts[i]+blk_size;
594 }
595 // Remember end of avoid_back_to_back offset
596 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
597 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
598 }
599 }
600
601 // When the next block starts a loop, we may insert pad NOP
602 // instructions. Since we cannot know our future alignment,
603 // assume the worst.
604 if (i < nblocks - 1) {
605 Block* nb = C->cfg()->get_block(i + 1);
606 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
607 if (max_loop_pad > 0) {
608 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
609 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
610 // If either is the last instruction in this block, bump by
611 // max_loop_pad in lock-step with blk_size, so sizing
612 // calculations in subsequent blocks still can conservatively
613 // detect that it may the last instruction in this block.
614 if (last_call_adr == blk_starts[i]+blk_size) {
615 last_call_adr += max_loop_pad;
616 }
617 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
618 last_avoid_back_to_back_adr += max_loop_pad;
619 }
620 blk_size += max_loop_pad;
621 block_worst_case_pad[i + 1] = max_loop_pad;
622 }
623 }
624
625 // Save block size; update total method size
626 blk_starts[i+1] = blk_starts[i]+blk_size;
627 }
628
629 // Step two, replace eligible long jumps.
630 bool progress = true;
631 uint last_may_be_short_branch_adr = max_juint;
632 while (has_short_branch_candidate && progress) {
633 progress = false;
634 has_short_branch_candidate = false;
635 int adjust_block_start = 0;
636 for (uint i = 0; i < nblocks; i++) {
637 Block* block = C->cfg()->get_block(i);
638 int idx = jmp_nidx[i];
639 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
640 if (mach != nullptr && mach->may_be_short_branch()) {
641 #ifdef ASSERT
642 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
643 int j;
644 // Find the branch; ignore trailing NOPs.
645 for (j = block->number_of_nodes()-1; j>=0; j--) {
646 Node* n = block->get_node(j);
647 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
648 break;
649 }
650 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
651 #endif
652 int br_size = jmp_size[i];
653 int br_offs = blk_starts[i] + jmp_offset[i];
654
655 // This requires the TRUE branch target be in succs[0]
656 uint bnum = block->non_connector_successor(0)->_pre_order;
657 int offset = blk_starts[bnum] - br_offs;
658 if (bnum > i) { // adjust following block's offset
659 offset -= adjust_block_start;
660 }
661
662 // This block can be a loop header, account for the padding
663 // in the previous block.
664 int block_padding = block_worst_case_pad[i];
665 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
666 // In the following code a nop could be inserted before
667 // the branch which will increase the backward distance.
668 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
669 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
670
671 if (needs_padding && offset <= 0)
672 offset -= nop_size;
673
674 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
675 // We've got a winner. Replace this branch.
676 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
677
678 // Update the jmp_size.
679 int new_size = replacement->size(C->regalloc());
680 int diff = br_size - new_size;
681 assert(diff >= (int)nop_size, "short_branch size should be smaller");
682 // Conservatively take into account padding between
683 // avoid_back_to_back branches. Previous branch could be
684 // converted into avoid_back_to_back branch during next
685 // rounds.
686 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
687 jmp_offset[i] += nop_size;
688 diff -= nop_size;
689 }
690 adjust_block_start += diff;
691 block->map_node(replacement, idx);
692 mach->subsume_by(replacement, C);
693 mach = replacement;
694 progress = true;
695
696 jmp_size[i] = new_size;
697 DEBUG_ONLY( jmp_target[i] = bnum; );
698 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
699 } else {
700 // The jump distance is not short, try again during next iteration.
701 has_short_branch_candidate = true;
702 }
703 } // (mach->may_be_short_branch())
704 if (mach != nullptr && (mach->may_be_short_branch() ||
705 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
706 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
707 }
708 blk_starts[i+1] -= adjust_block_start;
709 }
710 }
711
712 #ifdef ASSERT
713 for (uint i = 0; i < nblocks; i++) { // For all blocks
714 if (jmp_target[i] != 0) {
715 int br_size = jmp_size[i];
716 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
717 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
718 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]);
719 }
720 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
721 }
722 }
723 #endif
724
725 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
726 // after ScheduleAndBundle().
727
728 // ------------------
729 // Compute size for code buffer
730 code_size = blk_starts[nblocks];
731
732 // Relocation records
733 reloc_size += 1; // Relo entry for exception handler
734
735 // Adjust reloc_size to number of record of relocation info
736 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
737 // a relocation index.
738 // The CodeBuffer will expand the locs array if this estimate is too low.
739 reloc_size *= 10 / sizeof(relocInfo);
740
741 _buf_sizes._reloc = reloc_size;
742 _buf_sizes._code = code_size;
743 _buf_sizes._stub = stub_size;
744 }
745
746 //------------------------------FillLocArray-----------------------------------
747 // Create a bit of debug info and append it to the array. The mapping is from
748 // Java local or expression stack to constant, register or stack-slot. For
749 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
750 // entry has been taken care of and caller should skip it).
751 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
752 // This should never have accepted Bad before
753 assert(OptoReg::is_valid(regnum), "location must be valid");
754 return (OptoReg::is_reg(regnum))
755 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
756 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
757 }
758
759
760 ObjectValue*
761 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
762 for (int i = 0; i < objs->length(); i++) {
763 assert(objs->at(i)->is_object(), "corrupt object cache");
764 ObjectValue* sv = (ObjectValue*) objs->at(i);
765 if (sv->id() == id) {
766 return sv;
767 }
768 }
769 // Otherwise..
770 return nullptr;
771 }
772
773 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
774 ObjectValue* sv ) {
775 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
776 objs->append(sv);
777 }
778
779
780 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
781 GrowableArray<ScopeValue*> *array,
782 GrowableArray<ScopeValue*> *objs ) {
783 assert( local, "use _top instead of null" );
784 if (array->length() != idx) {
785 assert(array->length() == idx + 1, "Unexpected array count");
786 // Old functionality:
787 // return
788 // New functionality:
789 // Assert if the local is not top. In product mode let the new node
790 // override the old entry.
791 assert(local == C->top(), "LocArray collision");
792 if (local == C->top()) {
793 return;
794 }
795 array->pop();
796 }
797 const Type *t = local->bottom_type();
798
799 // Is it a safepoint scalar object node?
800 if (local->is_SafePointScalarObject()) {
801 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
802
803 ObjectValue* sv = (ObjectValue*) sv_for_node_id(objs, spobj->_idx);
804 if (sv == nullptr) {
805 ciKlass* cik = t->is_oopptr()->exact_klass();
806 assert(cik->is_instance_klass() ||
807 cik->is_array_klass(), "Not supported allocation.");
808 uint first_ind = spobj->first_index(sfpt->jvms());
809 // Nullable, scalarized inline types have an is_init input
810 // that needs to be checked before using the field values.
811 ScopeValue* is_init = nullptr;
812 if (cik->is_inlinetype()) {
813 Node* init_node = sfpt->in(first_ind++);
814 assert(init_node != nullptr, "is_init node not found");
815 if (!init_node->is_top()) {
816 const TypeInt* init_type = init_node->bottom_type()->is_int();
817 if (init_node->is_Con()) {
818 is_init = new ConstantIntValue(init_type->get_con());
819 } else {
820 OptoReg::Name init_reg = C->regalloc()->get_reg_first(init_node);
821 is_init = new_loc_value(C->regalloc(), init_reg, Location::normal);
822 }
823 }
824 }
825 sv = new ObjectValue(spobj->_idx,
826 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), is_init);
827 set_sv_for_object_node(objs, sv);
828
829 for (uint i = 0; i < spobj->n_fields(); i++) {
830 Node* fld_node = sfpt->in(first_ind+i);
831 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
832 }
833 }
834 array->append(sv);
835 return;
836 } else if (local->is_SafePointScalarMerge()) {
837 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
838 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
839
840 if (mv == NULL) {
841 GrowableArray<ScopeValue*> deps;
842
843 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
844 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
845 assert(deps.length() == 1, "missing value");
846
847 int selector_idx = smerge->selector_idx(sfpt->jvms());
848 (void)FillLocArray(1, NULL, sfpt->in(selector_idx), &deps, NULL);
849 assert(deps.length() == 2, "missing value");
850
851 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
852 set_sv_for_object_node(objs, mv);
853
854 for (uint i = 1; i < smerge->req(); i++) {
855 Node* obj_node = smerge->in(i);
856 (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs);
857 }
858 }
859 array->append(mv);
860 return;
861 }
862
863 // Grab the register number for the local
864 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
865 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
866 // Record the double as two float registers.
867 // The register mask for such a value always specifies two adjacent
868 // float registers, with the lower register number even.
869 // Normally, the allocation of high and low words to these registers
870 // is irrelevant, because nearly all operations on register pairs
871 // (e.g., StoreD) treat them as a single unit.
872 // Here, we assume in addition that the words in these two registers
873 // stored "naturally" (by operations like StoreD and double stores
874 // within the interpreter) such that the lower-numbered register
875 // is written to the lower memory address. This may seem like
876 // a machine dependency, but it is not--it is a requirement on
877 // the author of the <arch>.ad file to ensure that, for every
878 // even/odd double-register pair to which a double may be allocated,
879 // the word in the even single-register is stored to the first
880 // memory word. (Note that register numbers are completely
881 // arbitrary, and are not tied to any machine-level encodings.)
882 #ifdef _LP64
883 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
884 array->append(new ConstantIntValue((jint)0));
885 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
886 } else if ( t->base() == Type::Long ) {
887 array->append(new ConstantIntValue((jint)0));
888 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
889 } else if ( t->base() == Type::RawPtr ) {
890 // jsr/ret return address which must be restored into the full
891 // width 64-bit stack slot.
892 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
893 }
894 #else //_LP64
895 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
896 // Repack the double/long as two jints.
897 // The convention the interpreter uses is that the second local
898 // holds the first raw word of the native double representation.
899 // This is actually reasonable, since locals and stack arrays
900 // grow downwards in all implementations.
901 // (If, on some machine, the interpreter's Java locals or stack
902 // were to grow upwards, the embedded doubles would be word-swapped.)
903 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
904 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
905 }
906 #endif //_LP64
907 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
908 OptoReg::is_reg(regnum) ) {
909 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
910 ? Location::float_in_dbl : Location::normal ));
911 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
912 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
913 ? Location::int_in_long : Location::normal ));
914 } else if( t->base() == Type::NarrowOop ) {
915 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
916 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
917 t->base() == Type::VectorD || t->base() == Type::VectorX ||
918 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
919 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
920 } else if (C->regalloc()->is_oop(local)) {
921 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
922 t->base() == Type::AryPtr,
923 "Unexpected type: %s", t->msg());
924 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
925 } else {
926 assert(t->base() == Type::Int || t->base() == Type::Half ||
927 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
928 "Unexpected type: %s", t->msg());
929 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
930 }
931 return;
932 }
933
934 // No register. It must be constant data.
935 switch (t->base()) {
936 case Type::Half: // Second half of a double
937 ShouldNotReachHere(); // Caller should skip 2nd halves
938 break;
939 case Type::AnyPtr:
940 array->append(new ConstantOopWriteValue(nullptr));
941 break;
942 case Type::AryPtr:
943 case Type::InstPtr: // fall through
944 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
945 break;
946 case Type::NarrowOop:
947 if (t == TypeNarrowOop::NULL_PTR) {
948 array->append(new ConstantOopWriteValue(nullptr));
949 } else {
950 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
951 }
952 break;
953 case Type::Int:
954 array->append(new ConstantIntValue(t->is_int()->get_con()));
955 break;
956 case Type::RawPtr:
957 // A return address (T_ADDRESS).
958 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
959 #ifdef _LP64
960 // Must be restored to the full-width 64-bit stack slot.
961 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
962 #else
963 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
964 #endif
965 break;
966 case Type::FloatCon: {
967 float f = t->is_float_constant()->getf();
968 array->append(new ConstantIntValue(jint_cast(f)));
969 break;
970 }
971 case Type::DoubleCon: {
972 jdouble d = t->is_double_constant()->getd();
973 #ifdef _LP64
974 array->append(new ConstantIntValue((jint)0));
975 array->append(new ConstantDoubleValue(d));
976 #else
977 // Repack the double as two jints.
978 // The convention the interpreter uses is that the second local
979 // holds the first raw word of the native double representation.
980 // This is actually reasonable, since locals and stack arrays
981 // grow downwards in all implementations.
982 // (If, on some machine, the interpreter's Java locals or stack
983 // were to grow upwards, the embedded doubles would be word-swapped.)
984 jlong_accessor acc;
985 acc.long_value = jlong_cast(d);
986 array->append(new ConstantIntValue(acc.words[1]));
987 array->append(new ConstantIntValue(acc.words[0]));
988 #endif
989 break;
990 }
991 case Type::Long: {
992 jlong d = t->is_long()->get_con();
993 #ifdef _LP64
994 array->append(new ConstantIntValue((jint)0));
995 array->append(new ConstantLongValue(d));
996 #else
997 // Repack the long as two jints.
998 // The convention the interpreter uses is that the second local
999 // holds the first raw word of the native double representation.
1000 // This is actually reasonable, since locals and stack arrays
1001 // grow downwards in all implementations.
1002 // (If, on some machine, the interpreter's Java locals or stack
1003 // were to grow upwards, the embedded doubles would be word-swapped.)
1004 jlong_accessor acc;
1005 acc.long_value = d;
1006 array->append(new ConstantIntValue(acc.words[1]));
1007 array->append(new ConstantIntValue(acc.words[0]));
1008 #endif
1009 break;
1010 }
1011 case Type::Top: // Add an illegal value here
1012 array->append(new LocationValue(Location()));
1013 break;
1014 default:
1015 ShouldNotReachHere();
1016 break;
1017 }
1018 }
1019
1020 // Determine if this node starts a bundle
1021 bool PhaseOutput::starts_bundle(const Node *n) const {
1022 return (_node_bundling_limit > n->_idx &&
1023 _node_bundling_base[n->_idx].starts_bundle());
1024 }
1025
1026 // Determine if there is a monitor that has 'ov' as its owner.
1027 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
1028 for (int k = 0; k < monarray->length(); k++) {
1029 MonitorValue* mv = monarray->at(k);
1030 if (mv->owner() == ov) {
1031 return true;
1032 }
1033 }
1034
1035 return false;
1036 }
1037
1038 //--------------------------Process_OopMap_Node--------------------------------
1039 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1040 // Handle special safepoint nodes for synchronization
1041 MachSafePointNode *sfn = mach->as_MachSafePoint();
1042 MachCallNode *mcall;
1043
1044 int safepoint_pc_offset = current_offset;
1045 bool is_method_handle_invoke = false;
1046 bool return_oop = false;
1047 bool return_scalarized = false;
1048 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1049 bool arg_escape = false;
1050
1051 // Add the safepoint in the DebugInfoRecorder
1052 if( !mach->is_MachCall() ) {
1053 mcall = nullptr;
1054 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1055 } else {
1056 mcall = mach->as_MachCall();
1057
1058 // Is the call a MethodHandle call?
1059 if (mcall->is_MachCallJava()) {
1060 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1061 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1062 is_method_handle_invoke = true;
1063 }
1064 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1065 }
1066
1067 // Check if a call returns an object.
1068 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1069 return_oop = true;
1070 }
1071 if (mcall->returns_scalarized()) {
1072 return_scalarized = true;
1073 }
1074 safepoint_pc_offset += mcall->ret_addr_offset();
1075 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1076 }
1077
1078 // Loop over the JVMState list to add scope information
1079 // Do not skip safepoints with a null method, they need monitor info
1080 JVMState* youngest_jvms = sfn->jvms();
1081 int max_depth = youngest_jvms->depth();
1082
1083 // Allocate the object pool for scalar-replaced objects -- the map from
1084 // small-integer keys (which can be recorded in the local and ostack
1085 // arrays) to descriptions of the object state.
1086 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1087
1088 // Visit scopes from oldest to youngest.
1089 for (int depth = 1; depth <= max_depth; depth++) {
1090 JVMState* jvms = youngest_jvms->of_depth(depth);
1091 int idx;
1092 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1093 // Safepoints that do not have method() set only provide oop-map and monitor info
1094 // to support GC; these do not support deoptimization.
1095 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1096 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1097 int num_mon = jvms->nof_monitors();
1098 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1099 "JVMS local count must match that of the method");
1100
1101 // Add Local and Expression Stack Information
1102
1103 // Insert locals into the locarray
1104 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1105 for( idx = 0; idx < num_locs; idx++ ) {
1106 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1107 }
1108
1109 // Insert expression stack entries into the exparray
1110 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1111 for( idx = 0; idx < num_exps; idx++ ) {
1112 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1113 }
1114
1115 // Add in mappings of the monitors
1116 assert( !method ||
1117 !method->is_synchronized() ||
1118 method->is_native() ||
1119 num_mon > 0 ||
1120 !GenerateSynchronizationCode,
1121 "monitors must always exist for synchronized methods");
1122
1123 // Build the growable array of ScopeValues for exp stack
1124 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1125
1126 // Loop over monitors and insert into array
1127 for (idx = 0; idx < num_mon; idx++) {
1128 // Grab the node that defines this monitor
1129 Node* box_node = sfn->monitor_box(jvms, idx);
1130 Node* obj_node = sfn->monitor_obj(jvms, idx);
1131
1132 // Create ScopeValue for object
1133 ScopeValue *scval = nullptr;
1134
1135 if (obj_node->is_SafePointScalarObject()) {
1136 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1137 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1138 if (scval == nullptr) {
1139 const Type *t = spobj->bottom_type();
1140 ciKlass* cik = t->is_oopptr()->exact_klass();
1141 assert(cik->is_instance_klass() ||
1142 cik->is_array_klass(), "Not supported allocation.");
1143 ObjectValue* sv = new ObjectValue(spobj->_idx,
1144 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1145 PhaseOutput::set_sv_for_object_node(objs, sv);
1146
1147 uint first_ind = spobj->first_index(youngest_jvms);
1148 for (uint i = 0; i < spobj->n_fields(); i++) {
1149 Node* fld_node = sfn->in(first_ind+i);
1150 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1151 }
1152 scval = sv;
1153 }
1154 } else if (!obj_node->is_Con()) {
1155 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1156 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1157 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1158 } else {
1159 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1160 }
1161 } else {
1162 const TypePtr *tp = obj_node->get_ptr_type();
1163 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1164 }
1165
1166 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1167 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1168 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1169 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1170 }
1171
1172 // Mark ObjectValue nodes as root nodes if they are directly
1173 // referenced in the JVMS.
1174 for (int i = 0; i < objs->length(); i++) {
1175 ScopeValue* sv = objs->at(i);
1176 if (sv->is_object_merge()) {
1177 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1178
1179 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1180 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1181 bool is_root = locarray->contains(ov) || exparray->contains(ov) || contains_as_owner(monarray, ov);
1182 ov->set_root(is_root);
1183 }
1184 }
1185 }
1186
1187 // We dump the object pool first, since deoptimization reads it in first.
1188 C->debug_info()->dump_object_pool(objs);
1189
1190 // Build first class objects to pass to scope
1191 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1192 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1193 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1194
1195 // Make method available for all Safepoints
1196 ciMethod* scope_method = method ? method : C->method();
1197 // Describe the scope here
1198 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1199 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1200 // Now we can describe the scope.
1201 methodHandle null_mh;
1202 bool rethrow_exception = false;
1203 C->debug_info()->describe_scope(
1204 safepoint_pc_offset,
1205 null_mh,
1206 scope_method,
1207 jvms->bci(),
1208 jvms->should_reexecute(),
1209 rethrow_exception,
1210 is_method_handle_invoke,
1211 return_oop,
1212 return_scalarized,
1213 has_ea_local_in_scope,
1214 arg_escape,
1215 locvals,
1216 expvals,
1217 monvals
1218 );
1219 } // End jvms loop
1220
1221 // Mark the end of the scope set.
1222 C->debug_info()->end_safepoint(safepoint_pc_offset);
1223 }
1224
1225
1226
1227 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1228 class NonSafepointEmitter {
1229 Compile* C;
1230 JVMState* _pending_jvms;
1231 int _pending_offset;
1232
1233 void emit_non_safepoint();
1234
1235 public:
1236 NonSafepointEmitter(Compile* compile) {
1237 this->C = compile;
1238 _pending_jvms = nullptr;
1239 _pending_offset = 0;
1240 }
1241
1242 void observe_instruction(Node* n, int pc_offset) {
1243 if (!C->debug_info()->recording_non_safepoints()) return;
1244
1245 Node_Notes* nn = C->node_notes_at(n->_idx);
1246 if (nn == nullptr || nn->jvms() == nullptr) return;
1247 if (_pending_jvms != nullptr &&
1248 _pending_jvms->same_calls_as(nn->jvms())) {
1249 // Repeated JVMS? Stretch it up here.
1250 _pending_offset = pc_offset;
1251 } else {
1252 if (_pending_jvms != nullptr &&
1253 _pending_offset < pc_offset) {
1254 emit_non_safepoint();
1255 }
1256 _pending_jvms = nullptr;
1257 if (pc_offset > C->debug_info()->last_pc_offset()) {
1258 // This is the only way _pending_jvms can become non-null:
1259 _pending_jvms = nn->jvms();
1260 _pending_offset = pc_offset;
1261 }
1262 }
1263 }
1264
1265 // Stay out of the way of real safepoints:
1266 void observe_safepoint(JVMState* jvms, int pc_offset) {
1267 if (_pending_jvms != nullptr &&
1268 !_pending_jvms->same_calls_as(jvms) &&
1269 _pending_offset < pc_offset) {
1270 emit_non_safepoint();
1271 }
1272 _pending_jvms = nullptr;
1273 }
1274
1275 void flush_at_end() {
1276 if (_pending_jvms != nullptr) {
1277 emit_non_safepoint();
1278 }
1279 _pending_jvms = nullptr;
1280 }
1281 };
1282
1283 void NonSafepointEmitter::emit_non_safepoint() {
1284 JVMState* youngest_jvms = _pending_jvms;
1285 int pc_offset = _pending_offset;
1286
1287 // Clear it now:
1288 _pending_jvms = nullptr;
1289
1290 DebugInformationRecorder* debug_info = C->debug_info();
1291 assert(debug_info->recording_non_safepoints(), "sanity");
1292
1293 debug_info->add_non_safepoint(pc_offset);
1294 int max_depth = youngest_jvms->depth();
1295
1296 // Visit scopes from oldest to youngest.
1297 for (int depth = 1; depth <= max_depth; depth++) {
1298 JVMState* jvms = youngest_jvms->of_depth(depth);
1299 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1300 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1301 methodHandle null_mh;
1302 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1303 }
1304
1305 // Mark the end of the scope set.
1306 debug_info->end_non_safepoint(pc_offset);
1307 }
1308
1309 //------------------------------init_buffer------------------------------------
1310 void PhaseOutput::estimate_buffer_size(int& const_req) {
1311
1312 // Set the initially allocated size
1313 const_req = initial_const_capacity;
1314
1315 // The extra spacing after the code is necessary on some platforms.
1316 // Sometimes we need to patch in a jump after the last instruction,
1317 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1318
1319 // Compute the byte offset where we can store the deopt pc.
1320 if (C->fixed_slots() != 0) {
1321 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1322 }
1323
1324 // Compute prolog code size
1325 _method_size = 0;
1326 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1327 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1328
1329 if (C->has_mach_constant_base_node()) {
1330 uint add_size = 0;
1331 // Fill the constant table.
1332 // Note: This must happen before shorten_branches.
1333 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1334 Block* b = C->cfg()->get_block(i);
1335
1336 for (uint j = 0; j < b->number_of_nodes(); j++) {
1337 Node* n = b->get_node(j);
1338
1339 // If the node is a MachConstantNode evaluate the constant
1340 // value section.
1341 if (n->is_MachConstant()) {
1342 MachConstantNode* machcon = n->as_MachConstant();
1343 machcon->eval_constant(C);
1344 } else if (n->is_Mach()) {
1345 // On Power there are more nodes that issue constants.
1346 add_size += (n->as_Mach()->ins_num_consts() * 8);
1347 }
1348 }
1349 }
1350
1351 // Calculate the offsets of the constants and the size of the
1352 // constant table (including the padding to the next section).
1353 constant_table().calculate_offsets_and_size();
1354 const_req = constant_table().size() + add_size;
1355 }
1356
1357 // Initialize the space for the BufferBlob used to find and verify
1358 // instruction size in MachNode::emit_size()
1359 init_scratch_buffer_blob(const_req);
1360 }
1361
1362 CodeBuffer* PhaseOutput::init_buffer() {
1363 int stub_req = _buf_sizes._stub;
1364 int code_req = _buf_sizes._code;
1365 int const_req = _buf_sizes._const;
1366
1367 int pad_req = NativeCall::instruction_size;
1368
1369 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1370 stub_req += bs->estimate_stub_size();
1371
1372 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1373 // class HandlerImpl is platform-specific and defined in the *.ad files.
1374 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1375 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1376 stub_req += MAX_stubs_size; // ensure per-stub margin
1377 code_req += MAX_inst_size; // ensure per-instruction margin
1378
1379 if (StressCodeBuffers)
1380 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1381
1382 int total_req =
1383 const_req +
1384 code_req +
1385 pad_req +
1386 stub_req +
1387 exception_handler_req +
1388 deopt_handler_req; // deopt handler
1389
1390 if (C->has_method_handle_invokes())
1391 total_req += deopt_handler_req; // deopt MH handler
1392
1393 CodeBuffer* cb = code_buffer();
1394 cb->initialize(total_req, _buf_sizes._reloc);
1395
1396 // Have we run out of code space?
1397 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1398 C->record_failure("CodeCache is full");
1399 return nullptr;
1400 }
1401 // Configure the code buffer.
1402 cb->initialize_consts_size(const_req);
1403 cb->initialize_stubs_size(stub_req);
1404 cb->initialize_oop_recorder(C->env()->oop_recorder());
1405
1406 // fill in the nop array for bundling computations
1407 MachNode *_nop_list[Bundle::_nop_count];
1408 Bundle::initialize_nops(_nop_list);
1409
1410 return cb;
1411 }
1412
1413 //------------------------------fill_buffer------------------------------------
1414 void PhaseOutput::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1415 // blk_starts[] contains offsets calculated during short branches processing,
1416 // offsets should not be increased during following steps.
1417
1418 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1419 // of a loop. It is used to determine the padding for loop alignment.
1420 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1421
1422 compute_loop_first_inst_sizes();
1423
1424 // Create oopmap set.
1425 _oop_map_set = new OopMapSet();
1426
1427 // !!!!! This preserves old handling of oopmaps for now
1428 C->debug_info()->set_oopmaps(_oop_map_set);
1429
1430 uint nblocks = C->cfg()->number_of_blocks();
1431 // Count and start of implicit null check instructions
1432 uint inct_cnt = 0;
1433 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1434
1435 // Count and start of calls
1436 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1437
1438 uint return_offset = 0;
1439 int nop_size = (new MachNopNode())->size(C->regalloc());
1440
1441 int previous_offset = 0;
1442 int current_offset = 0;
1443 int last_call_offset = -1;
1444 int last_avoid_back_to_back_offset = -1;
1445 #ifdef ASSERT
1446 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1447 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1448 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1449 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1450 #endif
1451
1452 // Create an array of unused labels, one for each basic block, if printing is enabled
1453 #if defined(SUPPORT_OPTO_ASSEMBLY)
1454 int* node_offsets = nullptr;
1455 uint node_offset_limit = C->unique();
1456
1457 if (C->print_assembly()) {
1458 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1459 }
1460 if (node_offsets != nullptr) {
1461 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1462 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1463 }
1464 #endif
1465
1466 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1467
1468 // Emit the constant table.
1469 if (C->has_mach_constant_base_node()) {
1470 if (!constant_table().emit(*cb)) {
1471 C->record_failure("consts section overflow");
1472 return;
1473 }
1474 }
1475
1476 // Create an array of labels, one for each basic block
1477 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1478 for (uint i = 0; i <= nblocks; i++) {
1479 blk_labels[i].init();
1480 }
1481
1482 // Now fill in the code buffer
1483 Node* delay_slot = nullptr;
1484 for (uint i = 0; i < nblocks; i++) {
1485 Block* block = C->cfg()->get_block(i);
1486 _block = block;
1487 Node* head = block->head();
1488
1489 // If this block needs to start aligned (i.e, can be reached other
1490 // than by falling-thru from the previous block), then force the
1491 // start of a new bundle.
1492 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1493 cb->flush_bundle(true);
1494 }
1495
1496 #ifdef ASSERT
1497 if (!block->is_connector()) {
1498 stringStream st;
1499 block->dump_head(C->cfg(), &st);
1500 MacroAssembler(cb).block_comment(st.freeze());
1501 }
1502 jmp_target[i] = 0;
1503 jmp_offset[i] = 0;
1504 jmp_size[i] = 0;
1505 jmp_rule[i] = 0;
1506 #endif
1507 int blk_offset = current_offset;
1508
1509 // Define the label at the beginning of the basic block
1510 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1511
1512 uint last_inst = block->number_of_nodes();
1513
1514 // Emit block normally, except for last instruction.
1515 // Emit means "dump code bits into code buffer".
1516 for (uint j = 0; j<last_inst; j++) {
1517 _index = j;
1518
1519 // Get the node
1520 Node* n = block->get_node(j);
1521
1522 // See if delay slots are supported
1523 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1524 assert(delay_slot == nullptr, "no use of delay slot node");
1525 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1526
1527 delay_slot = n;
1528 continue;
1529 }
1530
1531 // If this starts a new instruction group, then flush the current one
1532 // (but allow split bundles)
1533 if (Pipeline::requires_bundling() && starts_bundle(n))
1534 cb->flush_bundle(false);
1535
1536 // Special handling for SafePoint/Call Nodes
1537 bool is_mcall = false;
1538 if (n->is_Mach()) {
1539 MachNode *mach = n->as_Mach();
1540 is_mcall = n->is_MachCall();
1541 bool is_sfn = n->is_MachSafePoint();
1542
1543 // If this requires all previous instructions be flushed, then do so
1544 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1545 cb->flush_bundle(true);
1546 current_offset = cb->insts_size();
1547 }
1548
1549 // A padding may be needed again since a previous instruction
1550 // could be moved to delay slot.
1551
1552 // align the instruction if necessary
1553 int padding = mach->compute_padding(current_offset);
1554 // Make sure safepoint node for polling is distinct from a call's
1555 // return by adding a nop if needed.
1556 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1557 padding = nop_size;
1558 }
1559 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1560 current_offset == last_avoid_back_to_back_offset) {
1561 // Avoid back to back some instructions.
1562 padding = nop_size;
1563 }
1564
1565 if (padding > 0) {
1566 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1567 int nops_cnt = padding / nop_size;
1568 MachNode *nop = new MachNopNode(nops_cnt);
1569 block->insert_node(nop, j++);
1570 last_inst++;
1571 C->cfg()->map_node_to_block(nop, block);
1572 // Ensure enough space.
1573 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1574 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1575 C->record_failure("CodeCache is full");
1576 return;
1577 }
1578 nop->emit(*cb, C->regalloc());
1579 cb->flush_bundle(true);
1580 current_offset = cb->insts_size();
1581 }
1582
1583 bool observe_safepoint = is_sfn;
1584 // Remember the start of the last call in a basic block
1585 if (is_mcall) {
1586 MachCallNode *mcall = mach->as_MachCall();
1587
1588 if (mcall->entry_point() != nullptr) {
1589 // This destination address is NOT PC-relative
1590 mcall->method_set((intptr_t)mcall->entry_point());
1591 }
1592
1593 // Save the return address
1594 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1595
1596 observe_safepoint = mcall->guaranteed_safepoint();
1597 }
1598
1599 // sfn will be valid whenever mcall is valid now because of inheritance
1600 if (observe_safepoint) {
1601 // Handle special safepoint nodes for synchronization
1602 if (!is_mcall) {
1603 MachSafePointNode *sfn = mach->as_MachSafePoint();
1604 // !!!!! Stubs only need an oopmap right now, so bail out
1605 if (sfn->jvms()->method() == nullptr) {
1606 // Write the oopmap directly to the code blob??!!
1607 continue;
1608 }
1609 } // End synchronization
1610
1611 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1612 current_offset);
1613 Process_OopMap_Node(mach, current_offset);
1614 } // End if safepoint
1615
1616 // If this is a null check, then add the start of the previous instruction to the list
1617 else if( mach->is_MachNullCheck() ) {
1618 inct_starts[inct_cnt++] = previous_offset;
1619 }
1620
1621 // If this is a branch, then fill in the label with the target BB's label
1622 else if (mach->is_MachBranch()) {
1623 // This requires the TRUE branch target be in succs[0]
1624 uint block_num = block->non_connector_successor(0)->_pre_order;
1625
1626 // Try to replace long branch if delay slot is not used,
1627 // it is mostly for back branches since forward branch's
1628 // distance is not updated yet.
1629 bool delay_slot_is_used = valid_bundle_info(n) &&
1630 C->output()->node_bundling(n)->use_unconditional_delay();
1631 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1632 assert(delay_slot == nullptr, "not expecting delay slot node");
1633 int br_size = n->size(C->regalloc());
1634 int offset = blk_starts[block_num] - current_offset;
1635 if (block_num >= i) {
1636 // Current and following block's offset are not
1637 // finalized yet, adjust distance by the difference
1638 // between calculated and final offsets of current block.
1639 offset -= (blk_starts[i] - blk_offset);
1640 }
1641 // In the following code a nop could be inserted before
1642 // the branch which will increase the backward distance.
1643 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1644 if (needs_padding && offset <= 0)
1645 offset -= nop_size;
1646
1647 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1648 // We've got a winner. Replace this branch.
1649 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1650
1651 // Update the jmp_size.
1652 int new_size = replacement->size(C->regalloc());
1653 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1654 // Insert padding between avoid_back_to_back branches.
1655 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1656 MachNode *nop = new MachNopNode();
1657 block->insert_node(nop, j++);
1658 C->cfg()->map_node_to_block(nop, block);
1659 last_inst++;
1660 nop->emit(*cb, C->regalloc());
1661 cb->flush_bundle(true);
1662 current_offset = cb->insts_size();
1663 }
1664 #ifdef ASSERT
1665 jmp_target[i] = block_num;
1666 jmp_offset[i] = current_offset - blk_offset;
1667 jmp_size[i] = new_size;
1668 jmp_rule[i] = mach->rule();
1669 #endif
1670 block->map_node(replacement, j);
1671 mach->subsume_by(replacement, C);
1672 n = replacement;
1673 mach = replacement;
1674 }
1675 }
1676 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1677 } else if (mach->ideal_Opcode() == Op_Jump) {
1678 for (uint h = 0; h < block->_num_succs; h++) {
1679 Block* succs_block = block->_succs[h];
1680 for (uint j = 1; j < succs_block->num_preds(); j++) {
1681 Node* jpn = succs_block->pred(j);
1682 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1683 uint block_num = succs_block->non_connector()->_pre_order;
1684 Label *blkLabel = &blk_labels[block_num];
1685 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1686 }
1687 }
1688 }
1689 }
1690 #ifdef ASSERT
1691 // Check that oop-store precedes the card-mark
1692 else if (mach->ideal_Opcode() == Op_StoreCM) {
1693 uint storeCM_idx = j;
1694 int count = 0;
1695 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1696 Node *oop_store = mach->in(prec); // Precedence edge
1697 if (oop_store == nullptr) continue;
1698 count++;
1699 uint i4;
1700 for (i4 = 0; i4 < last_inst; ++i4) {
1701 if (block->get_node(i4) == oop_store) {
1702 break;
1703 }
1704 }
1705 // Note: This test can provide a false failure if other precedence
1706 // edges have been added to the storeCMNode.
1707 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1708 }
1709 assert(count > 0, "storeCM expects at least one precedence edge");
1710 }
1711 #endif
1712 else if (!n->is_Proj()) {
1713 // Remember the beginning of the previous instruction, in case
1714 // it's followed by a flag-kill and a null-check. Happens on
1715 // Intel all the time, with add-to-memory kind of opcodes.
1716 previous_offset = current_offset;
1717 }
1718
1719 // Not an else-if!
1720 // If this is a trap based cmp then add its offset to the list.
1721 if (mach->is_TrapBasedCheckNode()) {
1722 inct_starts[inct_cnt++] = current_offset;
1723 }
1724 }
1725
1726 // Verify that there is sufficient space remaining
1727 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1728 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1729 C->record_failure("CodeCache is full");
1730 return;
1731 }
1732
1733 // Save the offset for the listing
1734 #if defined(SUPPORT_OPTO_ASSEMBLY)
1735 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1736 node_offsets[n->_idx] = cb->insts_size();
1737 }
1738 #endif
1739 assert(!C->failing(), "Should not reach here if failing.");
1740
1741 // "Normal" instruction case
1742 DEBUG_ONLY(uint instr_offset = cb->insts_size());
1743 n->emit(*cb, C->regalloc());
1744 current_offset = cb->insts_size();
1745
1746 // Above we only verified that there is enough space in the instruction section.
1747 // However, the instruction may emit stubs that cause code buffer expansion.
1748 // Bail out here if expansion failed due to a lack of code cache space.
1749 if (C->failing()) {
1750 return;
1751 }
1752
1753 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1754 "ret_addr_offset() not within emitted code");
1755 #ifdef ASSERT
1756 uint n_size = n->size(C->regalloc());
1757 if (n_size < (current_offset-instr_offset)) {
1758 MachNode* mach = n->as_Mach();
1759 n->dump();
1760 mach->dump_format(C->regalloc(), tty);
1761 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1762 Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty);
1763 tty->print_cr(" ------------------- ");
1764 BufferBlob* blob = this->scratch_buffer_blob();
1765 address blob_begin = blob->content_begin();
1766 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1767 assert(false, "wrong size of mach node");
1768 }
1769 #endif
1770 non_safepoints.observe_instruction(n, current_offset);
1771
1772 // mcall is last "call" that can be a safepoint
1773 // record it so we can see if a poll will directly follow it
1774 // in which case we'll need a pad to make the PcDesc sites unique
1775 // see 5010568. This can be slightly inaccurate but conservative
1776 // in the case that return address is not actually at current_offset.
1777 // This is a small price to pay.
1778
1779 if (is_mcall) {
1780 last_call_offset = current_offset;
1781 }
1782
1783 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1784 // Avoid back to back some instructions.
1785 last_avoid_back_to_back_offset = current_offset;
1786 }
1787
1788 // See if this instruction has a delay slot
1789 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1790 guarantee(delay_slot != nullptr, "expecting delay slot node");
1791
1792 // Back up 1 instruction
1793 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1794
1795 // Save the offset for the listing
1796 #if defined(SUPPORT_OPTO_ASSEMBLY)
1797 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1798 node_offsets[delay_slot->_idx] = cb->insts_size();
1799 }
1800 #endif
1801
1802 // Support a SafePoint in the delay slot
1803 if (delay_slot->is_MachSafePoint()) {
1804 MachNode *mach = delay_slot->as_Mach();
1805 // !!!!! Stubs only need an oopmap right now, so bail out
1806 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1807 // Write the oopmap directly to the code blob??!!
1808 delay_slot = nullptr;
1809 continue;
1810 }
1811
1812 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1813 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1814 adjusted_offset);
1815 // Generate an OopMap entry
1816 Process_OopMap_Node(mach, adjusted_offset);
1817 }
1818
1819 // Insert the delay slot instruction
1820 delay_slot->emit(*cb, C->regalloc());
1821
1822 // Don't reuse it
1823 delay_slot = nullptr;
1824 }
1825
1826 } // End for all instructions in block
1827
1828 // If the next block is the top of a loop, pad this block out to align
1829 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1830 if (i < nblocks-1) {
1831 Block *nb = C->cfg()->get_block(i + 1);
1832 int padding = nb->alignment_padding(current_offset);
1833 if( padding > 0 ) {
1834 MachNode *nop = new MachNopNode(padding / nop_size);
1835 block->insert_node(nop, block->number_of_nodes());
1836 C->cfg()->map_node_to_block(nop, block);
1837 nop->emit(*cb, C->regalloc());
1838 current_offset = cb->insts_size();
1839 }
1840 }
1841 // Verify that the distance for generated before forward
1842 // short branches is still valid.
1843 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1844
1845 // Save new block start offset
1846 blk_starts[i] = blk_offset;
1847 } // End of for all blocks
1848 blk_starts[nblocks] = current_offset;
1849
1850 non_safepoints.flush_at_end();
1851
1852 // Offset too large?
1853 if (C->failing()) return;
1854
1855 // Define a pseudo-label at the end of the code
1856 MacroAssembler(cb).bind( blk_labels[nblocks] );
1857
1858 // Compute the size of the first block
1859 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1860
1861 #ifdef ASSERT
1862 for (uint i = 0; i < nblocks; i++) { // For all blocks
1863 if (jmp_target[i] != 0) {
1864 int br_size = jmp_size[i];
1865 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1866 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1867 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]);
1868 assert(false, "Displacement too large for short jmp");
1869 }
1870 }
1871 }
1872 #endif
1873
1874 if (!cb->finalize_stubs()) {
1875 C->record_failure("CodeCache is full");
1876 return;
1877 }
1878
1879 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1880 bs->emit_stubs(*cb);
1881 if (C->failing()) return;
1882
1883 // Fill in stubs.
1884 _stub_list.emit(*cb);
1885 if (C->failing()) return;
1886
1887 #ifndef PRODUCT
1888 // Information on the size of the method, without the extraneous code
1889 Scheduling::increment_method_size(cb->insts_size());
1890 #endif
1891
1892 // ------------------
1893 // Fill in exception table entries.
1894 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1895
1896 // Only java methods have exception handlers and deopt handlers
1897 // class HandlerImpl is platform-specific and defined in the *.ad files.
1898 if (C->method()) {
1899 // Emit the exception handler code.
1900 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1901 if (C->failing()) {
1902 return; // CodeBuffer::expand failed
1903 }
1904 // Emit the deopt handler code.
1905 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1906
1907 // Emit the MethodHandle deopt handler code (if required).
1908 if (C->has_method_handle_invokes() && !C->failing()) {
1909 // We can use the same code as for the normal deopt handler, we
1910 // just need a different entry point address.
1911 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1912 }
1913 }
1914
1915 // One last check for failed CodeBuffer::expand:
1916 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1917 C->record_failure("CodeCache is full");
1918 return;
1919 }
1920
1921 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1922 if (C->print_assembly()) {
1923 tty->cr();
1924 tty->print_cr("============================= C2-compiled nmethod ==============================");
1925 }
1926 #endif
1927
1928 #if defined(SUPPORT_OPTO_ASSEMBLY)
1929 // Dump the assembly code, including basic-block numbers
1930 if (C->print_assembly()) {
1931 ttyLocker ttyl; // keep the following output all in one block
1932 if (!VMThread::should_terminate()) { // test this under the tty lock
1933 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1934 // We call them first and write to a stringStream, then we retake the lock to
1935 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1936 ResourceMark rm;
1937 stringStream method_metadata_str;
1938 if (C->method() != nullptr) {
1939 C->method()->print_metadata(&method_metadata_str);
1940 }
1941 stringStream dump_asm_str;
1942 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1943
1944 NoSafepointVerifier nsv;
1945 ttyLocker ttyl2;
1946 // This output goes directly to the tty, not the compiler log.
1947 // To enable tools to match it up with the compilation activity,
1948 // be sure to tag this tty output with the compile ID.
1949 if (xtty != nullptr) {
1950 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1951 C->is_osr_compilation() ? " compile_kind='osr'" : "");
1952 }
1953 if (C->method() != nullptr) {
1954 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1955 tty->print_raw(method_metadata_str.freeze());
1956 } else if (C->stub_name() != nullptr) {
1957 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1958 }
1959 tty->cr();
1960 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1961 tty->print_raw(dump_asm_str.freeze());
1962 tty->print_cr("--------------------------------------------------------------------------------");
1963 if (xtty != nullptr) {
1964 xtty->tail("opto_assembly");
1965 }
1966 }
1967 }
1968 #endif
1969 }
1970
1971 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1972 _inc_table.set_size(cnt);
1973
1974 uint inct_cnt = 0;
1975 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1976 Block* block = C->cfg()->get_block(i);
1977 Node *n = nullptr;
1978 int j;
1979
1980 // Find the branch; ignore trailing NOPs.
1981 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1982 n = block->get_node(j);
1983 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1984 break;
1985 }
1986 }
1987
1988 // If we didn't find anything, continue
1989 if (j < 0) {
1990 continue;
1991 }
1992
1993 // Compute ExceptionHandlerTable subtable entry and add it
1994 // (skip empty blocks)
1995 if (n->is_Catch()) {
1996
1997 // Get the offset of the return from the call
1998 uint call_return = call_returns[block->_pre_order];
1999 #ifdef ASSERT
2000 assert( call_return > 0, "no call seen for this basic block" );
2001 while (block->get_node(--j)->is_MachProj()) ;
2002 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
2003 #endif
2004 // last instruction is a CatchNode, find it's CatchProjNodes
2005 int nof_succs = block->_num_succs;
2006 // allocate space
2007 GrowableArray<intptr_t> handler_bcis(nof_succs);
2008 GrowableArray<intptr_t> handler_pcos(nof_succs);
2009 // iterate through all successors
2010 for (int j = 0; j < nof_succs; j++) {
2011 Block* s = block->_succs[j];
2012 bool found_p = false;
2013 for (uint k = 1; k < s->num_preds(); k++) {
2014 Node* pk = s->pred(k);
2015 if (pk->is_CatchProj() && pk->in(0) == n) {
2016 const CatchProjNode* p = pk->as_CatchProj();
2017 found_p = true;
2018 // add the corresponding handler bci & pco information
2019 if (p->_con != CatchProjNode::fall_through_index) {
2020 // p leads to an exception handler (and is not fall through)
2021 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2022 // no duplicates, please
2023 if (!handler_bcis.contains(p->handler_bci())) {
2024 uint block_num = s->non_connector()->_pre_order;
2025 handler_bcis.append(p->handler_bci());
2026 handler_pcos.append(blk_labels[block_num].loc_pos());
2027 }
2028 }
2029 }
2030 }
2031 assert(found_p, "no matching predecessor found");
2032 // Note: Due to empty block removal, one block may have
2033 // several CatchProj inputs, from the same Catch.
2034 }
2035
2036 // Set the offset of the return from the call
2037 assert(handler_bcis.find(-1) != -1, "must have default handler");
2038 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2039 continue;
2040 }
2041
2042 // Handle implicit null exception table updates
2043 if (n->is_MachNullCheck()) {
2044 uint block_num = block->non_connector_successor(0)->_pre_order;
2045 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2046 continue;
2047 }
2048 // Handle implicit exception table updates: trap instructions.
2049 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2050 uint block_num = block->non_connector_successor(0)->_pre_order;
2051 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2052 continue;
2053 }
2054 } // End of for all blocks fill in exception table entries
2055 }
2056
2057 // Static Variables
2058 #ifndef PRODUCT
2059 uint Scheduling::_total_nop_size = 0;
2060 uint Scheduling::_total_method_size = 0;
2061 uint Scheduling::_total_branches = 0;
2062 uint Scheduling::_total_unconditional_delays = 0;
2063 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2064 #endif
2065
2066 // Initializer for class Scheduling
2067
2068 Scheduling::Scheduling(Arena *arena, Compile &compile)
2069 : _arena(arena),
2070 _cfg(compile.cfg()),
2071 _regalloc(compile.regalloc()),
2072 _scheduled(arena),
2073 _available(arena),
2074 _reg_node(arena),
2075 _pinch_free_list(arena),
2076 _next_node(nullptr),
2077 _bundle_instr_count(0),
2078 _bundle_cycle_number(0),
2079 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2080 #ifndef PRODUCT
2081 , _branches(0)
2082 , _unconditional_delays(0)
2083 #endif
2084 {
2085 // Create a MachNopNode
2086 _nop = new MachNopNode();
2087
2088 // Now that the nops are in the array, save the count
2089 // (but allow entries for the nops)
2090 _node_bundling_limit = compile.unique();
2091 uint node_max = _regalloc->node_regs_max_index();
2092
2093 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2094
2095 // This one is persistent within the Compile class
2096 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2097
2098 // Allocate space for fixed-size arrays
2099 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2100 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2101
2102 // Clear the arrays
2103 for (uint i = 0; i < node_max; i++) {
2104 ::new (&_node_bundling_base[i]) Bundle();
2105 }
2106 memset(_uses, 0, node_max * sizeof(short));
2107 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2108
2109 // Clear the bundling information
2110 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2111
2112 // Get the last node
2113 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2114
2115 _next_node = block->get_node(block->number_of_nodes() - 1);
2116 }
2117
2118 #ifndef PRODUCT
2119 // Scheduling destructor
2120 Scheduling::~Scheduling() {
2121 _total_branches += _branches;
2122 _total_unconditional_delays += _unconditional_delays;
2123 }
2124 #endif
2125
2126 // Step ahead "i" cycles
2127 void Scheduling::step(uint i) {
2128
2129 Bundle *bundle = node_bundling(_next_node);
2130 bundle->set_starts_bundle();
2131
2132 // Update the bundle record, but leave the flags information alone
2133 if (_bundle_instr_count > 0) {
2134 bundle->set_instr_count(_bundle_instr_count);
2135 bundle->set_resources_used(_bundle_use.resourcesUsed());
2136 }
2137
2138 // Update the state information
2139 _bundle_instr_count = 0;
2140 _bundle_cycle_number += i;
2141 _bundle_use.step(i);
2142 }
2143
2144 void Scheduling::step_and_clear() {
2145 Bundle *bundle = node_bundling(_next_node);
2146 bundle->set_starts_bundle();
2147
2148 // Update the bundle record
2149 if (_bundle_instr_count > 0) {
2150 bundle->set_instr_count(_bundle_instr_count);
2151 bundle->set_resources_used(_bundle_use.resourcesUsed());
2152
2153 _bundle_cycle_number += 1;
2154 }
2155
2156 // Clear the bundling information
2157 _bundle_instr_count = 0;
2158 _bundle_use.reset();
2159
2160 memcpy(_bundle_use_elements,
2161 Pipeline_Use::elaborated_elements,
2162 sizeof(Pipeline_Use::elaborated_elements));
2163 }
2164
2165 // Perform instruction scheduling and bundling over the sequence of
2166 // instructions in backwards order.
2167 void PhaseOutput::ScheduleAndBundle() {
2168
2169 // Don't optimize this if it isn't a method
2170 if (!C->method())
2171 return;
2172
2173 // Don't optimize this if scheduling is disabled
2174 if (!C->do_scheduling())
2175 return;
2176
2177 // Scheduling code works only with pairs (8 bytes) maximum.
2178 // And when the scalable vector register is used, we may spill/unspill
2179 // the whole reg regardless of the max vector size.
2180 if (C->max_vector_size() > 8 ||
2181 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2182 return;
2183 }
2184
2185 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2186
2187 // Create a data structure for all the scheduling information
2188 Scheduling scheduling(Thread::current()->resource_area(), *C);
2189
2190 // Walk backwards over each basic block, computing the needed alignment
2191 // Walk over all the basic blocks
2192 scheduling.DoScheduling();
2193
2194 #ifndef PRODUCT
2195 if (C->trace_opto_output()) {
2196 // Buffer and print all at once
2197 ResourceMark rm;
2198 stringStream ss;
2199 ss.print("\n---- After ScheduleAndBundle ----\n");
2200 print_scheduling(&ss);
2201 tty->print("%s", ss.as_string());
2202 }
2203 #endif
2204 }
2205
2206 #ifndef PRODUCT
2207 // Separated out so that it can be called directly from debugger
2208 void PhaseOutput::print_scheduling() {
2209 print_scheduling(tty);
2210 }
2211
2212 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2213 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2214 output_stream->print("\nBB#%03d:\n", i);
2215 Block* block = C->cfg()->get_block(i);
2216 for (uint j = 0; j < block->number_of_nodes(); j++) {
2217 Node* n = block->get_node(j);
2218 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2219 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2220 n->dump("\n", false, output_stream);
2221 }
2222 }
2223 }
2224 #endif
2225
2226 // See if this node fits into the present instruction bundle
2227 bool Scheduling::NodeFitsInBundle(Node *n) {
2228 uint n_idx = n->_idx;
2229
2230 // If this is the unconditional delay instruction, then it fits
2231 if (n == _unconditional_delay_slot) {
2232 #ifndef PRODUCT
2233 if (_cfg->C->trace_opto_output())
2234 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2235 #endif
2236 return (true);
2237 }
2238
2239 // If the node cannot be scheduled this cycle, skip it
2240 if (_current_latency[n_idx] > _bundle_cycle_number) {
2241 #ifndef PRODUCT
2242 if (_cfg->C->trace_opto_output())
2243 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2244 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2245 #endif
2246 return (false);
2247 }
2248
2249 const Pipeline *node_pipeline = n->pipeline();
2250
2251 uint instruction_count = node_pipeline->instructionCount();
2252 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2253 instruction_count = 0;
2254 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2255 instruction_count++;
2256
2257 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2258 #ifndef PRODUCT
2259 if (_cfg->C->trace_opto_output())
2260 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2261 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2262 #endif
2263 return (false);
2264 }
2265
2266 // Don't allow non-machine nodes to be handled this way
2267 if (!n->is_Mach() && instruction_count == 0)
2268 return (false);
2269
2270 // See if there is any overlap
2271 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2272
2273 if (delay > 0) {
2274 #ifndef PRODUCT
2275 if (_cfg->C->trace_opto_output())
2276 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2277 #endif
2278 return false;
2279 }
2280
2281 #ifndef PRODUCT
2282 if (_cfg->C->trace_opto_output())
2283 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2284 #endif
2285
2286 return true;
2287 }
2288
2289 Node * Scheduling::ChooseNodeToBundle() {
2290 uint siz = _available.size();
2291
2292 if (siz == 0) {
2293
2294 #ifndef PRODUCT
2295 if (_cfg->C->trace_opto_output())
2296 tty->print("# ChooseNodeToBundle: null\n");
2297 #endif
2298 return (nullptr);
2299 }
2300
2301 // Fast path, if only 1 instruction in the bundle
2302 if (siz == 1) {
2303 #ifndef PRODUCT
2304 if (_cfg->C->trace_opto_output()) {
2305 tty->print("# ChooseNodeToBundle (only 1): ");
2306 _available[0]->dump();
2307 }
2308 #endif
2309 return (_available[0]);
2310 }
2311
2312 // Don't bother, if the bundle is already full
2313 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2314 for ( uint i = 0; i < siz; i++ ) {
2315 Node *n = _available[i];
2316
2317 // Skip projections, we'll handle them another way
2318 if (n->is_Proj())
2319 continue;
2320
2321 // This presupposed that instructions are inserted into the
2322 // available list in a legality order; i.e. instructions that
2323 // must be inserted first are at the head of the list
2324 if (NodeFitsInBundle(n)) {
2325 #ifndef PRODUCT
2326 if (_cfg->C->trace_opto_output()) {
2327 tty->print("# ChooseNodeToBundle: ");
2328 n->dump();
2329 }
2330 #endif
2331 return (n);
2332 }
2333 }
2334 }
2335
2336 // Nothing fits in this bundle, choose the highest priority
2337 #ifndef PRODUCT
2338 if (_cfg->C->trace_opto_output()) {
2339 tty->print("# ChooseNodeToBundle: ");
2340 _available[0]->dump();
2341 }
2342 #endif
2343
2344 return _available[0];
2345 }
2346
2347 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2348 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2349
2350 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2351 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2352 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2353 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2354
2355 // Comparison between stack -> reg and stack -> reg
2356 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2357 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2358 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2359 }
2360
2361 // Comparison between reg -> stack and reg -> stack
2362 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2363 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2364 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2365 }
2366
2367 return 0; // Not comparable
2368 }
2369
2370 void Scheduling::AddNodeToAvailableList(Node *n) {
2371 assert( !n->is_Proj(), "projections never directly made available" );
2372 #ifndef PRODUCT
2373 if (_cfg->C->trace_opto_output()) {
2374 tty->print("# AddNodeToAvailableList: ");
2375 n->dump();
2376 }
2377 #endif
2378
2379 int latency = _current_latency[n->_idx];
2380
2381 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2382 // for stack spilling or unspilling have the same latency, we sort
2383 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2384 // have more opportunities to do ld/st merging
2385 uint i;
2386 for (i = 0; i < _available.size(); i++) {
2387 if (_current_latency[_available[i]->_idx] > latency) {
2388 break;
2389 } else if (_current_latency[_available[i]->_idx] == latency &&
2390 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2391 compare_two_spill_nodes(n, _available[i]) > 0) {
2392 break;
2393 }
2394 }
2395
2396 // Special Check for compares following branches
2397 if( n->is_Mach() && _scheduled.size() > 0 ) {
2398 int op = n->as_Mach()->ideal_Opcode();
2399 Node *last = _scheduled[0];
2400 if( last->is_MachIf() && last->in(1) == n &&
2401 ( op == Op_CmpI ||
2402 op == Op_CmpU ||
2403 op == Op_CmpUL ||
2404 op == Op_CmpP ||
2405 op == Op_CmpF ||
2406 op == Op_CmpD ||
2407 op == Op_CmpL ) ) {
2408
2409 // Recalculate position, moving to front of same latency
2410 for ( i=0 ; i < _available.size(); i++ )
2411 if (_current_latency[_available[i]->_idx] >= latency)
2412 break;
2413 }
2414 }
2415
2416 // Insert the node in the available list
2417 _available.insert(i, n);
2418
2419 #ifndef PRODUCT
2420 if (_cfg->C->trace_opto_output())
2421 dump_available();
2422 #endif
2423 }
2424
2425 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2426 for ( uint i=0; i < n->len(); i++ ) {
2427 Node *def = n->in(i);
2428 if (!def) continue;
2429 if( def->is_Proj() ) // If this is a machine projection, then
2430 def = def->in(0); // propagate usage thru to the base instruction
2431
2432 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2433 continue;
2434 }
2435
2436 // Compute the latency
2437 uint l = _bundle_cycle_number + n->latency(i);
2438 if (_current_latency[def->_idx] < l)
2439 _current_latency[def->_idx] = l;
2440
2441 // If this does not have uses then schedule it
2442 if ((--_uses[def->_idx]) == 0)
2443 AddNodeToAvailableList(def);
2444 }
2445 }
2446
2447 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2448 #ifndef PRODUCT
2449 if (_cfg->C->trace_opto_output()) {
2450 tty->print("# AddNodeToBundle: ");
2451 n->dump();
2452 }
2453 #endif
2454
2455 // Remove this from the available list
2456 uint i;
2457 for (i = 0; i < _available.size(); i++)
2458 if (_available[i] == n)
2459 break;
2460 assert(i < _available.size(), "entry in _available list not found");
2461 _available.remove(i);
2462
2463 // See if this fits in the current bundle
2464 const Pipeline *node_pipeline = n->pipeline();
2465 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2466
2467 // Check for instructions to be placed in the delay slot. We
2468 // do this before we actually schedule the current instruction,
2469 // because the delay slot follows the current instruction.
2470 if (Pipeline::_branch_has_delay_slot &&
2471 node_pipeline->hasBranchDelay() &&
2472 !_unconditional_delay_slot) {
2473
2474 uint siz = _available.size();
2475
2476 // Conditional branches can support an instruction that
2477 // is unconditionally executed and not dependent by the
2478 // branch, OR a conditionally executed instruction if
2479 // the branch is taken. In practice, this means that
2480 // the first instruction at the branch target is
2481 // copied to the delay slot, and the branch goes to
2482 // the instruction after that at the branch target
2483 if ( n->is_MachBranch() ) {
2484
2485 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2486 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2487
2488 #ifndef PRODUCT
2489 _branches++;
2490 #endif
2491
2492 // At least 1 instruction is on the available list
2493 // that is not dependent on the branch
2494 for (uint i = 0; i < siz; i++) {
2495 Node *d = _available[i];
2496 const Pipeline *avail_pipeline = d->pipeline();
2497
2498 // Don't allow safepoints in the branch shadow, that will
2499 // cause a number of difficulties
2500 if ( avail_pipeline->instructionCount() == 1 &&
2501 !avail_pipeline->hasMultipleBundles() &&
2502 !avail_pipeline->hasBranchDelay() &&
2503 Pipeline::instr_has_unit_size() &&
2504 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2505 NodeFitsInBundle(d) &&
2506 !node_bundling(d)->used_in_delay()) {
2507
2508 if (d->is_Mach() && !d->is_MachSafePoint()) {
2509 // A node that fits in the delay slot was found, so we need to
2510 // set the appropriate bits in the bundle pipeline information so
2511 // that it correctly indicates resource usage. Later, when we
2512 // attempt to add this instruction to the bundle, we will skip
2513 // setting the resource usage.
2514 _unconditional_delay_slot = d;
2515 node_bundling(n)->set_use_unconditional_delay();
2516 node_bundling(d)->set_used_in_unconditional_delay();
2517 _bundle_use.add_usage(avail_pipeline->resourceUse());
2518 _current_latency[d->_idx] = _bundle_cycle_number;
2519 _next_node = d;
2520 ++_bundle_instr_count;
2521 #ifndef PRODUCT
2522 _unconditional_delays++;
2523 #endif
2524 break;
2525 }
2526 }
2527 }
2528 }
2529
2530 // No delay slot, add a nop to the usage
2531 if (!_unconditional_delay_slot) {
2532 // See if adding an instruction in the delay slot will overflow
2533 // the bundle.
2534 if (!NodeFitsInBundle(_nop)) {
2535 #ifndef PRODUCT
2536 if (_cfg->C->trace_opto_output())
2537 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2538 #endif
2539 step(1);
2540 }
2541
2542 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2543 _next_node = _nop;
2544 ++_bundle_instr_count;
2545 }
2546
2547 // See if the instruction in the delay slot requires a
2548 // step of the bundles
2549 if (!NodeFitsInBundle(n)) {
2550 #ifndef PRODUCT
2551 if (_cfg->C->trace_opto_output())
2552 tty->print("# *** STEP(branch won't fit) ***\n");
2553 #endif
2554 // Update the state information
2555 _bundle_instr_count = 0;
2556 _bundle_cycle_number += 1;
2557 _bundle_use.step(1);
2558 }
2559 }
2560
2561 // Get the number of instructions
2562 uint instruction_count = node_pipeline->instructionCount();
2563 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2564 instruction_count = 0;
2565
2566 // Compute the latency information
2567 uint delay = 0;
2568
2569 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2570 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2571 if (relative_latency < 0)
2572 relative_latency = 0;
2573
2574 delay = _bundle_use.full_latency(relative_latency, node_usage);
2575
2576 // Does not fit in this bundle, start a new one
2577 if (delay > 0) {
2578 step(delay);
2579
2580 #ifndef PRODUCT
2581 if (_cfg->C->trace_opto_output())
2582 tty->print("# *** STEP(%d) ***\n", delay);
2583 #endif
2584 }
2585 }
2586
2587 // If this was placed in the delay slot, ignore it
2588 if (n != _unconditional_delay_slot) {
2589
2590 if (delay == 0) {
2591 if (node_pipeline->hasMultipleBundles()) {
2592 #ifndef PRODUCT
2593 if (_cfg->C->trace_opto_output())
2594 tty->print("# *** STEP(multiple instructions) ***\n");
2595 #endif
2596 step(1);
2597 }
2598
2599 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2600 #ifndef PRODUCT
2601 if (_cfg->C->trace_opto_output())
2602 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2603 instruction_count + _bundle_instr_count,
2604 Pipeline::_max_instrs_per_cycle);
2605 #endif
2606 step(1);
2607 }
2608 }
2609
2610 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2611 _bundle_instr_count++;
2612
2613 // Set the node's latency
2614 _current_latency[n->_idx] = _bundle_cycle_number;
2615
2616 // Now merge the functional unit information
2617 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2618 _bundle_use.add_usage(node_usage);
2619
2620 // Increment the number of instructions in this bundle
2621 _bundle_instr_count += instruction_count;
2622
2623 // Remember this node for later
2624 if (n->is_Mach())
2625 _next_node = n;
2626 }
2627
2628 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2629 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2630 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2631 // into the block. All other scheduled nodes get put in the schedule here.
2632 int op = n->Opcode();
2633 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2634 (op != Op_Node && // Not an unused antidepedence node and
2635 // not an unallocated boxlock
2636 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2637
2638 // Push any trailing projections
2639 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2640 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2641 Node *foi = n->fast_out(i);
2642 if( foi->is_Proj() )
2643 _scheduled.push(foi);
2644 }
2645 }
2646
2647 // Put the instruction in the schedule list
2648 _scheduled.push(n);
2649 }
2650
2651 #ifndef PRODUCT
2652 if (_cfg->C->trace_opto_output())
2653 dump_available();
2654 #endif
2655
2656 // Walk all the definitions, decrementing use counts, and
2657 // if a definition has a 0 use count, place it in the available list.
2658 DecrementUseCounts(n,bb);
2659 }
2660
2661 // This method sets the use count within a basic block. We will ignore all
2662 // uses outside the current basic block. As we are doing a backwards walk,
2663 // any node we reach that has a use count of 0 may be scheduled. This also
2664 // avoids the problem of cyclic references from phi nodes, as long as phi
2665 // nodes are at the front of the basic block. This method also initializes
2666 // the available list to the set of instructions that have no uses within this
2667 // basic block.
2668 void Scheduling::ComputeUseCount(const Block *bb) {
2669 #ifndef PRODUCT
2670 if (_cfg->C->trace_opto_output())
2671 tty->print("# -> ComputeUseCount\n");
2672 #endif
2673
2674 // Clear the list of available and scheduled instructions, just in case
2675 _available.clear();
2676 _scheduled.clear();
2677
2678 // No delay slot specified
2679 _unconditional_delay_slot = nullptr;
2680
2681 #ifdef ASSERT
2682 for( uint i=0; i < bb->number_of_nodes(); i++ )
2683 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2684 #endif
2685
2686 // Force the _uses count to never go to zero for unscheduable pieces
2687 // of the block
2688 for( uint k = 0; k < _bb_start; k++ )
2689 _uses[bb->get_node(k)->_idx] = 1;
2690 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2691 _uses[bb->get_node(l)->_idx] = 1;
2692
2693 // Iterate backwards over the instructions in the block. Don't count the
2694 // branch projections at end or the block header instructions.
2695 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2696 Node *n = bb->get_node(j);
2697 if( n->is_Proj() ) continue; // Projections handled another way
2698
2699 // Account for all uses
2700 for ( uint k = 0; k < n->len(); k++ ) {
2701 Node *inp = n->in(k);
2702 if (!inp) continue;
2703 assert(inp != n, "no cycles allowed" );
2704 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2705 if (inp->is_Proj()) { // Skip through Proj's
2706 inp = inp->in(0);
2707 }
2708 ++_uses[inp->_idx]; // Count 1 block-local use
2709 }
2710 }
2711
2712 // If this instruction has a 0 use count, then it is available
2713 if (!_uses[n->_idx]) {
2714 _current_latency[n->_idx] = _bundle_cycle_number;
2715 AddNodeToAvailableList(n);
2716 }
2717
2718 #ifndef PRODUCT
2719 if (_cfg->C->trace_opto_output()) {
2720 tty->print("# uses: %3d: ", _uses[n->_idx]);
2721 n->dump();
2722 }
2723 #endif
2724 }
2725
2726 #ifndef PRODUCT
2727 if (_cfg->C->trace_opto_output())
2728 tty->print("# <- ComputeUseCount\n");
2729 #endif
2730 }
2731
2732 // This routine performs scheduling on each basic block in reverse order,
2733 // using instruction latencies and taking into account function unit
2734 // availability.
2735 void Scheduling::DoScheduling() {
2736 #ifndef PRODUCT
2737 if (_cfg->C->trace_opto_output())
2738 tty->print("# -> DoScheduling\n");
2739 #endif
2740
2741 Block *succ_bb = nullptr;
2742 Block *bb;
2743 Compile* C = Compile::current();
2744
2745 // Walk over all the basic blocks in reverse order
2746 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2747 bb = _cfg->get_block(i);
2748
2749 #ifndef PRODUCT
2750 if (_cfg->C->trace_opto_output()) {
2751 tty->print("# Schedule BB#%03d (initial)\n", i);
2752 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2753 bb->get_node(j)->dump();
2754 }
2755 }
2756 #endif
2757
2758 // On the head node, skip processing
2759 if (bb == _cfg->get_root_block()) {
2760 continue;
2761 }
2762
2763 // Skip empty, connector blocks
2764 if (bb->is_connector())
2765 continue;
2766
2767 // If the following block is not the sole successor of
2768 // this one, then reset the pipeline information
2769 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2770 #ifndef PRODUCT
2771 if (_cfg->C->trace_opto_output()) {
2772 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2773 _next_node->_idx, _bundle_instr_count);
2774 }
2775 #endif
2776 step_and_clear();
2777 }
2778
2779 // Leave untouched the starting instruction, any Phis, a CreateEx node
2780 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2781 _bb_end = bb->number_of_nodes()-1;
2782 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2783 Node *n = bb->get_node(_bb_start);
2784 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2785 // Also, MachIdealNodes do not get scheduled
2786 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2787 MachNode *mach = n->as_Mach();
2788 int iop = mach->ideal_Opcode();
2789 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2790 if( iop == Op_Con ) continue; // Do not schedule Top
2791 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2792 mach->pipeline() == MachNode::pipeline_class() &&
2793 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2794 continue;
2795 break; // Funny loop structure to be sure...
2796 }
2797 // Compute last "interesting" instruction in block - last instruction we
2798 // might schedule. _bb_end points just after last schedulable inst. We
2799 // normally schedule conditional branches (despite them being forced last
2800 // in the block), because they have delay slots we can fill. Calls all
2801 // have their delay slots filled in the template expansions, so we don't
2802 // bother scheduling them.
2803 Node *last = bb->get_node(_bb_end);
2804 // Ignore trailing NOPs.
2805 while (_bb_end > 0 && last->is_Mach() &&
2806 last->as_Mach()->ideal_Opcode() == Op_Con) {
2807 last = bb->get_node(--_bb_end);
2808 }
2809 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2810 if( last->is_Catch() ||
2811 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2812 // There might be a prior call. Skip it.
2813 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2814 } else if( last->is_MachNullCheck() ) {
2815 // Backup so the last null-checked memory instruction is
2816 // outside the schedulable range. Skip over the nullcheck,
2817 // projection, and the memory nodes.
2818 Node *mem = last->in(1);
2819 do {
2820 _bb_end--;
2821 } while (mem != bb->get_node(_bb_end));
2822 } else {
2823 // Set _bb_end to point after last schedulable inst.
2824 _bb_end++;
2825 }
2826
2827 assert( _bb_start <= _bb_end, "inverted block ends" );
2828
2829 // Compute the register antidependencies for the basic block
2830 ComputeRegisterAntidependencies(bb);
2831 if (C->failing()) return; // too many D-U pinch points
2832
2833 // Compute the usage within the block, and set the list of all nodes
2834 // in the block that have no uses within the block.
2835 ComputeUseCount(bb);
2836
2837 // Schedule the remaining instructions in the block
2838 while ( _available.size() > 0 ) {
2839 Node *n = ChooseNodeToBundle();
2840 guarantee(n != nullptr, "no nodes available");
2841 AddNodeToBundle(n,bb);
2842 }
2843
2844 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2845 #ifdef ASSERT
2846 for( uint l = _bb_start; l < _bb_end; l++ ) {
2847 Node *n = bb->get_node(l);
2848 uint m;
2849 for( m = 0; m < _bb_end-_bb_start; m++ )
2850 if( _scheduled[m] == n )
2851 break;
2852 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2853 }
2854 #endif
2855
2856 // Now copy the instructions (in reverse order) back to the block
2857 for ( uint k = _bb_start; k < _bb_end; k++ )
2858 bb->map_node(_scheduled[_bb_end-k-1], k);
2859
2860 #ifndef PRODUCT
2861 if (_cfg->C->trace_opto_output()) {
2862 tty->print("# Schedule BB#%03d (final)\n", i);
2863 uint current = 0;
2864 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2865 Node *n = bb->get_node(j);
2866 if( valid_bundle_info(n) ) {
2867 Bundle *bundle = node_bundling(n);
2868 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2869 tty->print("*** Bundle: ");
2870 bundle->dump();
2871 }
2872 n->dump();
2873 }
2874 }
2875 }
2876 #endif
2877 #ifdef ASSERT
2878 verify_good_schedule(bb,"after block local scheduling");
2879 #endif
2880 }
2881
2882 #ifndef PRODUCT
2883 if (_cfg->C->trace_opto_output())
2884 tty->print("# <- DoScheduling\n");
2885 #endif
2886
2887 // Record final node-bundling array location
2888 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2889
2890 } // end DoScheduling
2891
2892 // Verify that no live-range used in the block is killed in the block by a
2893 // wrong DEF. This doesn't verify live-ranges that span blocks.
2894
2895 // Check for edge existence. Used to avoid adding redundant precedence edges.
2896 static bool edge_from_to( Node *from, Node *to ) {
2897 for( uint i=0; i<from->len(); i++ )
2898 if( from->in(i) == to )
2899 return true;
2900 return false;
2901 }
2902
2903 #ifdef ASSERT
2904 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2905 // Check for bad kills
2906 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2907 Node *prior_use = _reg_node[def];
2908 if( prior_use && !edge_from_to(prior_use,n) ) {
2909 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2910 n->dump();
2911 tty->print_cr("...");
2912 prior_use->dump();
2913 assert(edge_from_to(prior_use,n), "%s", msg);
2914 }
2915 _reg_node.map(def,nullptr); // Kill live USEs
2916 }
2917 }
2918
2919 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2920
2921 // Zap to something reasonable for the verify code
2922 _reg_node.clear();
2923
2924 // Walk over the block backwards. Check to make sure each DEF doesn't
2925 // kill a live value (other than the one it's supposed to). Add each
2926 // USE to the live set.
2927 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2928 Node *n = b->get_node(i);
2929 int n_op = n->Opcode();
2930 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2931 // Fat-proj kills a slew of registers
2932 RegMaskIterator rmi(n->out_RegMask());
2933 while (rmi.has_next()) {
2934 OptoReg::Name kill = rmi.next();
2935 verify_do_def(n, kill, msg);
2936 }
2937 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2938 // Get DEF'd registers the normal way
2939 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2940 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2941 }
2942
2943 // Now make all USEs live
2944 for( uint i=1; i<n->req(); i++ ) {
2945 Node *def = n->in(i);
2946 assert(def != 0, "input edge required");
2947 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2948 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2949 if( OptoReg::is_valid(reg_lo) ) {
2950 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2951 _reg_node.map(reg_lo,n);
2952 }
2953 if( OptoReg::is_valid(reg_hi) ) {
2954 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2955 _reg_node.map(reg_hi,n);
2956 }
2957 }
2958
2959 }
2960
2961 // Zap to something reasonable for the Antidependence code
2962 _reg_node.clear();
2963 }
2964 #endif
2965
2966 // Conditionally add precedence edges. Avoid putting edges on Projs.
2967 static void add_prec_edge_from_to( Node *from, Node *to ) {
2968 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2969 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2970 from = from->in(0);
2971 }
2972 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2973 !edge_from_to( from, to ) ) // Avoid duplicate edge
2974 from->add_prec(to);
2975 }
2976
2977 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2978 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2979 return;
2980
2981 if (OptoReg::is_reg(def_reg)) {
2982 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2983 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2984 // This is one of the high slots of a vector register.
2985 // ScheduleAndBundle already checked there are no live wide
2986 // vectors in this method so it can be safely ignored.
2987 return;
2988 }
2989 }
2990
2991 Node *pinch = _reg_node[def_reg]; // Get pinch point
2992 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2993 is_def ) { // Check for a true def (not a kill)
2994 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2995 return;
2996 }
2997
2998 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2999 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
3000
3001 // After some number of kills there _may_ be a later def
3002 Node *later_def = nullptr;
3003
3004 Compile* C = Compile::current();
3005
3006 // Finding a kill requires a real pinch-point.
3007 // Check for not already having a pinch-point.
3008 // Pinch points are Op_Node's.
3009 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
3010 later_def = pinch; // Must be def/kill as optimistic pinch-point
3011 if ( _pinch_free_list.size() > 0) {
3012 pinch = _pinch_free_list.pop();
3013 } else {
3014 pinch = new Node(1); // Pinch point to-be
3015 }
3016 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
3017 DEBUG_ONLY( pinch->dump(); );
3018 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
3019 _cfg->C->record_method_not_compilable("too many D-U pinch points");
3020 return;
3021 }
3022 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3023 _reg_node.map(def_reg,pinch); // Record pinch-point
3024 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3025 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3026 pinch->init_req(0, C->top()); // set not null for the next call
3027 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3028 later_def = nullptr; // and no later def
3029 }
3030 pinch->set_req(0,later_def); // Hook later def so we can find it
3031 } else { // Else have valid pinch point
3032 if( pinch->in(0) ) // If there is a later-def
3033 later_def = pinch->in(0); // Get it
3034 }
3035
3036 // Add output-dependence edge from later def to kill
3037 if( later_def ) // If there is some original def
3038 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3039
3040 // See if current kill is also a use, and so is forced to be the pinch-point.
3041 if( pinch->Opcode() == Op_Node ) {
3042 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3043 for( uint i=1; i<uses->req(); i++ ) {
3044 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3045 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3046 // Yes, found a use/kill pinch-point
3047 pinch->set_req(0,nullptr); //
3048 pinch->replace_by(kill); // Move anti-dep edges up
3049 pinch = kill;
3050 _reg_node.map(def_reg,pinch);
3051 return;
3052 }
3053 }
3054 }
3055
3056 // Add edge from kill to pinch-point
3057 add_prec_edge_from_to(kill,pinch);
3058 }
3059
3060 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3061 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3062 return;
3063 Node *pinch = _reg_node[use_reg]; // Get pinch point
3064 // Check for no later def_reg/kill in block
3065 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3066 // Use has to be block-local as well
3067 _cfg->get_block_for_node(use) == b) {
3068 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3069 pinch->req() == 1 ) { // pinch not yet in block?
3070 pinch->del_req(0); // yank pointer to later-def, also set flag
3071 // Insert the pinch-point in the block just after the last use
3072 b->insert_node(pinch, b->find_node(use) + 1);
3073 _bb_end++; // Increase size scheduled region in block
3074 }
3075
3076 add_prec_edge_from_to(pinch,use);
3077 }
3078 }
3079
3080 // We insert antidependences between the reads and following write of
3081 // allocated registers to prevent illegal code motion. Hopefully, the
3082 // number of added references should be fairly small, especially as we
3083 // are only adding references within the current basic block.
3084 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3085
3086 #ifdef ASSERT
3087 verify_good_schedule(b,"before block local scheduling");
3088 #endif
3089
3090 // A valid schedule, for each register independently, is an endless cycle
3091 // of: a def, then some uses (connected to the def by true dependencies),
3092 // then some kills (defs with no uses), finally the cycle repeats with a new
3093 // def. The uses are allowed to float relative to each other, as are the
3094 // kills. No use is allowed to slide past a kill (or def). This requires
3095 // antidependencies between all uses of a single def and all kills that
3096 // follow, up to the next def. More edges are redundant, because later defs
3097 // & kills are already serialized with true or antidependencies. To keep
3098 // the edge count down, we add a 'pinch point' node if there's more than
3099 // one use or more than one kill/def.
3100
3101 // We add dependencies in one bottom-up pass.
3102
3103 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3104
3105 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3106 // register. If not, we record the DEF/KILL in _reg_node, the
3107 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3108 // "pinch point", a new Node that's in the graph but not in the block.
3109 // We put edges from the prior and current DEF/KILLs to the pinch point.
3110 // We put the pinch point in _reg_node. If there's already a pinch point
3111 // we merely add an edge from the current DEF/KILL to the pinch point.
3112
3113 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3114 // put an edge from the pinch point to the USE.
3115
3116 // To be expedient, the _reg_node array is pre-allocated for the whole
3117 // compilation. _reg_node is lazily initialized; it either contains a null,
3118 // or a valid def/kill/pinch-point, or a leftover node from some prior
3119 // block. Leftover node from some prior block is treated like a null (no
3120 // prior def, so no anti-dependence needed). Valid def is distinguished by
3121 // it being in the current block.
3122 bool fat_proj_seen = false;
3123 uint last_safept = _bb_end-1;
3124 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3125 Node* last_safept_node = end_node;
3126 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3127 Node *n = b->get_node(i);
3128 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3129 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3130 // Fat-proj kills a slew of registers
3131 // This can add edges to 'n' and obscure whether or not it was a def,
3132 // hence the is_def flag.
3133 fat_proj_seen = true;
3134 RegMaskIterator rmi(n->out_RegMask());
3135 while (rmi.has_next()) {
3136 OptoReg::Name kill = rmi.next();
3137 anti_do_def(b, n, kill, is_def);
3138 }
3139 } else {
3140 // Get DEF'd registers the normal way
3141 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3142 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3143 }
3144
3145 // Kill projections on a branch should appear to occur on the
3146 // branch, not afterwards, so grab the masks from the projections
3147 // and process them.
3148 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3149 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3150 Node* use = n->fast_out(i);
3151 if (use->is_Proj()) {
3152 RegMaskIterator rmi(use->out_RegMask());
3153 while (rmi.has_next()) {
3154 OptoReg::Name kill = rmi.next();
3155 anti_do_def(b, n, kill, false);
3156 }
3157 }
3158 }
3159 }
3160
3161 // Check each register used by this instruction for a following DEF/KILL
3162 // that must occur afterward and requires an anti-dependence edge.
3163 for( uint j=0; j<n->req(); j++ ) {
3164 Node *def = n->in(j);
3165 if( def ) {
3166 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3167 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3168 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3169 }
3170 }
3171 // Do not allow defs of new derived values to float above GC
3172 // points unless the base is definitely available at the GC point.
3173
3174 Node *m = b->get_node(i);
3175
3176 // Add precedence edge from following safepoint to use of derived pointer
3177 if( last_safept_node != end_node &&
3178 m != last_safept_node) {
3179 for (uint k = 1; k < m->req(); k++) {
3180 const Type *t = m->in(k)->bottom_type();
3181 if( t->isa_oop_ptr() &&
3182 t->is_ptr()->offset() != 0 ) {
3183 last_safept_node->add_prec( m );
3184 break;
3185 }
3186 }
3187
3188 // Do not allow a CheckCastPP node whose input is a raw pointer to
3189 // float past a safepoint. This can occur when a buffered inline
3190 // type is allocated in a loop and the CheckCastPP from that
3191 // allocation is reused outside the loop. If the use inside the
3192 // loop is scalarized the CheckCastPP will no longer be connected
3193 // to the loop safepoint. See JDK-8264340.
3194 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3195 Node *def = m->in(1);
3196 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3197 last_safept_node->add_prec(m);
3198 }
3199 }
3200 }
3201
3202 if( n->jvms() ) { // Precedence edge from derived to safept
3203 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3204 if( b->get_node(last_safept) != last_safept_node ) {
3205 last_safept = b->find_node(last_safept_node);
3206 }
3207 for( uint j=last_safept; j > i; j-- ) {
3208 Node *mach = b->get_node(j);
3209 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3210 mach->add_prec( n );
3211 }
3212 last_safept = i;
3213 last_safept_node = m;
3214 }
3215 }
3216
3217 if (fat_proj_seen) {
3218 // Garbage collect pinch nodes that were not consumed.
3219 // They are usually created by a fat kill MachProj for a call.
3220 garbage_collect_pinch_nodes();
3221 }
3222 }
3223
3224 // Garbage collect pinch nodes for reuse by other blocks.
3225 //
3226 // The block scheduler's insertion of anti-dependence
3227 // edges creates many pinch nodes when the block contains
3228 // 2 or more Calls. A pinch node is used to prevent a
3229 // combinatorial explosion of edges. If a set of kills for a
3230 // register is anti-dependent on a set of uses (or defs), rather
3231 // than adding an edge in the graph between each pair of kill
3232 // and use (or def), a pinch is inserted between them:
3233 //
3234 // use1 use2 use3
3235 // \ | /
3236 // \ | /
3237 // pinch
3238 // / | \
3239 // / | \
3240 // kill1 kill2 kill3
3241 //
3242 // One pinch node is created per register killed when
3243 // the second call is encountered during a backwards pass
3244 // over the block. Most of these pinch nodes are never
3245 // wired into the graph because the register is never
3246 // used or def'ed in the block.
3247 //
3248 void Scheduling::garbage_collect_pinch_nodes() {
3249 #ifndef PRODUCT
3250 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3251 #endif
3252 int trace_cnt = 0;
3253 for (uint k = 0; k < _reg_node.max(); k++) {
3254 Node* pinch = _reg_node[k];
3255 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3256 // no predecence input edges
3257 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3258 cleanup_pinch(pinch);
3259 _pinch_free_list.push(pinch);
3260 _reg_node.map(k, nullptr);
3261 #ifndef PRODUCT
3262 if (_cfg->C->trace_opto_output()) {
3263 trace_cnt++;
3264 if (trace_cnt > 40) {
3265 tty->print("\n");
3266 trace_cnt = 0;
3267 }
3268 tty->print(" %d", pinch->_idx);
3269 }
3270 #endif
3271 }
3272 }
3273 #ifndef PRODUCT
3274 if (_cfg->C->trace_opto_output()) tty->print("\n");
3275 #endif
3276 }
3277
3278 // Clean up a pinch node for reuse.
3279 void Scheduling::cleanup_pinch( Node *pinch ) {
3280 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3281
3282 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3283 Node* use = pinch->last_out(i);
3284 uint uses_found = 0;
3285 for (uint j = use->req(); j < use->len(); j++) {
3286 if (use->in(j) == pinch) {
3287 use->rm_prec(j);
3288 uses_found++;
3289 }
3290 }
3291 assert(uses_found > 0, "must be a precedence edge");
3292 i -= uses_found; // we deleted 1 or more copies of this edge
3293 }
3294 // May have a later_def entry
3295 pinch->set_req(0, nullptr);
3296 }
3297
3298 #ifndef PRODUCT
3299
3300 void Scheduling::dump_available() const {
3301 tty->print("#Availist ");
3302 for (uint i = 0; i < _available.size(); i++)
3303 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3304 tty->cr();
3305 }
3306
3307 // Print Scheduling Statistics
3308 void Scheduling::print_statistics() {
3309 // Print the size added by nops for bundling
3310 tty->print("Nops added %d bytes to total of %d bytes",
3311 _total_nop_size, _total_method_size);
3312 if (_total_method_size > 0)
3313 tty->print(", for %.2f%%",
3314 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3315 tty->print("\n");
3316
3317 // Print the number of branch shadows filled
3318 if (Pipeline::_branch_has_delay_slot) {
3319 tty->print("Of %d branches, %d had unconditional delay slots filled",
3320 _total_branches, _total_unconditional_delays);
3321 if (_total_branches > 0)
3322 tty->print(", for %.2f%%",
3323 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3324 tty->print("\n");
3325 }
3326
3327 uint total_instructions = 0, total_bundles = 0;
3328
3329 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3330 uint bundle_count = _total_instructions_per_bundle[i];
3331 total_instructions += bundle_count * i;
3332 total_bundles += bundle_count;
3333 }
3334
3335 if (total_bundles > 0)
3336 tty->print("Average ILP (excluding nops) is %.2f\n",
3337 ((double)total_instructions) / ((double)total_bundles));
3338 }
3339 #endif
3340
3341 //-----------------------init_scratch_buffer_blob------------------------------
3342 // Construct a temporary BufferBlob and cache it for this compile.
3343 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3344 // If there is already a scratch buffer blob allocated and the
3345 // constant section is big enough, use it. Otherwise free the
3346 // current and allocate a new one.
3347 BufferBlob* blob = scratch_buffer_blob();
3348 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3349 // Use the current blob.
3350 } else {
3351 if (blob != nullptr) {
3352 BufferBlob::free(blob);
3353 }
3354
3355 ResourceMark rm;
3356 _scratch_const_size = const_size;
3357 int size = C2Compiler::initial_code_buffer_size(const_size);
3358 if (C->has_scalarized_args()) {
3359 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3360 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3361 ciMethod* method = C->method();
3362 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3363 int arg_num = 0;
3364 if (!method->is_static()) {
3365 if (method->is_scalarized_arg(arg_num)) {
3366 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3367 }
3368 arg_num++;
3369 }
3370 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3371 if (method->is_scalarized_arg(arg_num)) {
3372 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3373 }
3374 arg_num++;
3375 }
3376 }
3377 blob = BufferBlob::create("Compile::scratch_buffer", size);
3378 // Record the buffer blob for next time.
3379 set_scratch_buffer_blob(blob);
3380 // Have we run out of code space?
3381 if (scratch_buffer_blob() == nullptr) {
3382 // Let CompilerBroker disable further compilations.
3383 C->record_failure("Not enough space for scratch buffer in CodeCache");
3384 return;
3385 }
3386 }
3387
3388 // Initialize the relocation buffers
3389 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3390 set_scratch_locs_memory(locs_buf);
3391 }
3392
3393
3394 //-----------------------scratch_emit_size-------------------------------------
3395 // Helper function that computes size by emitting code
3396 uint PhaseOutput::scratch_emit_size(const Node* n) {
3397 // Start scratch_emit_size section.
3398 set_in_scratch_emit_size(true);
3399
3400 // Emit into a trash buffer and count bytes emitted.
3401 // This is a pretty expensive way to compute a size,
3402 // but it works well enough if seldom used.
3403 // All common fixed-size instructions are given a size
3404 // method by the AD file.
3405 // Note that the scratch buffer blob and locs memory are
3406 // allocated at the beginning of the compile task, and
3407 // may be shared by several calls to scratch_emit_size.
3408 // The allocation of the scratch buffer blob is particularly
3409 // expensive, since it has to grab the code cache lock.
3410 BufferBlob* blob = this->scratch_buffer_blob();
3411 assert(blob != nullptr, "Initialize BufferBlob at start");
3412 assert(blob->size() > MAX_inst_size, "sanity");
3413 relocInfo* locs_buf = scratch_locs_memory();
3414 address blob_begin = blob->content_begin();
3415 address blob_end = (address)locs_buf;
3416 assert(blob->contains(blob_end), "sanity");
3417 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3418 buf.initialize_consts_size(_scratch_const_size);
3419 buf.initialize_stubs_size(MAX_stubs_size);
3420 assert(locs_buf != nullptr, "sanity");
3421 int lsize = MAX_locs_size / 3;
3422 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3423 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3424 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3425 // Mark as scratch buffer.
3426 buf.consts()->set_scratch_emit();
3427 buf.insts()->set_scratch_emit();
3428 buf.stubs()->set_scratch_emit();
3429
3430 // Do the emission.
3431
3432 Label fakeL; // Fake label for branch instructions.
3433 Label* saveL = nullptr;
3434 uint save_bnum = 0;
3435 bool is_branch = n->is_MachBranch();
3436 if (is_branch) {
3437 MacroAssembler masm(&buf);
3438 masm.bind(fakeL);
3439 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3440 n->as_MachBranch()->label_set(&fakeL, 0);
3441 }
3442 n->emit(buf, C->regalloc());
3443
3444 // Emitting into the scratch buffer should not fail
3445 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3446
3447 // Restore label.
3448 if (is_branch) {
3449 n->as_MachBranch()->label_set(saveL, save_bnum);
3450 }
3451
3452 // End scratch_emit_size section.
3453 set_in_scratch_emit_size(false);
3454
3455 return buf.insts_size();
3456 }
3457
3458 void PhaseOutput::install() {
3459 if (!C->should_install_code()) {
3460 return;
3461 } else if (C->stub_function() != nullptr) {
3462 install_stub(C->stub_name());
3463 } else {
3464 install_code(C->method(),
3465 C->entry_bci(),
3466 CompileBroker::compiler2(),
3467 C->has_unsafe_access(),
3468 SharedRuntime::is_wide_vector(C->max_vector_size()),
3469 C->rtm_state());
3470 }
3471 }
3472
3473 void PhaseOutput::install_code(ciMethod* target,
3474 int entry_bci,
3475 AbstractCompiler* compiler,
3476 bool has_unsafe_access,
3477 bool has_wide_vectors,
3478 RTMState rtm_state) {
3479 // Check if we want to skip execution of all compiled code.
3480 {
3481 #ifndef PRODUCT
3482 if (OptoNoExecute) {
3483 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3484 return;
3485 }
3486 #endif
3487 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3488
3489 if (C->is_osr_compilation()) {
3490 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3491 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3492 } else {
3493 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3494 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3495 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3496 }
3497 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3498 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3499 }
3500 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3501 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3502 }
3503 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3504 }
3505
3506 C->env()->register_method(target,
3507 entry_bci,
3508 &_code_offsets,
3509 _orig_pc_slot_offset_in_bytes,
3510 code_buffer(),
3511 frame_size_in_words(),
3512 _oop_map_set,
3513 &_handler_table,
3514 inc_table(),
3515 compiler,
3516 has_unsafe_access,
3517 SharedRuntime::is_wide_vector(C->max_vector_size()),
3518 C->has_monitors(),
3519 0,
3520 C->rtm_state());
3521
3522 if (C->log() != nullptr) { // Print code cache state into compiler log
3523 C->log()->code_cache_state();
3524 }
3525 }
3526 }
3527 void PhaseOutput::install_stub(const char* stub_name) {
3528 // Entry point will be accessed using stub_entry_point();
3529 if (code_buffer() == nullptr) {
3530 Matcher::soft_match_failure();
3531 } else {
3532 if (PrintAssembly && (WizardMode || Verbose))
3533 tty->print_cr("### Stub::%s", stub_name);
3534
3535 if (!C->failing()) {
3536 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3537
3538 // Make the NMethod
3539 // For now we mark the frame as never safe for profile stackwalking
3540 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3541 code_buffer(),
3542 CodeOffsets::frame_never_safe,
3543 // _code_offsets.value(CodeOffsets::Frame_Complete),
3544 frame_size_in_words(),
3545 oop_map_set(),
3546 false);
3547 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3548
3549 C->set_stub_entry_point(rs->entry_point());
3550 }
3551 }
3552 }
3553
3554 // Support for bundling info
3555 Bundle* PhaseOutput::node_bundling(const Node *n) {
3556 assert(valid_bundle_info(n), "oob");
3557 return &_node_bundling_base[n->_idx];
3558 }
3559
3560 bool PhaseOutput::valid_bundle_info(const Node *n) {
3561 return (_node_bundling_limit > n->_idx);
3562 }
3563
3564 //------------------------------frame_size_in_words-----------------------------
3565 // frame_slots in units of words
3566 int PhaseOutput::frame_size_in_words() const {
3567 // shift is 0 in LP32 and 1 in LP64
3568 const int shift = (LogBytesPerWord - LogBytesPerInt);
3569 int words = _frame_slots >> shift;
3570 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3571 return words;
3572 }
3573
3574 // To bang the stack of this compiled method we use the stack size
3575 // that the interpreter would need in case of a deoptimization. This
3576 // removes the need to bang the stack in the deoptimization blob which
3577 // in turn simplifies stack overflow handling.
3578 int PhaseOutput::bang_size_in_bytes() const {
3579 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3580 }
3581
3582 //------------------------------dump_asm---------------------------------------
3583 // Dump formatted assembly
3584 #if defined(SUPPORT_OPTO_ASSEMBLY)
3585 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3586
3587 int pc_digits = 3; // #chars required for pc
3588 int sb_chars = 3; // #chars for "start bundle" indicator
3589 int tab_size = 8;
3590 if (pcs != nullptr) {
3591 int max_pc = 0;
3592 for (uint i = 0; i < pc_limit; i++) {
3593 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3594 }
3595 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3596 }
3597 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3598
3599 bool cut_short = false;
3600 st->print_cr("#");
3601 st->print("# "); C->tf()->dump_on(st); st->cr();
3602 st->print_cr("#");
3603
3604 // For all blocks
3605 int pc = 0x0; // Program counter
3606 char starts_bundle = ' ';
3607 C->regalloc()->dump_frame();
3608
3609 Node *n = nullptr;
3610 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3611 if (VMThread::should_terminate()) {
3612 cut_short = true;
3613 break;
3614 }
3615 Block* block = C->cfg()->get_block(i);
3616 if (block->is_connector() && !Verbose) {
3617 continue;
3618 }
3619 n = block->head();
3620 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3621 pc = pcs[n->_idx];
3622 st->print("%*.*x", pc_digits, pc_digits, pc);
3623 }
3624 st->fill_to(prefix_len);
3625 block->dump_head(C->cfg(), st);
3626 if (block->is_connector()) {
3627 st->fill_to(prefix_len);
3628 st->print_cr("# Empty connector block");
3629 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3630 st->fill_to(prefix_len);
3631 st->print_cr("# Block is sole successor of call");
3632 }
3633
3634 // For all instructions
3635 Node *delay = nullptr;
3636 for (uint j = 0; j < block->number_of_nodes(); j++) {
3637 if (VMThread::should_terminate()) {
3638 cut_short = true;
3639 break;
3640 }
3641 n = block->get_node(j);
3642 if (valid_bundle_info(n)) {
3643 Bundle* bundle = node_bundling(n);
3644 if (bundle->used_in_unconditional_delay()) {
3645 delay = n;
3646 continue;
3647 }
3648 if (bundle->starts_bundle()) {
3649 starts_bundle = '+';
3650 }
3651 }
3652
3653 if (WizardMode) {
3654 n->dump();
3655 }
3656
3657 if( !n->is_Region() && // Dont print in the Assembly
3658 !n->is_Phi() && // a few noisely useless nodes
3659 !n->is_Proj() &&
3660 !n->is_MachTemp() &&
3661 !n->is_SafePointScalarObject() &&
3662 !n->is_Catch() && // Would be nice to print exception table targets
3663 !n->is_MergeMem() && // Not very interesting
3664 !n->is_top() && // Debug info table constants
3665 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3666 ) {
3667 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3668 pc = pcs[n->_idx];
3669 st->print("%*.*x", pc_digits, pc_digits, pc);
3670 } else {
3671 st->fill_to(pc_digits);
3672 }
3673 st->print(" %c ", starts_bundle);
3674 starts_bundle = ' ';
3675 st->fill_to(prefix_len);
3676 n->format(C->regalloc(), st);
3677 st->cr();
3678 }
3679
3680 // If we have an instruction with a delay slot, and have seen a delay,
3681 // then back up and print it
3682 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3683 // Coverity finding - Explicit null dereferenced.
3684 guarantee(delay != nullptr, "no unconditional delay instruction");
3685 if (WizardMode) delay->dump();
3686
3687 if (node_bundling(delay)->starts_bundle())
3688 starts_bundle = '+';
3689 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3690 pc = pcs[n->_idx];
3691 st->print("%*.*x", pc_digits, pc_digits, pc);
3692 } else {
3693 st->fill_to(pc_digits);
3694 }
3695 st->print(" %c ", starts_bundle);
3696 starts_bundle = ' ';
3697 st->fill_to(prefix_len);
3698 delay->format(C->regalloc(), st);
3699 st->cr();
3700 delay = nullptr;
3701 }
3702
3703 // Dump the exception table as well
3704 if( n->is_Catch() && (Verbose || WizardMode) ) {
3705 // Print the exception table for this offset
3706 _handler_table.print_subtable_for(pc);
3707 }
3708 st->bol(); // Make sure we start on a new line
3709 }
3710 st->cr(); // one empty line between blocks
3711 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3712 } // End of per-block dump
3713
3714 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3715 }
3716 #endif
3717
3718 #ifndef PRODUCT
3719 void PhaseOutput::print_statistics() {
3720 Scheduling::print_statistics();
3721 }
3722 #endif