1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/debugInfo.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/compilerDirectives.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "gc/shared/barrierSet.hpp"
36 #include "gc/shared/gc_globals.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/allocation.inline.hpp"
39 #include "memory/allocation.hpp"
40 #include "opto/ad.hpp"
41 #include "opto/block.hpp"
42 #include "opto/c2compiler.hpp"
43 #include "opto/c2_MacroAssembler.hpp"
44 #include "opto/callnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/node.hpp"
49 #include "opto/optoreg.hpp"
50 #include "opto/output.hpp"
51 #include "opto/regalloc.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/subnode.hpp"
54 #include "opto/type.hpp"
55 #include "runtime/handles.inline.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "utilities/macros.hpp"
58 #include "utilities/powerOfTwo.hpp"
59 #include "utilities/xmlstream.hpp"
60
61 #ifndef PRODUCT
62 #define DEBUG_ARG(x) , x
63 #else
64 #define DEBUG_ARG(x)
65 #endif
66
67 //------------------------------Scheduling----------------------------------
68 // This class contains all the information necessary to implement instruction
69 // scheduling and bundling.
70 class Scheduling {
71
72 private:
73 // Arena to use
74 Arena *_arena;
75
76 // Control-Flow Graph info
77 PhaseCFG *_cfg;
78
79 // Register Allocation info
80 PhaseRegAlloc *_regalloc;
81
82 // Number of nodes in the method
83 uint _node_bundling_limit;
84
85 // List of scheduled nodes. Generated in reverse order
86 Node_List _scheduled;
87
88 // List of nodes currently available for choosing for scheduling
89 Node_List _available;
90
91 // For each instruction beginning a bundle, the number of following
92 // nodes to be bundled with it.
93 Bundle *_node_bundling_base;
94
95 // Mapping from register to Node
96 Node_List _reg_node;
97
98 // Free list for pinch nodes.
99 Node_List _pinch_free_list;
100
101 // Number of uses of this node within the containing basic block.
102 short *_uses;
103
104 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
105 // front, branch+proj at end. Also skips Catch/CProj (same as
106 // branch-at-end), plus just-prior exception-throwing call.
107 uint _bb_start, _bb_end;
108
109 // Latency from the end of the basic block as scheduled
110 unsigned short *_current_latency;
111
112 // Remember the next node
113 Node *_next_node;
114
115 // Use this for an unconditional branch delay slot
116 Node *_unconditional_delay_slot;
117
118 // Pointer to a Nop
119 MachNopNode *_nop;
120
121 // Length of the current bundle, in instructions
122 uint _bundle_instr_count;
123
124 // Current Cycle number, for computing latencies and bundling
125 uint _bundle_cycle_number;
126
127 // Bundle information
128 Pipeline_Use_Element _bundle_use_elements[resource_count];
129 Pipeline_Use _bundle_use;
130
131 // Dump the available list
132 void dump_available() const;
133
134 public:
135 Scheduling(Arena *arena, Compile &compile);
136
137 // Destructor
138 NOT_PRODUCT( ~Scheduling(); )
139
140 // Step ahead "i" cycles
141 void step(uint i);
142
143 // Step ahead 1 cycle, and clear the bundle state (for example,
144 // at a branch target)
145 void step_and_clear();
146
147 Bundle* node_bundling(const Node *n) {
148 assert(valid_bundle_info(n), "oob");
149 return (&_node_bundling_base[n->_idx]);
150 }
151
152 bool valid_bundle_info(const Node *n) const {
153 return (_node_bundling_limit > n->_idx);
154 }
155
156 bool starts_bundle(const Node *n) const {
157 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
158 }
159
160 // Do the scheduling
161 void DoScheduling();
162
163 // Compute the register antidependencies within a basic block
164 void ComputeRegisterAntidependencies(Block *bb);
165 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
166 void verify_good_schedule( Block *b, const char *msg );
167 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
168 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
169
170 // Add a node to the current bundle
171 void AddNodeToBundle(Node *n, const Block *bb);
172
173 // Return an integer less than, equal to, or greater than zero
174 // if the stack offset of the first argument is respectively
175 // less than, equal to, or greater than the second.
176 int compare_two_spill_nodes(Node* first, Node* second);
177
178 // Add a node to the list of available nodes
179 void AddNodeToAvailableList(Node *n);
180
181 // Compute the local use count for the nodes in a block, and compute
182 // the list of instructions with no uses in the block as available
183 void ComputeUseCount(const Block *bb);
184
185 // Choose an instruction from the available list to add to the bundle
186 Node * ChooseNodeToBundle();
187
188 // See if this Node fits into the currently accumulating bundle
189 bool NodeFitsInBundle(Node *n);
190
191 // Decrement the use count for a node
192 void DecrementUseCounts(Node *n, const Block *bb);
193
194 // Garbage collect pinch nodes for reuse by other blocks.
195 void garbage_collect_pinch_nodes();
196 // Clean up a pinch node for reuse (helper for above).
197 void cleanup_pinch( Node *pinch );
198
199 // Information for statistics gathering
200 #ifndef PRODUCT
201 private:
202 // Gather information on size of nops relative to total
203 uint _branches, _unconditional_delays;
204
205 static uint _total_nop_size, _total_method_size;
206 static uint _total_branches, _total_unconditional_delays;
207 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
208
209 public:
210 static void print_statistics();
211
212 static void increment_instructions_per_bundle(uint i) {
213 _total_instructions_per_bundle[i]++;
214 }
215
216 static void increment_nop_size(uint s) {
217 _total_nop_size += s;
218 }
219
220 static void increment_method_size(uint s) {
221 _total_method_size += s;
222 }
223 #endif
224
225 };
226
227 PhaseOutput::PhaseOutput()
228 : Phase(Phase::Output),
229 _code_buffer("Compile::Fill_buffer"),
230 _first_block_size(0),
231 _handler_table(),
232 _inc_table(),
233 _stub_list(),
234 _oop_map_set(nullptr),
235 _scratch_buffer_blob(nullptr),
236 _scratch_locs_memory(nullptr),
237 _scratch_const_size(-1),
238 _in_scratch_emit_size(false),
239 _frame_slots(0),
240 _code_offsets(),
241 _node_bundling_limit(0),
242 _node_bundling_base(nullptr),
243 _orig_pc_slot(0),
244 _orig_pc_slot_offset_in_bytes(0),
245 _buf_sizes(),
246 _block(nullptr),
247 _index(0) {
248 C->set_output(this);
249 if (C->stub_name() == nullptr) {
250 int fixed_slots = C->fixed_slots();
251 if (C->needs_stack_repair()) {
252 fixed_slots -= 2;
253 }
254 // TODO 8284443 Only reserve extra slot if needed
255 if (InlineTypeReturnedAsFields) {
256 fixed_slots -= 2;
257 }
258 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
259 }
260 }
261
262 PhaseOutput::~PhaseOutput() {
263 C->set_output(nullptr);
264 if (_scratch_buffer_blob != nullptr) {
265 BufferBlob::free(_scratch_buffer_blob);
266 }
267 }
268
269 void PhaseOutput::perform_mach_node_analysis() {
270 // Late barrier analysis must be done after schedule and bundle
271 // Otherwise liveness based spilling will fail
272 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
273 bs->late_barrier_analysis();
274
275 pd_perform_mach_node_analysis();
276
277 C->print_method(CompilerPhaseType::PHASE_MACH_ANALYSIS, 3);
278 }
279
280 // Convert Nodes to instruction bits and pass off to the VM
281 void PhaseOutput::Output() {
282 // RootNode goes
283 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
284
285 // The number of new nodes (mostly MachNop) is proportional to
286 // the number of java calls and inner loops which are aligned.
287 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
288 C->inner_loops()*(OptoLoopAlignment-1)),
289 "out of nodes before code generation" ) ) {
290 return;
291 }
292 // Make sure I can find the Start Node
293 Block *entry = C->cfg()->get_block(1);
294 Block *broot = C->cfg()->get_root_block();
295
296 const StartNode *start = entry->head()->as_Start();
297
298 // Replace StartNode with prolog
299 Label verified_entry;
300 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
301 entry->map_node(prolog, 0);
302 C->cfg()->map_node_to_block(prolog, entry);
303 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
304
305 // Virtual methods need an unverified entry point
306 if (C->is_osr_compilation()) {
307 if (PoisonOSREntry) {
308 // TODO: Should use a ShouldNotReachHereNode...
309 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
310 }
311 } else {
312 if (C->method()) {
313 if (C->method()->has_scalarized_args()) {
314 // Add entry point to unpack all inline type arguments
315 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
316 if (!C->method()->is_static()) {
317 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
318 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
319 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
320 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
321 }
322 } else if (!C->method()->is_static()) {
323 // Insert unvalidated entry point
324 C->cfg()->insert(broot, 0, new MachUEPNode());
325 }
326 }
327 }
328
329 // Break before main entry point
330 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
331 (OptoBreakpoint && C->is_method_compilation()) ||
332 (OptoBreakpointOSR && C->is_osr_compilation()) ||
333 (OptoBreakpointC2R && !C->method()) ) {
334 // checking for C->method() means that OptoBreakpoint does not apply to
335 // runtime stubs or frame converters
336 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
337 }
338
339 // Insert epilogs before every return
340 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
341 Block* block = C->cfg()->get_block(i);
342 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
343 Node* m = block->end();
344 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
345 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
346 block->add_inst(epilog);
347 C->cfg()->map_node_to_block(epilog, block);
348 }
349 }
350 }
351
352 // Keeper of sizing aspects
353 _buf_sizes = BufferSizingData();
354
355 // Initialize code buffer
356 estimate_buffer_size(_buf_sizes._const);
357 if (C->failing()) return;
358
359 // Pre-compute the length of blocks and replace
360 // long branches with short if machine supports it.
361 // Must be done before ScheduleAndBundle due to SPARC delay slots
362 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
363 blk_starts[0] = 0;
364 shorten_branches(blk_starts);
365
366 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
367 // Compute the offsets of the entry points required by the inline type calling convention
368 if (!C->method()->is_static()) {
369 // We have entries at the beginning of the method, implemented by the first 4 nodes.
370 // Entry (unverified) @ offset 0
371 // Verified_Inline_Entry_RO
372 // Inline_Entry (unverified)
373 // Verified_Inline_Entry
374 uint offset = 0;
375 _code_offsets.set_value(CodeOffsets::Entry, offset);
376
377 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
378 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
379
380 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
381 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
382
383 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
384 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
385 } else {
386 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
387 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
388 }
389 }
390
391 ScheduleAndBundle();
392 if (C->failing()) {
393 return;
394 }
395
396 perform_mach_node_analysis();
397
398 // Complete sizing of codebuffer
399 CodeBuffer* cb = init_buffer();
400 if (cb == nullptr || C->failing()) {
401 return;
402 }
403
404 BuildOopMaps();
405
406 if (C->failing()) {
407 return;
408 }
409
410 C2_MacroAssembler masm(cb);
411 fill_buffer(&masm, blk_starts);
412 }
413
414 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
415 // Determine if we need to generate a stack overflow check.
416 // Do it if the method is not a stub function and
417 // has java calls or has frame size > vm_page_size/8.
418 // The debug VM checks that deoptimization doesn't trigger an
419 // unexpected stack overflow (compiled method stack banging should
420 // guarantee it doesn't happen) so we always need the stack bang in
421 // a debug VM.
422 return (C->stub_function() == nullptr &&
423 (C->has_java_calls() || frame_size_in_bytes > (int)(os::vm_page_size())>>3
424 DEBUG_ONLY(|| true)));
425 }
426
427 bool PhaseOutput::need_register_stack_bang() const {
428 // Determine if we need to generate a register stack overflow check.
429 // This is only used on architectures which have split register
430 // and memory stacks (ie. IA64).
431 // Bang if the method is not a stub function and has java calls
432 return (C->stub_function() == nullptr && C->has_java_calls());
433 }
434
435
436 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
437 // of a loop. When aligning a loop we need to provide enough instructions
438 // in cpu's fetch buffer to feed decoders. The loop alignment could be
439 // avoided if we have enough instructions in fetch buffer at the head of a loop.
440 // By default, the size is set to 999999 by Block's constructor so that
441 // a loop will be aligned if the size is not reset here.
442 //
443 // Note: Mach instructions could contain several HW instructions
444 // so the size is estimated only.
445 //
446 void PhaseOutput::compute_loop_first_inst_sizes() {
447 // The next condition is used to gate the loop alignment optimization.
448 // Don't aligned a loop if there are enough instructions at the head of a loop
449 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
450 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
451 // equal to 11 bytes which is the largest address NOP instruction.
452 if (MaxLoopPad < OptoLoopAlignment - 1) {
453 uint last_block = C->cfg()->number_of_blocks() - 1;
454 for (uint i = 1; i <= last_block; i++) {
455 Block* block = C->cfg()->get_block(i);
456 // Check the first loop's block which requires an alignment.
457 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
458 uint sum_size = 0;
459 uint inst_cnt = NumberOfLoopInstrToAlign;
460 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
461
462 // Check subsequent fallthrough blocks if the loop's first
463 // block(s) does not have enough instructions.
464 Block *nb = block;
465 while(inst_cnt > 0 &&
466 i < last_block &&
467 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
468 !nb->has_successor(block)) {
469 i++;
470 nb = C->cfg()->get_block(i);
471 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
472 } // while( inst_cnt > 0 && i < last_block )
473
474 block->set_first_inst_size(sum_size);
475 } // f( b->head()->is_Loop() )
476 } // for( i <= last_block )
477 } // if( MaxLoopPad < OptoLoopAlignment-1 )
478 }
479
480 // The architecture description provides short branch variants for some long
481 // branch instructions. Replace eligible long branches with short branches.
482 void PhaseOutput::shorten_branches(uint* blk_starts) {
483
484 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
485
486 // Compute size of each block, method size, and relocation information size
487 uint nblocks = C->cfg()->number_of_blocks();
488
489 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
490 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
491 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
492
493 // Collect worst case block paddings
494 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
495 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
496
497 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
498 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
499
500 bool has_short_branch_candidate = false;
501
502 // Initialize the sizes to 0
503 int code_size = 0; // Size in bytes of generated code
504 int stub_size = 0; // Size in bytes of all stub entries
505 // Size in bytes of all relocation entries, including those in local stubs.
506 // Start with 2-bytes of reloc info for the unvalidated entry point
507 int reloc_size = 1; // Number of relocation entries
508
509 // Make three passes. The first computes pessimistic blk_starts,
510 // relative jmp_offset and reloc_size information. The second performs
511 // short branch substitution using the pessimistic sizing. The
512 // third inserts nops where needed.
513
514 // Step one, perform a pessimistic sizing pass.
515 uint last_call_adr = max_juint;
516 uint last_avoid_back_to_back_adr = max_juint;
517 uint nop_size = (new MachNopNode())->size(C->regalloc());
518 for (uint i = 0; i < nblocks; i++) { // For all blocks
519 Block* block = C->cfg()->get_block(i);
520 _block = block;
521
522 // During short branch replacement, we store the relative (to blk_starts)
523 // offset of jump in jmp_offset, rather than the absolute offset of jump.
524 // This is so that we do not need to recompute sizes of all nodes when
525 // we compute correct blk_starts in our next sizing pass.
526 jmp_offset[i] = 0;
527 jmp_size[i] = 0;
528 jmp_nidx[i] = -1;
529 DEBUG_ONLY( jmp_target[i] = 0; )
530 DEBUG_ONLY( jmp_rule[i] = 0; )
531
532 // Sum all instruction sizes to compute block size
533 uint last_inst = block->number_of_nodes();
534 uint blk_size = 0;
535 for (uint j = 0; j < last_inst; j++) {
536 _index = j;
537 Node* nj = block->get_node(_index);
538 // Handle machine instruction nodes
539 if (nj->is_Mach()) {
540 MachNode* mach = nj->as_Mach();
541 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
542 reloc_size += mach->reloc();
543 if (mach->is_MachCall()) {
544 // add size information for trampoline stub
545 // class CallStubImpl is platform-specific and defined in the *.ad files.
546 stub_size += CallStubImpl::size_call_trampoline();
547 reloc_size += CallStubImpl::reloc_call_trampoline();
548
549 MachCallNode *mcall = mach->as_MachCall();
550 // This destination address is NOT PC-relative
551
552 if (mcall->entry_point() != nullptr) {
553 mcall->method_set((intptr_t)mcall->entry_point());
554 }
555
556 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
557 stub_size += CompiledDirectCall::to_interp_stub_size();
558 reloc_size += CompiledDirectCall::reloc_to_interp_stub();
559 }
560 } else if (mach->is_MachSafePoint()) {
561 // If call/safepoint are adjacent, account for possible
562 // nop to disambiguate the two safepoints.
563 // ScheduleAndBundle() can rearrange nodes in a block,
564 // check for all offsets inside this block.
565 if (last_call_adr >= blk_starts[i]) {
566 blk_size += nop_size;
567 }
568 }
569 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
570 // Nop is inserted between "avoid back to back" instructions.
571 // ScheduleAndBundle() can rearrange nodes in a block,
572 // check for all offsets inside this block.
573 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
574 blk_size += nop_size;
575 }
576 }
577 if (mach->may_be_short_branch()) {
578 if (!nj->is_MachBranch()) {
579 #ifndef PRODUCT
580 nj->dump(3);
581 #endif
582 Unimplemented();
583 }
584 assert(jmp_nidx[i] == -1, "block should have only one branch");
585 jmp_offset[i] = blk_size;
586 jmp_size[i] = nj->size(C->regalloc());
587 jmp_nidx[i] = j;
588 has_short_branch_candidate = true;
589 }
590 }
591 blk_size += nj->size(C->regalloc());
592 // Remember end of call offset
593 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
594 last_call_adr = blk_starts[i]+blk_size;
595 }
596 // Remember end of avoid_back_to_back offset
597 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
598 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
599 }
600 }
601
602 // When the next block starts a loop, we may insert pad NOP
603 // instructions. Since we cannot know our future alignment,
604 // assume the worst.
605 if (i < nblocks - 1) {
606 Block* nb = C->cfg()->get_block(i + 1);
607 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
608 if (max_loop_pad > 0) {
609 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
610 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
611 // If either is the last instruction in this block, bump by
612 // max_loop_pad in lock-step with blk_size, so sizing
613 // calculations in subsequent blocks still can conservatively
614 // detect that it may the last instruction in this block.
615 if (last_call_adr == blk_starts[i]+blk_size) {
616 last_call_adr += max_loop_pad;
617 }
618 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
619 last_avoid_back_to_back_adr += max_loop_pad;
620 }
621 blk_size += max_loop_pad;
622 block_worst_case_pad[i + 1] = max_loop_pad;
623 }
624 }
625
626 // Save block size; update total method size
627 blk_starts[i+1] = blk_starts[i]+blk_size;
628 }
629
630 // Step two, replace eligible long jumps.
631 bool progress = true;
632 uint last_may_be_short_branch_adr = max_juint;
633 while (has_short_branch_candidate && progress) {
634 progress = false;
635 has_short_branch_candidate = false;
636 int adjust_block_start = 0;
637 for (uint i = 0; i < nblocks; i++) {
638 Block* block = C->cfg()->get_block(i);
639 int idx = jmp_nidx[i];
640 MachNode* mach = (idx == -1) ? nullptr: block->get_node(idx)->as_Mach();
641 if (mach != nullptr && mach->may_be_short_branch()) {
642 #ifdef ASSERT
643 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
644 int j;
645 // Find the branch; ignore trailing NOPs.
646 for (j = block->number_of_nodes()-1; j>=0; j--) {
647 Node* n = block->get_node(j);
648 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
649 break;
650 }
651 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
652 #endif
653 int br_size = jmp_size[i];
654 int br_offs = blk_starts[i] + jmp_offset[i];
655
656 // This requires the TRUE branch target be in succs[0]
657 uint bnum = block->non_connector_successor(0)->_pre_order;
658 int offset = blk_starts[bnum] - br_offs;
659 if (bnum > i) { // adjust following block's offset
660 offset -= adjust_block_start;
661 }
662
663 // This block can be a loop header, account for the padding
664 // in the previous block.
665 int block_padding = block_worst_case_pad[i];
666 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
667 // In the following code a nop could be inserted before
668 // the branch which will increase the backward distance.
669 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
670 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
671
672 if (needs_padding && offset <= 0)
673 offset -= nop_size;
674
675 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
676 // We've got a winner. Replace this branch.
677 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
678
679 // Update the jmp_size.
680 int new_size = replacement->size(C->regalloc());
681 int diff = br_size - new_size;
682 assert(diff >= (int)nop_size, "short_branch size should be smaller");
683 // Conservatively take into account padding between
684 // avoid_back_to_back branches. Previous branch could be
685 // converted into avoid_back_to_back branch during next
686 // rounds.
687 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
688 jmp_offset[i] += nop_size;
689 diff -= nop_size;
690 }
691 adjust_block_start += diff;
692 block->map_node(replacement, idx);
693 mach->subsume_by(replacement, C);
694 mach = replacement;
695 progress = true;
696
697 jmp_size[i] = new_size;
698 DEBUG_ONLY( jmp_target[i] = bnum; );
699 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
700 } else {
701 // The jump distance is not short, try again during next iteration.
702 has_short_branch_candidate = true;
703 }
704 } // (mach->may_be_short_branch())
705 if (mach != nullptr && (mach->may_be_short_branch() ||
706 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
707 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
708 }
709 blk_starts[i+1] -= adjust_block_start;
710 }
711 }
712
713 #ifdef ASSERT
714 for (uint i = 0; i < nblocks; i++) { // For all blocks
715 if (jmp_target[i] != 0) {
716 int br_size = jmp_size[i];
717 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
718 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
719 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
720 }
721 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
722 }
723 }
724 #endif
725
726 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
727 // after ScheduleAndBundle().
728
729 // ------------------
730 // Compute size for code buffer
731 code_size = blk_starts[nblocks];
732
733 // Relocation records
734 reloc_size += 1; // Relo entry for exception handler
735
736 // Adjust reloc_size to number of record of relocation info
737 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
738 // a relocation index.
739 // The CodeBuffer will expand the locs array if this estimate is too low.
740 reloc_size *= 10 / sizeof(relocInfo);
741
742 _buf_sizes._reloc = reloc_size;
743 _buf_sizes._code = code_size;
744 _buf_sizes._stub = stub_size;
745 }
746
747 //------------------------------FillLocArray-----------------------------------
748 // Create a bit of debug info and append it to the array. The mapping is from
749 // Java local or expression stack to constant, register or stack-slot. For
750 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
751 // entry has been taken care of and caller should skip it).
752 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
753 // This should never have accepted Bad before
754 assert(OptoReg::is_valid(regnum), "location must be valid");
755 return (OptoReg::is_reg(regnum))
756 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
757 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
758 }
759
760
761 ObjectValue*
762 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
763 for (int i = 0; i < objs->length(); i++) {
764 assert(objs->at(i)->is_object(), "corrupt object cache");
765 ObjectValue* sv = objs->at(i)->as_ObjectValue();
766 if (sv->id() == id) {
767 return sv;
768 }
769 }
770 // Otherwise..
771 return nullptr;
772 }
773
774 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
775 ObjectValue* sv ) {
776 assert(sv_for_node_id(objs, sv->id()) == nullptr, "Precondition");
777 objs->append(sv);
778 }
779
780
781 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
782 GrowableArray<ScopeValue*> *array,
783 GrowableArray<ScopeValue*> *objs ) {
784 assert( local, "use _top instead of null" );
785 if (array->length() != idx) {
786 assert(array->length() == idx + 1, "Unexpected array count");
787 // Old functionality:
788 // return
789 // New functionality:
790 // Assert if the local is not top. In product mode let the new node
791 // override the old entry.
792 assert(local == C->top(), "LocArray collision");
793 if (local == C->top()) {
794 return;
795 }
796 array->pop();
797 }
798 const Type *t = local->bottom_type();
799
800 // Is it a safepoint scalar object node?
801 if (local->is_SafePointScalarObject()) {
802 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
803
804 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
805 if (sv == nullptr) {
806 ciKlass* cik = t->is_oopptr()->exact_klass();
807 assert(cik->is_instance_klass() ||
808 cik->is_array_klass(), "Not supported allocation.");
809 uint first_ind = spobj->first_index(sfpt->jvms());
810 // Nullable, scalarized inline types have an is_init input
811 // that needs to be checked before using the field values.
812 ScopeValue* is_init = nullptr;
813 if (cik->is_inlinetype()) {
814 Node* init_node = sfpt->in(first_ind++);
815 assert(init_node != nullptr, "is_init node not found");
816 if (!init_node->is_top()) {
817 const TypeInt* init_type = init_node->bottom_type()->is_int();
818 if (init_node->is_Con()) {
819 is_init = new ConstantIntValue(init_type->get_con());
820 } else {
821 OptoReg::Name init_reg = C->regalloc()->get_reg_first(init_node);
822 is_init = new_loc_value(C->regalloc(), init_reg, Location::normal);
823 }
824 }
825 }
826 sv = new ObjectValue(spobj->_idx,
827 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), true, is_init);
828 set_sv_for_object_node(objs, sv);
829
830 for (uint i = 0; i < spobj->n_fields(); i++) {
831 Node* fld_node = sfpt->in(first_ind+i);
832 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
833 }
834 }
835 array->append(sv);
836 return;
837 } else if (local->is_SafePointScalarMerge()) {
838 SafePointScalarMergeNode* smerge = local->as_SafePointScalarMerge();
839 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
840
841 if (mv == nullptr) {
842 GrowableArray<ScopeValue*> deps;
843
844 int merge_pointer_idx = smerge->merge_pointer_idx(sfpt->jvms());
845 (void)FillLocArray(0, sfpt, sfpt->in(merge_pointer_idx), &deps, objs);
846 assert(deps.length() == 1, "missing value");
847
848 int selector_idx = smerge->selector_idx(sfpt->jvms());
849 (void)FillLocArray(1, nullptr, sfpt->in(selector_idx), &deps, nullptr);
850 assert(deps.length() == 2, "missing value");
851
852 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
853 set_sv_for_object_node(objs, mv);
854
855 for (uint i = 1; i < smerge->req(); i++) {
856 Node* obj_node = smerge->in(i);
857 (void)FillLocArray(mv->possible_objects()->length(), sfpt, obj_node, mv->possible_objects(), objs);
858 }
859 }
860 array->append(mv);
861 return;
862 }
863
864 // Grab the register number for the local
865 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
866 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
867 // Record the double as two float registers.
868 // The register mask for such a value always specifies two adjacent
869 // float registers, with the lower register number even.
870 // Normally, the allocation of high and low words to these registers
871 // is irrelevant, because nearly all operations on register pairs
872 // (e.g., StoreD) treat them as a single unit.
873 // Here, we assume in addition that the words in these two registers
874 // stored "naturally" (by operations like StoreD and double stores
875 // within the interpreter) such that the lower-numbered register
876 // is written to the lower memory address. This may seem like
877 // a machine dependency, but it is not--it is a requirement on
878 // the author of the <arch>.ad file to ensure that, for every
879 // even/odd double-register pair to which a double may be allocated,
880 // the word in the even single-register is stored to the first
881 // memory word. (Note that register numbers are completely
882 // arbitrary, and are not tied to any machine-level encodings.)
883 #ifdef _LP64
884 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
885 array->append(new ConstantIntValue((jint)0));
886 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
887 } else if ( t->base() == Type::Long ) {
888 array->append(new ConstantIntValue((jint)0));
889 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
890 } else if ( t->base() == Type::RawPtr ) {
891 // jsr/ret return address which must be restored into the full
892 // width 64-bit stack slot.
893 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
894 }
895 #else //_LP64
896 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
897 // Repack the double/long as two jints.
898 // The convention the interpreter uses is that the second local
899 // holds the first raw word of the native double representation.
900 // This is actually reasonable, since locals and stack arrays
901 // grow downwards in all implementations.
902 // (If, on some machine, the interpreter's Java locals or stack
903 // were to grow upwards, the embedded doubles would be word-swapped.)
904 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
905 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
906 }
907 #endif //_LP64
908 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
909 OptoReg::is_reg(regnum) ) {
910 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
911 ? Location::float_in_dbl : Location::normal ));
912 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
913 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
914 ? Location::int_in_long : Location::normal ));
915 } else if( t->base() == Type::NarrowOop ) {
916 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
917 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
918 t->base() == Type::VectorD || t->base() == Type::VectorX ||
919 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
920 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
921 } else if (C->regalloc()->is_oop(local)) {
922 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
923 t->base() == Type::AryPtr,
924 "Unexpected type: %s", t->msg());
925 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
926 } else {
927 assert(t->base() == Type::Int || t->base() == Type::Half ||
928 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
929 "Unexpected type: %s", t->msg());
930 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
931 }
932 return;
933 }
934
935 // No register. It must be constant data.
936 switch (t->base()) {
937 case Type::Half: // Second half of a double
938 ShouldNotReachHere(); // Caller should skip 2nd halves
939 break;
940 case Type::AnyPtr:
941 array->append(new ConstantOopWriteValue(nullptr));
942 break;
943 case Type::AryPtr:
944 case Type::InstPtr: // fall through
945 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
946 break;
947 case Type::NarrowOop:
948 if (t == TypeNarrowOop::NULL_PTR) {
949 array->append(new ConstantOopWriteValue(nullptr));
950 } else {
951 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
952 }
953 break;
954 case Type::Int:
955 array->append(new ConstantIntValue(t->is_int()->get_con()));
956 break;
957 case Type::RawPtr:
958 // A return address (T_ADDRESS).
959 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
960 #ifdef _LP64
961 // Must be restored to the full-width 64-bit stack slot.
962 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
963 #else
964 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
965 #endif
966 break;
967 case Type::FloatCon: {
968 float f = t->is_float_constant()->getf();
969 array->append(new ConstantIntValue(jint_cast(f)));
970 break;
971 }
972 case Type::DoubleCon: {
973 jdouble d = t->is_double_constant()->getd();
974 #ifdef _LP64
975 array->append(new ConstantIntValue((jint)0));
976 array->append(new ConstantDoubleValue(d));
977 #else
978 // Repack the double as two jints.
979 // The convention the interpreter uses is that the second local
980 // holds the first raw word of the native double representation.
981 // This is actually reasonable, since locals and stack arrays
982 // grow downwards in all implementations.
983 // (If, on some machine, the interpreter's Java locals or stack
984 // were to grow upwards, the embedded doubles would be word-swapped.)
985 jlong_accessor acc;
986 acc.long_value = jlong_cast(d);
987 array->append(new ConstantIntValue(acc.words[1]));
988 array->append(new ConstantIntValue(acc.words[0]));
989 #endif
990 break;
991 }
992 case Type::Long: {
993 jlong d = t->is_long()->get_con();
994 #ifdef _LP64
995 array->append(new ConstantIntValue((jint)0));
996 array->append(new ConstantLongValue(d));
997 #else
998 // Repack the long as two jints.
999 // The convention the interpreter uses is that the second local
1000 // holds the first raw word of the native double representation.
1001 // This is actually reasonable, since locals and stack arrays
1002 // grow downwards in all implementations.
1003 // (If, on some machine, the interpreter's Java locals or stack
1004 // were to grow upwards, the embedded doubles would be word-swapped.)
1005 jlong_accessor acc;
1006 acc.long_value = d;
1007 array->append(new ConstantIntValue(acc.words[1]));
1008 array->append(new ConstantIntValue(acc.words[0]));
1009 #endif
1010 break;
1011 }
1012 case Type::Top: // Add an illegal value here
1013 array->append(new LocationValue(Location()));
1014 break;
1015 default:
1016 ShouldNotReachHere();
1017 break;
1018 }
1019 }
1020
1021 // Determine if this node starts a bundle
1022 bool PhaseOutput::starts_bundle(const Node *n) const {
1023 return (_node_bundling_limit > n->_idx &&
1024 _node_bundling_base[n->_idx].starts_bundle());
1025 }
1026
1027 // Determine if there is a monitor that has 'ov' as its owner.
1028 bool PhaseOutput::contains_as_owner(GrowableArray<MonitorValue*> *monarray, ObjectValue *ov) const {
1029 for (int k = 0; k < monarray->length(); k++) {
1030 MonitorValue* mv = monarray->at(k);
1031 if (mv->owner() == ov) {
1032 return true;
1033 }
1034 }
1035
1036 return false;
1037 }
1038
1039 // Determine if there is a scalar replaced object description represented by 'ov'.
1040 bool PhaseOutput::contains_as_scalarized_obj(JVMState* jvms, MachSafePointNode* sfn,
1041 GrowableArray<ScopeValue*>* objs,
1042 ObjectValue* ov) const {
1043 for (int i = 0; i < jvms->scl_size(); i++) {
1044 Node* n = sfn->scalarized_obj(jvms, i);
1045 // Other kinds of nodes that we may encounter here, for instance constants
1046 // representing values of fields of objects scalarized, aren't relevant for
1047 // us, since they don't map to ObjectValue.
1048 if (!n->is_SafePointScalarObject()) {
1049 continue;
1050 }
1051
1052 ObjectValue* other = sv_for_node_id(objs, n->_idx);
1053 if (ov == other) {
1054 return true;
1055 }
1056 }
1057 return false;
1058 }
1059
1060 //--------------------------Process_OopMap_Node--------------------------------
1061 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1062 // Handle special safepoint nodes for synchronization
1063 MachSafePointNode *sfn = mach->as_MachSafePoint();
1064 MachCallNode *mcall;
1065
1066 int safepoint_pc_offset = current_offset;
1067 bool is_method_handle_invoke = false;
1068 bool return_oop = false;
1069 bool return_scalarized = false;
1070 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1071 bool arg_escape = false;
1072
1073 // Add the safepoint in the DebugInfoRecorder
1074 if( !mach->is_MachCall() ) {
1075 mcall = nullptr;
1076 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1077 } else {
1078 mcall = mach->as_MachCall();
1079
1080 // Is the call a MethodHandle call?
1081 if (mcall->is_MachCallJava()) {
1082 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1083 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1084 is_method_handle_invoke = true;
1085 }
1086 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1087 }
1088
1089 // Check if a call returns an object.
1090 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1091 return_oop = true;
1092 }
1093 if (mcall->returns_scalarized()) {
1094 return_scalarized = true;
1095 }
1096 safepoint_pc_offset += mcall->ret_addr_offset();
1097 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1098 }
1099
1100 // Loop over the JVMState list to add scope information
1101 // Do not skip safepoints with a null method, they need monitor info
1102 JVMState* youngest_jvms = sfn->jvms();
1103 int max_depth = youngest_jvms->depth();
1104
1105 // Allocate the object pool for scalar-replaced objects -- the map from
1106 // small-integer keys (which can be recorded in the local and ostack
1107 // arrays) to descriptions of the object state.
1108 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1109
1110 // Visit scopes from oldest to youngest.
1111 for (int depth = 1; depth <= max_depth; depth++) {
1112 JVMState* jvms = youngest_jvms->of_depth(depth);
1113 int idx;
1114 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1115 // Safepoints that do not have method() set only provide oop-map and monitor info
1116 // to support GC; these do not support deoptimization.
1117 int num_locs = (method == nullptr) ? 0 : jvms->loc_size();
1118 int num_exps = (method == nullptr) ? 0 : jvms->stk_size();
1119 int num_mon = jvms->nof_monitors();
1120 assert(method == nullptr || jvms->bci() < 0 || num_locs == method->max_locals(),
1121 "JVMS local count must match that of the method");
1122
1123 // Add Local and Expression Stack Information
1124
1125 // Insert locals into the locarray
1126 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1127 for( idx = 0; idx < num_locs; idx++ ) {
1128 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1129 }
1130
1131 // Insert expression stack entries into the exparray
1132 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1133 for( idx = 0; idx < num_exps; idx++ ) {
1134 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1135 }
1136
1137 // Add in mappings of the monitors
1138 assert( !method ||
1139 !method->is_synchronized() ||
1140 method->is_native() ||
1141 num_mon > 0 ||
1142 !GenerateSynchronizationCode,
1143 "monitors must always exist for synchronized methods");
1144
1145 // Build the growable array of ScopeValues for exp stack
1146 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1147
1148 // Loop over monitors and insert into array
1149 for (idx = 0; idx < num_mon; idx++) {
1150 // Grab the node that defines this monitor
1151 Node* box_node = sfn->monitor_box(jvms, idx);
1152 Node* obj_node = sfn->monitor_obj(jvms, idx);
1153
1154 // Create ScopeValue for object
1155 ScopeValue *scval = nullptr;
1156
1157 if (obj_node->is_SafePointScalarObject()) {
1158 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1159 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1160 if (scval == nullptr) {
1161 const Type *t = spobj->bottom_type();
1162 ciKlass* cik = t->is_oopptr()->exact_klass();
1163 assert(cik->is_instance_klass() ||
1164 cik->is_array_klass(), "Not supported allocation.");
1165 ObjectValue* sv = new ObjectValue(spobj->_idx,
1166 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1167 PhaseOutput::set_sv_for_object_node(objs, sv);
1168
1169 uint first_ind = spobj->first_index(youngest_jvms);
1170 for (uint i = 0; i < spobj->n_fields(); i++) {
1171 Node* fld_node = sfn->in(first_ind+i);
1172 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1173 }
1174 scval = sv;
1175 }
1176 } else if (obj_node->is_SafePointScalarMerge()) {
1177 SafePointScalarMergeNode* smerge = obj_node->as_SafePointScalarMerge();
1178 ObjectMergeValue* mv = (ObjectMergeValue*) sv_for_node_id(objs, smerge->_idx);
1179
1180 if (mv == nullptr) {
1181 GrowableArray<ScopeValue*> deps;
1182
1183 int merge_pointer_idx = smerge->merge_pointer_idx(youngest_jvms);
1184 FillLocArray(0, sfn, sfn->in(merge_pointer_idx), &deps, objs);
1185 assert(deps.length() == 1, "missing value");
1186
1187 int selector_idx = smerge->selector_idx(youngest_jvms);
1188 FillLocArray(1, nullptr, sfn->in(selector_idx), &deps, nullptr);
1189 assert(deps.length() == 2, "missing value");
1190
1191 mv = new ObjectMergeValue(smerge->_idx, deps.at(0), deps.at(1));
1192 set_sv_for_object_node(objs, mv);
1193
1194 for (uint i = 1; i < smerge->req(); i++) {
1195 Node* obj_node = smerge->in(i);
1196 FillLocArray(mv->possible_objects()->length(), sfn, obj_node, mv->possible_objects(), objs);
1197 }
1198 }
1199 scval = mv;
1200 } else if (!obj_node->is_Con()) {
1201 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1202 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1203 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1204 } else {
1205 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1206 }
1207 } else {
1208 const TypePtr *tp = obj_node->get_ptr_type();
1209 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1210 }
1211
1212 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1213 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1214 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1215 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1216 }
1217
1218 // Mark ObjectValue nodes as root nodes if they are directly
1219 // referenced in the JVMS.
1220 for (int i = 0; i < objs->length(); i++) {
1221 ScopeValue* sv = objs->at(i);
1222 if (sv->is_object_merge()) {
1223 ObjectMergeValue* merge = sv->as_ObjectMergeValue();
1224
1225 for (int j = 0; j< merge->possible_objects()->length(); j++) {
1226 ObjectValue* ov = merge->possible_objects()->at(j)->as_ObjectValue();
1227 bool is_root = locarray->contains(ov) ||
1228 exparray->contains(ov) ||
1229 contains_as_owner(monarray, ov) ||
1230 contains_as_scalarized_obj(jvms, sfn, objs, ov);
1231 ov->set_root(is_root);
1232 }
1233 }
1234 }
1235
1236 // We dump the object pool first, since deoptimization reads it in first.
1237 C->debug_info()->dump_object_pool(objs);
1238
1239 // Build first class objects to pass to scope
1240 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1241 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1242 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1243
1244 // Make method available for all Safepoints
1245 ciMethod* scope_method = method ? method : C->method();
1246 // Describe the scope here
1247 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1248 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1249 // Now we can describe the scope.
1250 methodHandle null_mh;
1251 bool rethrow_exception = false;
1252 C->debug_info()->describe_scope(
1253 safepoint_pc_offset,
1254 null_mh,
1255 scope_method,
1256 jvms->bci(),
1257 jvms->should_reexecute(),
1258 rethrow_exception,
1259 is_method_handle_invoke,
1260 return_oop,
1261 return_scalarized,
1262 has_ea_local_in_scope,
1263 arg_escape,
1264 locvals,
1265 expvals,
1266 monvals
1267 );
1268 } // End jvms loop
1269
1270 // Mark the end of the scope set.
1271 C->debug_info()->end_safepoint(safepoint_pc_offset);
1272 }
1273
1274
1275
1276 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1277 class NonSafepointEmitter {
1278 Compile* C;
1279 JVMState* _pending_jvms;
1280 int _pending_offset;
1281
1282 void emit_non_safepoint();
1283
1284 public:
1285 NonSafepointEmitter(Compile* compile) {
1286 this->C = compile;
1287 _pending_jvms = nullptr;
1288 _pending_offset = 0;
1289 }
1290
1291 void observe_instruction(Node* n, int pc_offset) {
1292 if (!C->debug_info()->recording_non_safepoints()) return;
1293
1294 Node_Notes* nn = C->node_notes_at(n->_idx);
1295 if (nn == nullptr || nn->jvms() == nullptr) return;
1296 if (_pending_jvms != nullptr &&
1297 _pending_jvms->same_calls_as(nn->jvms())) {
1298 // Repeated JVMS? Stretch it up here.
1299 _pending_offset = pc_offset;
1300 } else {
1301 if (_pending_jvms != nullptr &&
1302 _pending_offset < pc_offset) {
1303 emit_non_safepoint();
1304 }
1305 _pending_jvms = nullptr;
1306 if (pc_offset > C->debug_info()->last_pc_offset()) {
1307 // This is the only way _pending_jvms can become non-null:
1308 _pending_jvms = nn->jvms();
1309 _pending_offset = pc_offset;
1310 }
1311 }
1312 }
1313
1314 // Stay out of the way of real safepoints:
1315 void observe_safepoint(JVMState* jvms, int pc_offset) {
1316 if (_pending_jvms != nullptr &&
1317 !_pending_jvms->same_calls_as(jvms) &&
1318 _pending_offset < pc_offset) {
1319 emit_non_safepoint();
1320 }
1321 _pending_jvms = nullptr;
1322 }
1323
1324 void flush_at_end() {
1325 if (_pending_jvms != nullptr) {
1326 emit_non_safepoint();
1327 }
1328 _pending_jvms = nullptr;
1329 }
1330 };
1331
1332 void NonSafepointEmitter::emit_non_safepoint() {
1333 JVMState* youngest_jvms = _pending_jvms;
1334 int pc_offset = _pending_offset;
1335
1336 // Clear it now:
1337 _pending_jvms = nullptr;
1338
1339 DebugInformationRecorder* debug_info = C->debug_info();
1340 assert(debug_info->recording_non_safepoints(), "sanity");
1341
1342 debug_info->add_non_safepoint(pc_offset);
1343 int max_depth = youngest_jvms->depth();
1344
1345 // Visit scopes from oldest to youngest.
1346 for (int depth = 1; depth <= max_depth; depth++) {
1347 JVMState* jvms = youngest_jvms->of_depth(depth);
1348 ciMethod* method = jvms->has_method() ? jvms->method() : nullptr;
1349 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1350 methodHandle null_mh;
1351 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1352 }
1353
1354 // Mark the end of the scope set.
1355 debug_info->end_non_safepoint(pc_offset);
1356 }
1357
1358 //------------------------------init_buffer------------------------------------
1359 void PhaseOutput::estimate_buffer_size(int& const_req) {
1360
1361 // Set the initially allocated size
1362 const_req = initial_const_capacity;
1363
1364 // The extra spacing after the code is necessary on some platforms.
1365 // Sometimes we need to patch in a jump after the last instruction,
1366 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1367
1368 // Compute the byte offset where we can store the deopt pc.
1369 if (C->fixed_slots() != 0) {
1370 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1371 }
1372
1373 // Compute prolog code size
1374 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1375 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1376
1377 if (C->has_mach_constant_base_node()) {
1378 uint add_size = 0;
1379 // Fill the constant table.
1380 // Note: This must happen before shorten_branches.
1381 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1382 Block* b = C->cfg()->get_block(i);
1383
1384 for (uint j = 0; j < b->number_of_nodes(); j++) {
1385 Node* n = b->get_node(j);
1386
1387 // If the node is a MachConstantNode evaluate the constant
1388 // value section.
1389 if (n->is_MachConstant()) {
1390 MachConstantNode* machcon = n->as_MachConstant();
1391 machcon->eval_constant(C);
1392 } else if (n->is_Mach()) {
1393 // On Power there are more nodes that issue constants.
1394 add_size += (n->as_Mach()->ins_num_consts() * 8);
1395 }
1396 }
1397 }
1398
1399 // Calculate the offsets of the constants and the size of the
1400 // constant table (including the padding to the next section).
1401 constant_table().calculate_offsets_and_size();
1402 const_req = constant_table().size() + add_size;
1403 }
1404
1405 // Initialize the space for the BufferBlob used to find and verify
1406 // instruction size in MachNode::emit_size()
1407 init_scratch_buffer_blob(const_req);
1408 }
1409
1410 CodeBuffer* PhaseOutput::init_buffer() {
1411 int stub_req = _buf_sizes._stub;
1412 int code_req = _buf_sizes._code;
1413 int const_req = _buf_sizes._const;
1414
1415 int pad_req = NativeCall::byte_size();
1416
1417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1418 stub_req += bs->estimate_stub_size();
1419
1420 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1421 // class HandlerImpl is platform-specific and defined in the *.ad files.
1422 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1423 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1424 stub_req += MAX_stubs_size; // ensure per-stub margin
1425 code_req += MAX_inst_size; // ensure per-instruction margin
1426
1427 if (StressCodeBuffers)
1428 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1429
1430 int total_req =
1431 const_req +
1432 code_req +
1433 pad_req +
1434 stub_req +
1435 exception_handler_req +
1436 deopt_handler_req; // deopt handler
1437
1438 if (C->has_method_handle_invokes())
1439 total_req += deopt_handler_req; // deopt MH handler
1440
1441 CodeBuffer* cb = code_buffer();
1442 cb->initialize(total_req, _buf_sizes._reloc);
1443
1444 // Have we run out of code space?
1445 if ((cb->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1446 C->record_failure("CodeCache is full");
1447 return nullptr;
1448 }
1449 // Configure the code buffer.
1450 cb->initialize_consts_size(const_req);
1451 cb->initialize_stubs_size(stub_req);
1452 cb->initialize_oop_recorder(C->env()->oop_recorder());
1453
1454 // fill in the nop array for bundling computations
1455 MachNode *_nop_list[Bundle::_nop_count];
1456 Bundle::initialize_nops(_nop_list);
1457
1458 return cb;
1459 }
1460
1461 //------------------------------fill_buffer------------------------------------
1462 void PhaseOutput::fill_buffer(C2_MacroAssembler* masm, uint* blk_starts) {
1463 // blk_starts[] contains offsets calculated during short branches processing,
1464 // offsets should not be increased during following steps.
1465
1466 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1467 // of a loop. It is used to determine the padding for loop alignment.
1468 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1469
1470 compute_loop_first_inst_sizes();
1471
1472 // Create oopmap set.
1473 _oop_map_set = new OopMapSet();
1474
1475 // !!!!! This preserves old handling of oopmaps for now
1476 C->debug_info()->set_oopmaps(_oop_map_set);
1477
1478 uint nblocks = C->cfg()->number_of_blocks();
1479 // Count and start of implicit null check instructions
1480 uint inct_cnt = 0;
1481 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1482
1483 // Count and start of calls
1484 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1485
1486 uint return_offset = 0;
1487 int nop_size = (new MachNopNode())->size(C->regalloc());
1488
1489 int previous_offset = 0;
1490 int current_offset = 0;
1491 int last_call_offset = -1;
1492 int last_avoid_back_to_back_offset = -1;
1493 #ifdef ASSERT
1494 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1495 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1496 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1497 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1498 #endif
1499
1500 // Create an array of unused labels, one for each basic block, if printing is enabled
1501 #if defined(SUPPORT_OPTO_ASSEMBLY)
1502 int* node_offsets = nullptr;
1503 uint node_offset_limit = C->unique();
1504
1505 if (C->print_assembly()) {
1506 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1507 }
1508 if (node_offsets != nullptr) {
1509 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1510 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1511 }
1512 #endif
1513
1514 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1515
1516 // Emit the constant table.
1517 if (C->has_mach_constant_base_node()) {
1518 if (!constant_table().emit(masm)) {
1519 C->record_failure("consts section overflow");
1520 return;
1521 }
1522 }
1523
1524 // Create an array of labels, one for each basic block
1525 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1526 for (uint i = 0; i <= nblocks; i++) {
1527 blk_labels[i].init();
1528 }
1529
1530 // Now fill in the code buffer
1531 Node* delay_slot = nullptr;
1532 for (uint i = 0; i < nblocks; i++) {
1533 Block* block = C->cfg()->get_block(i);
1534 _block = block;
1535 Node* head = block->head();
1536
1537 // If this block needs to start aligned (i.e, can be reached other
1538 // than by falling-thru from the previous block), then force the
1539 // start of a new bundle.
1540 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1541 masm->code()->flush_bundle(true);
1542 }
1543
1544 #ifdef ASSERT
1545 if (!block->is_connector()) {
1546 stringStream st;
1547 block->dump_head(C->cfg(), &st);
1548 masm->block_comment(st.freeze());
1549 }
1550 jmp_target[i] = 0;
1551 jmp_offset[i] = 0;
1552 jmp_size[i] = 0;
1553 jmp_rule[i] = 0;
1554 #endif
1555 int blk_offset = current_offset;
1556
1557 // Define the label at the beginning of the basic block
1558 masm->bind(blk_labels[block->_pre_order]);
1559
1560 uint last_inst = block->number_of_nodes();
1561
1562 // Emit block normally, except for last instruction.
1563 // Emit means "dump code bits into code buffer".
1564 for (uint j = 0; j<last_inst; j++) {
1565 _index = j;
1566
1567 // Get the node
1568 Node* n = block->get_node(j);
1569
1570 // See if delay slots are supported
1571 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1572 assert(delay_slot == nullptr, "no use of delay slot node");
1573 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1574
1575 delay_slot = n;
1576 continue;
1577 }
1578
1579 // If this starts a new instruction group, then flush the current one
1580 // (but allow split bundles)
1581 if (Pipeline::requires_bundling() && starts_bundle(n))
1582 masm->code()->flush_bundle(false);
1583
1584 // Special handling for SafePoint/Call Nodes
1585 bool is_mcall = false;
1586 if (n->is_Mach()) {
1587 MachNode *mach = n->as_Mach();
1588 is_mcall = n->is_MachCall();
1589 bool is_sfn = n->is_MachSafePoint();
1590
1591 // If this requires all previous instructions be flushed, then do so
1592 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1593 masm->code()->flush_bundle(true);
1594 current_offset = masm->offset();
1595 }
1596
1597 // A padding may be needed again since a previous instruction
1598 // could be moved to delay slot.
1599
1600 // align the instruction if necessary
1601 int padding = mach->compute_padding(current_offset);
1602 // Make sure safepoint node for polling is distinct from a call's
1603 // return by adding a nop if needed.
1604 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1605 padding = nop_size;
1606 }
1607 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1608 current_offset == last_avoid_back_to_back_offset) {
1609 // Avoid back to back some instructions.
1610 padding = nop_size;
1611 }
1612
1613 if (padding > 0) {
1614 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1615 int nops_cnt = padding / nop_size;
1616 MachNode *nop = new MachNopNode(nops_cnt);
1617 block->insert_node(nop, j++);
1618 last_inst++;
1619 C->cfg()->map_node_to_block(nop, block);
1620 // Ensure enough space.
1621 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1622 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1623 C->record_failure("CodeCache is full");
1624 return;
1625 }
1626 nop->emit(masm, C->regalloc());
1627 masm->code()->flush_bundle(true);
1628 current_offset = masm->offset();
1629 }
1630
1631 bool observe_safepoint = is_sfn;
1632 // Remember the start of the last call in a basic block
1633 if (is_mcall) {
1634 MachCallNode *mcall = mach->as_MachCall();
1635
1636 if (mcall->entry_point() != nullptr) {
1637 // This destination address is NOT PC-relative
1638 mcall->method_set((intptr_t)mcall->entry_point());
1639 }
1640
1641 // Save the return address
1642 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1643
1644 observe_safepoint = mcall->guaranteed_safepoint();
1645 }
1646
1647 // sfn will be valid whenever mcall is valid now because of inheritance
1648 if (observe_safepoint) {
1649 // Handle special safepoint nodes for synchronization
1650 if (!is_mcall) {
1651 MachSafePointNode *sfn = mach->as_MachSafePoint();
1652 // !!!!! Stubs only need an oopmap right now, so bail out
1653 if (sfn->jvms()->method() == nullptr) {
1654 // Write the oopmap directly to the code blob??!!
1655 continue;
1656 }
1657 } // End synchronization
1658
1659 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1660 current_offset);
1661 Process_OopMap_Node(mach, current_offset);
1662 } // End if safepoint
1663
1664 // If this is a null check, then add the start of the previous instruction to the list
1665 else if( mach->is_MachNullCheck() ) {
1666 inct_starts[inct_cnt++] = previous_offset;
1667 }
1668
1669 // If this is a branch, then fill in the label with the target BB's label
1670 else if (mach->is_MachBranch()) {
1671 // This requires the TRUE branch target be in succs[0]
1672 uint block_num = block->non_connector_successor(0)->_pre_order;
1673
1674 // Try to replace long branch if delay slot is not used,
1675 // it is mostly for back branches since forward branch's
1676 // distance is not updated yet.
1677 bool delay_slot_is_used = valid_bundle_info(n) &&
1678 C->output()->node_bundling(n)->use_unconditional_delay();
1679 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1680 assert(delay_slot == nullptr, "not expecting delay slot node");
1681 int br_size = n->size(C->regalloc());
1682 int offset = blk_starts[block_num] - current_offset;
1683 if (block_num >= i) {
1684 // Current and following block's offset are not
1685 // finalized yet, adjust distance by the difference
1686 // between calculated and final offsets of current block.
1687 offset -= (blk_starts[i] - blk_offset);
1688 }
1689 // In the following code a nop could be inserted before
1690 // the branch which will increase the backward distance.
1691 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1692 if (needs_padding && offset <= 0)
1693 offset -= nop_size;
1694
1695 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1696 // We've got a winner. Replace this branch.
1697 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1698
1699 // Update the jmp_size.
1700 int new_size = replacement->size(C->regalloc());
1701 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1702 // Insert padding between avoid_back_to_back branches.
1703 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1704 MachNode *nop = new MachNopNode();
1705 block->insert_node(nop, j++);
1706 C->cfg()->map_node_to_block(nop, block);
1707 last_inst++;
1708 nop->emit(masm, C->regalloc());
1709 masm->code()->flush_bundle(true);
1710 current_offset = masm->offset();
1711 }
1712 #ifdef ASSERT
1713 jmp_target[i] = block_num;
1714 jmp_offset[i] = current_offset - blk_offset;
1715 jmp_size[i] = new_size;
1716 jmp_rule[i] = mach->rule();
1717 #endif
1718 block->map_node(replacement, j);
1719 mach->subsume_by(replacement, C);
1720 n = replacement;
1721 mach = replacement;
1722 }
1723 }
1724 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1725 } else if (mach->ideal_Opcode() == Op_Jump) {
1726 for (uint h = 0; h < block->_num_succs; h++) {
1727 Block* succs_block = block->_succs[h];
1728 for (uint j = 1; j < succs_block->num_preds(); j++) {
1729 Node* jpn = succs_block->pred(j);
1730 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1731 uint block_num = succs_block->non_connector()->_pre_order;
1732 Label *blkLabel = &blk_labels[block_num];
1733 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1734 }
1735 }
1736 }
1737 }
1738 #ifdef ASSERT
1739 // Check that oop-store precedes the card-mark
1740 else if (mach->ideal_Opcode() == Op_StoreCM) {
1741 uint storeCM_idx = j;
1742 int count = 0;
1743 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1744 Node *oop_store = mach->in(prec); // Precedence edge
1745 if (oop_store == nullptr) continue;
1746 count++;
1747 uint i4;
1748 for (i4 = 0; i4 < last_inst; ++i4) {
1749 if (block->get_node(i4) == oop_store) {
1750 break;
1751 }
1752 }
1753 // Note: This test can provide a false failure if other precedence
1754 // edges have been added to the storeCMNode.
1755 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1756 }
1757 assert(count > 0, "storeCM expects at least one precedence edge");
1758 }
1759 #endif
1760 else if (!n->is_Proj()) {
1761 // Remember the beginning of the previous instruction, in case
1762 // it's followed by a flag-kill and a null-check. Happens on
1763 // Intel all the time, with add-to-memory kind of opcodes.
1764 previous_offset = current_offset;
1765 }
1766
1767 // Not an else-if!
1768 // If this is a trap based cmp then add its offset to the list.
1769 if (mach->is_TrapBasedCheckNode()) {
1770 inct_starts[inct_cnt++] = current_offset;
1771 }
1772 }
1773
1774 // Verify that there is sufficient space remaining
1775 masm->code()->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1776 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1777 C->record_failure("CodeCache is full");
1778 return;
1779 }
1780
1781 // Save the offset for the listing
1782 #if defined(SUPPORT_OPTO_ASSEMBLY)
1783 if ((node_offsets != nullptr) && (n->_idx < node_offset_limit)) {
1784 node_offsets[n->_idx] = masm->offset();
1785 }
1786 #endif
1787 assert(!C->failing(), "Should not reach here if failing.");
1788
1789 // "Normal" instruction case
1790 DEBUG_ONLY(uint instr_offset = masm->offset());
1791 n->emit(masm, C->regalloc());
1792 current_offset = masm->offset();
1793
1794 // Above we only verified that there is enough space in the instruction section.
1795 // However, the instruction may emit stubs that cause code buffer expansion.
1796 // Bail out here if expansion failed due to a lack of code cache space.
1797 if (C->failing()) {
1798 return;
1799 }
1800
1801 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1802 "ret_addr_offset() not within emitted code");
1803 #ifdef ASSERT
1804 uint n_size = n->size(C->regalloc());
1805 if (n_size < (current_offset-instr_offset)) {
1806 MachNode* mach = n->as_Mach();
1807 n->dump();
1808 mach->dump_format(C->regalloc(), tty);
1809 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1810 Disassembler::decode(masm->code()->insts_begin() + instr_offset, masm->code()->insts_begin() + current_offset + 1, tty);
1811 tty->print_cr(" ------------------- ");
1812 BufferBlob* blob = this->scratch_buffer_blob();
1813 address blob_begin = blob->content_begin();
1814 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1815 assert(false, "wrong size of mach node");
1816 }
1817 #endif
1818 non_safepoints.observe_instruction(n, current_offset);
1819
1820 // mcall is last "call" that can be a safepoint
1821 // record it so we can see if a poll will directly follow it
1822 // in which case we'll need a pad to make the PcDesc sites unique
1823 // see 5010568. This can be slightly inaccurate but conservative
1824 // in the case that return address is not actually at current_offset.
1825 // This is a small price to pay.
1826
1827 if (is_mcall) {
1828 last_call_offset = current_offset;
1829 }
1830
1831 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1832 // Avoid back to back some instructions.
1833 last_avoid_back_to_back_offset = current_offset;
1834 }
1835
1836 // See if this instruction has a delay slot
1837 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1838 guarantee(delay_slot != nullptr, "expecting delay slot node");
1839
1840 // Back up 1 instruction
1841 masm->code()->set_insts_end(masm->code()->insts_end() - Pipeline::instr_unit_size());
1842
1843 // Save the offset for the listing
1844 #if defined(SUPPORT_OPTO_ASSEMBLY)
1845 if ((node_offsets != nullptr) && (delay_slot->_idx < node_offset_limit)) {
1846 node_offsets[delay_slot->_idx] = masm->offset();
1847 }
1848 #endif
1849
1850 // Support a SafePoint in the delay slot
1851 if (delay_slot->is_MachSafePoint()) {
1852 MachNode *mach = delay_slot->as_Mach();
1853 // !!!!! Stubs only need an oopmap right now, so bail out
1854 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == nullptr) {
1855 // Write the oopmap directly to the code blob??!!
1856 delay_slot = nullptr;
1857 continue;
1858 }
1859
1860 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1861 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1862 adjusted_offset);
1863 // Generate an OopMap entry
1864 Process_OopMap_Node(mach, adjusted_offset);
1865 }
1866
1867 // Insert the delay slot instruction
1868 delay_slot->emit(masm, C->regalloc());
1869
1870 // Don't reuse it
1871 delay_slot = nullptr;
1872 }
1873
1874 } // End for all instructions in block
1875
1876 // If the next block is the top of a loop, pad this block out to align
1877 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1878 if (i < nblocks-1) {
1879 Block *nb = C->cfg()->get_block(i + 1);
1880 int padding = nb->alignment_padding(current_offset);
1881 if( padding > 0 ) {
1882 MachNode *nop = new MachNopNode(padding / nop_size);
1883 block->insert_node(nop, block->number_of_nodes());
1884 C->cfg()->map_node_to_block(nop, block);
1885 nop->emit(masm, C->regalloc());
1886 current_offset = masm->offset();
1887 }
1888 }
1889 // Verify that the distance for generated before forward
1890 // short branches is still valid.
1891 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1892
1893 // Save new block start offset
1894 blk_starts[i] = blk_offset;
1895 } // End of for all blocks
1896 blk_starts[nblocks] = current_offset;
1897
1898 non_safepoints.flush_at_end();
1899
1900 // Offset too large?
1901 if (C->failing()) return;
1902
1903 // Define a pseudo-label at the end of the code
1904 masm->bind( blk_labels[nblocks] );
1905
1906 // Compute the size of the first block
1907 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1908
1909 #ifdef ASSERT
1910 for (uint i = 0; i < nblocks; i++) { // For all blocks
1911 if (jmp_target[i] != 0) {
1912 int br_size = jmp_size[i];
1913 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1914 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1915 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]);
1916 assert(false, "Displacement too large for short jmp");
1917 }
1918 }
1919 }
1920 #endif
1921
1922 if (!masm->code()->finalize_stubs()) {
1923 C->record_failure("CodeCache is full");
1924 return;
1925 }
1926
1927 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1928 bs->emit_stubs(*masm->code());
1929 if (C->failing()) return;
1930
1931 // Fill in stubs.
1932 assert(masm->inst_mark() == nullptr, "should be.");
1933 _stub_list.emit(*masm);
1934 if (C->failing()) return;
1935
1936 #ifndef PRODUCT
1937 // Information on the size of the method, without the extraneous code
1938 Scheduling::increment_method_size(masm->offset());
1939 #endif
1940
1941 // ------------------
1942 // Fill in exception table entries.
1943 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1944
1945 // Only java methods have exception handlers and deopt handlers
1946 // class HandlerImpl is platform-specific and defined in the *.ad files.
1947 if (C->method()) {
1948 // Emit the exception handler code.
1949 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(masm));
1950 if (C->failing()) {
1951 return; // CodeBuffer::expand failed
1952 }
1953 // Emit the deopt handler code.
1954 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(masm));
1955
1956 // Emit the MethodHandle deopt handler code (if required).
1957 if (C->has_method_handle_invokes() && !C->failing()) {
1958 // We can use the same code as for the normal deopt handler, we
1959 // just need a different entry point address.
1960 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(masm));
1961 }
1962 }
1963
1964 // One last check for failed CodeBuffer::expand:
1965 if ((masm->code()->blob() == nullptr) || (!CompileBroker::should_compile_new_jobs())) {
1966 C->record_failure("CodeCache is full");
1967 return;
1968 }
1969
1970 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1971 if (C->print_assembly()) {
1972 tty->cr();
1973 tty->print_cr("============================= C2-compiled nmethod ==============================");
1974 }
1975 #endif
1976
1977 #if defined(SUPPORT_OPTO_ASSEMBLY)
1978 // Dump the assembly code, including basic-block numbers
1979 if (C->print_assembly()) {
1980 ttyLocker ttyl; // keep the following output all in one block
1981 if (!VMThread::should_terminate()) { // test this under the tty lock
1982 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1983 // We call them first and write to a stringStream, then we retake the lock to
1984 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1985 ResourceMark rm;
1986 stringStream method_metadata_str;
1987 if (C->method() != nullptr) {
1988 C->method()->print_metadata(&method_metadata_str);
1989 }
1990 stringStream dump_asm_str;
1991 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1992
1993 NoSafepointVerifier nsv;
1994 ttyLocker ttyl2;
1995 // This output goes directly to the tty, not the compiler log.
1996 // To enable tools to match it up with the compilation activity,
1997 // be sure to tag this tty output with the compile ID.
1998 if (xtty != nullptr) {
1999 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
2000 C->is_osr_compilation() ? " compile_kind='osr'" : "");
2001 }
2002 if (C->method() != nullptr) {
2003 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
2004 tty->print_raw(method_metadata_str.freeze());
2005 } else if (C->stub_name() != nullptr) {
2006 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
2007 }
2008 tty->cr();
2009 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
2010 tty->print_raw(dump_asm_str.freeze());
2011 tty->print_cr("--------------------------------------------------------------------------------");
2012 if (xtty != nullptr) {
2013 xtty->tail("opto_assembly");
2014 }
2015 }
2016 }
2017 #endif
2018 }
2019
2020 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
2021 _inc_table.set_size(cnt);
2022
2023 uint inct_cnt = 0;
2024 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2025 Block* block = C->cfg()->get_block(i);
2026 Node *n = nullptr;
2027 int j;
2028
2029 // Find the branch; ignore trailing NOPs.
2030 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
2031 n = block->get_node(j);
2032 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
2033 break;
2034 }
2035 }
2036
2037 // If we didn't find anything, continue
2038 if (j < 0) {
2039 continue;
2040 }
2041
2042 // Compute ExceptionHandlerTable subtable entry and add it
2043 // (skip empty blocks)
2044 if (n->is_Catch()) {
2045
2046 // Get the offset of the return from the call
2047 uint call_return = call_returns[block->_pre_order];
2048 #ifdef ASSERT
2049 assert( call_return > 0, "no call seen for this basic block" );
2050 while (block->get_node(--j)->is_MachProj()) ;
2051 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
2052 #endif
2053 // last instruction is a CatchNode, find it's CatchProjNodes
2054 int nof_succs = block->_num_succs;
2055 // allocate space
2056 GrowableArray<intptr_t> handler_bcis(nof_succs);
2057 GrowableArray<intptr_t> handler_pcos(nof_succs);
2058 // iterate through all successors
2059 for (int j = 0; j < nof_succs; j++) {
2060 Block* s = block->_succs[j];
2061 bool found_p = false;
2062 for (uint k = 1; k < s->num_preds(); k++) {
2063 Node* pk = s->pred(k);
2064 if (pk->is_CatchProj() && pk->in(0) == n) {
2065 const CatchProjNode* p = pk->as_CatchProj();
2066 found_p = true;
2067 // add the corresponding handler bci & pco information
2068 if (p->_con != CatchProjNode::fall_through_index) {
2069 // p leads to an exception handler (and is not fall through)
2070 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2071 // no duplicates, please
2072 if (!handler_bcis.contains(p->handler_bci())) {
2073 uint block_num = s->non_connector()->_pre_order;
2074 handler_bcis.append(p->handler_bci());
2075 handler_pcos.append(blk_labels[block_num].loc_pos());
2076 }
2077 }
2078 }
2079 }
2080 assert(found_p, "no matching predecessor found");
2081 // Note: Due to empty block removal, one block may have
2082 // several CatchProj inputs, from the same Catch.
2083 }
2084
2085 // Set the offset of the return from the call
2086 assert(handler_bcis.find(-1) != -1, "must have default handler");
2087 _handler_table.add_subtable(call_return, &handler_bcis, nullptr, &handler_pcos);
2088 continue;
2089 }
2090
2091 // Handle implicit null exception table updates
2092 if (n->is_MachNullCheck()) {
2093 uint block_num = block->non_connector_successor(0)->_pre_order;
2094 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2095 continue;
2096 }
2097 // Handle implicit exception table updates: trap instructions.
2098 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2099 uint block_num = block->non_connector_successor(0)->_pre_order;
2100 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2101 continue;
2102 }
2103 } // End of for all blocks fill in exception table entries
2104 }
2105
2106 // Static Variables
2107 #ifndef PRODUCT
2108 uint Scheduling::_total_nop_size = 0;
2109 uint Scheduling::_total_method_size = 0;
2110 uint Scheduling::_total_branches = 0;
2111 uint Scheduling::_total_unconditional_delays = 0;
2112 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2113 #endif
2114
2115 // Initializer for class Scheduling
2116
2117 Scheduling::Scheduling(Arena *arena, Compile &compile)
2118 : _arena(arena),
2119 _cfg(compile.cfg()),
2120 _regalloc(compile.regalloc()),
2121 _scheduled(arena),
2122 _available(arena),
2123 _reg_node(arena),
2124 _pinch_free_list(arena),
2125 _next_node(nullptr),
2126 _bundle_instr_count(0),
2127 _bundle_cycle_number(0),
2128 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2129 #ifndef PRODUCT
2130 , _branches(0)
2131 , _unconditional_delays(0)
2132 #endif
2133 {
2134 // Create a MachNopNode
2135 _nop = new MachNopNode();
2136
2137 // Now that the nops are in the array, save the count
2138 // (but allow entries for the nops)
2139 _node_bundling_limit = compile.unique();
2140 uint node_max = _regalloc->node_regs_max_index();
2141
2142 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2143
2144 // This one is persistent within the Compile class
2145 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2146
2147 // Allocate space for fixed-size arrays
2148 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2149 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2150
2151 // Clear the arrays
2152 for (uint i = 0; i < node_max; i++) {
2153 ::new (&_node_bundling_base[i]) Bundle();
2154 }
2155 memset(_uses, 0, node_max * sizeof(short));
2156 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2157
2158 // Clear the bundling information
2159 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2160
2161 // Get the last node
2162 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2163
2164 _next_node = block->get_node(block->number_of_nodes() - 1);
2165 }
2166
2167 #ifndef PRODUCT
2168 // Scheduling destructor
2169 Scheduling::~Scheduling() {
2170 _total_branches += _branches;
2171 _total_unconditional_delays += _unconditional_delays;
2172 }
2173 #endif
2174
2175 // Step ahead "i" cycles
2176 void Scheduling::step(uint i) {
2177
2178 Bundle *bundle = node_bundling(_next_node);
2179 bundle->set_starts_bundle();
2180
2181 // Update the bundle record, but leave the flags information alone
2182 if (_bundle_instr_count > 0) {
2183 bundle->set_instr_count(_bundle_instr_count);
2184 bundle->set_resources_used(_bundle_use.resourcesUsed());
2185 }
2186
2187 // Update the state information
2188 _bundle_instr_count = 0;
2189 _bundle_cycle_number += i;
2190 _bundle_use.step(i);
2191 }
2192
2193 void Scheduling::step_and_clear() {
2194 Bundle *bundle = node_bundling(_next_node);
2195 bundle->set_starts_bundle();
2196
2197 // Update the bundle record
2198 if (_bundle_instr_count > 0) {
2199 bundle->set_instr_count(_bundle_instr_count);
2200 bundle->set_resources_used(_bundle_use.resourcesUsed());
2201
2202 _bundle_cycle_number += 1;
2203 }
2204
2205 // Clear the bundling information
2206 _bundle_instr_count = 0;
2207 _bundle_use.reset();
2208
2209 memcpy(_bundle_use_elements,
2210 Pipeline_Use::elaborated_elements,
2211 sizeof(Pipeline_Use::elaborated_elements));
2212 }
2213
2214 // Perform instruction scheduling and bundling over the sequence of
2215 // instructions in backwards order.
2216 void PhaseOutput::ScheduleAndBundle() {
2217
2218 // Don't optimize this if it isn't a method
2219 if (!C->method())
2220 return;
2221
2222 // Don't optimize this if scheduling is disabled
2223 if (!C->do_scheduling())
2224 return;
2225
2226 // Scheduling code works only with pairs (8 bytes) maximum.
2227 // And when the scalable vector register is used, we may spill/unspill
2228 // the whole reg regardless of the max vector size.
2229 if (C->max_vector_size() > 8 ||
2230 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2231 return;
2232 }
2233
2234 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2235
2236 // Create a data structure for all the scheduling information
2237 Scheduling scheduling(Thread::current()->resource_area(), *C);
2238
2239 // Walk backwards over each basic block, computing the needed alignment
2240 // Walk over all the basic blocks
2241 scheduling.DoScheduling();
2242
2243 #ifndef PRODUCT
2244 if (C->trace_opto_output()) {
2245 // Buffer and print all at once
2246 ResourceMark rm;
2247 stringStream ss;
2248 ss.print("\n---- After ScheduleAndBundle ----\n");
2249 print_scheduling(&ss);
2250 tty->print("%s", ss.as_string());
2251 }
2252 #endif
2253 }
2254
2255 #ifndef PRODUCT
2256 // Separated out so that it can be called directly from debugger
2257 void PhaseOutput::print_scheduling() {
2258 print_scheduling(tty);
2259 }
2260
2261 void PhaseOutput::print_scheduling(outputStream* output_stream) {
2262 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2263 output_stream->print("\nBB#%03d:\n", i);
2264 Block* block = C->cfg()->get_block(i);
2265 for (uint j = 0; j < block->number_of_nodes(); j++) {
2266 Node* n = block->get_node(j);
2267 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2268 output_stream->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2269 n->dump("\n", false, output_stream);
2270 }
2271 }
2272 }
2273 #endif
2274
2275 // See if this node fits into the present instruction bundle
2276 bool Scheduling::NodeFitsInBundle(Node *n) {
2277 uint n_idx = n->_idx;
2278
2279 // If this is the unconditional delay instruction, then it fits
2280 if (n == _unconditional_delay_slot) {
2281 #ifndef PRODUCT
2282 if (_cfg->C->trace_opto_output())
2283 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2284 #endif
2285 return (true);
2286 }
2287
2288 // If the node cannot be scheduled this cycle, skip it
2289 if (_current_latency[n_idx] > _bundle_cycle_number) {
2290 #ifndef PRODUCT
2291 if (_cfg->C->trace_opto_output())
2292 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2293 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2294 #endif
2295 return (false);
2296 }
2297
2298 const Pipeline *node_pipeline = n->pipeline();
2299
2300 uint instruction_count = node_pipeline->instructionCount();
2301 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2302 instruction_count = 0;
2303 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2304 instruction_count++;
2305
2306 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2307 #ifndef PRODUCT
2308 if (_cfg->C->trace_opto_output())
2309 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2310 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2311 #endif
2312 return (false);
2313 }
2314
2315 // Don't allow non-machine nodes to be handled this way
2316 if (!n->is_Mach() && instruction_count == 0)
2317 return (false);
2318
2319 // See if there is any overlap
2320 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2321
2322 if (delay > 0) {
2323 #ifndef PRODUCT
2324 if (_cfg->C->trace_opto_output())
2325 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2326 #endif
2327 return false;
2328 }
2329
2330 #ifndef PRODUCT
2331 if (_cfg->C->trace_opto_output())
2332 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2333 #endif
2334
2335 return true;
2336 }
2337
2338 Node * Scheduling::ChooseNodeToBundle() {
2339 uint siz = _available.size();
2340
2341 if (siz == 0) {
2342
2343 #ifndef PRODUCT
2344 if (_cfg->C->trace_opto_output())
2345 tty->print("# ChooseNodeToBundle: null\n");
2346 #endif
2347 return (nullptr);
2348 }
2349
2350 // Fast path, if only 1 instruction in the bundle
2351 if (siz == 1) {
2352 #ifndef PRODUCT
2353 if (_cfg->C->trace_opto_output()) {
2354 tty->print("# ChooseNodeToBundle (only 1): ");
2355 _available[0]->dump();
2356 }
2357 #endif
2358 return (_available[0]);
2359 }
2360
2361 // Don't bother, if the bundle is already full
2362 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2363 for ( uint i = 0; i < siz; i++ ) {
2364 Node *n = _available[i];
2365
2366 // Skip projections, we'll handle them another way
2367 if (n->is_Proj())
2368 continue;
2369
2370 // This presupposed that instructions are inserted into the
2371 // available list in a legality order; i.e. instructions that
2372 // must be inserted first are at the head of the list
2373 if (NodeFitsInBundle(n)) {
2374 #ifndef PRODUCT
2375 if (_cfg->C->trace_opto_output()) {
2376 tty->print("# ChooseNodeToBundle: ");
2377 n->dump();
2378 }
2379 #endif
2380 return (n);
2381 }
2382 }
2383 }
2384
2385 // Nothing fits in this bundle, choose the highest priority
2386 #ifndef PRODUCT
2387 if (_cfg->C->trace_opto_output()) {
2388 tty->print("# ChooseNodeToBundle: ");
2389 _available[0]->dump();
2390 }
2391 #endif
2392
2393 return _available[0];
2394 }
2395
2396 int Scheduling::compare_two_spill_nodes(Node* first, Node* second) {
2397 assert(first->is_MachSpillCopy() && second->is_MachSpillCopy(), "");
2398
2399 OptoReg::Name first_src_lo = _regalloc->get_reg_first(first->in(1));
2400 OptoReg::Name first_dst_lo = _regalloc->get_reg_first(first);
2401 OptoReg::Name second_src_lo = _regalloc->get_reg_first(second->in(1));
2402 OptoReg::Name second_dst_lo = _regalloc->get_reg_first(second);
2403
2404 // Comparison between stack -> reg and stack -> reg
2405 if (OptoReg::is_stack(first_src_lo) && OptoReg::is_stack(second_src_lo) &&
2406 OptoReg::is_reg(first_dst_lo) && OptoReg::is_reg(second_dst_lo)) {
2407 return _regalloc->reg2offset(first_src_lo) - _regalloc->reg2offset(second_src_lo);
2408 }
2409
2410 // Comparison between reg -> stack and reg -> stack
2411 if (OptoReg::is_stack(first_dst_lo) && OptoReg::is_stack(second_dst_lo) &&
2412 OptoReg::is_reg(first_src_lo) && OptoReg::is_reg(second_src_lo)) {
2413 return _regalloc->reg2offset(first_dst_lo) - _regalloc->reg2offset(second_dst_lo);
2414 }
2415
2416 return 0; // Not comparable
2417 }
2418
2419 void Scheduling::AddNodeToAvailableList(Node *n) {
2420 assert( !n->is_Proj(), "projections never directly made available" );
2421 #ifndef PRODUCT
2422 if (_cfg->C->trace_opto_output()) {
2423 tty->print("# AddNodeToAvailableList: ");
2424 n->dump();
2425 }
2426 #endif
2427
2428 int latency = _current_latency[n->_idx];
2429
2430 // Insert in latency order (insertion sort). If two MachSpillCopyNodes
2431 // for stack spilling or unspilling have the same latency, we sort
2432 // them in the order of stack offset. Some ports (e.g. aarch64) may also
2433 // have more opportunities to do ld/st merging
2434 uint i;
2435 for (i = 0; i < _available.size(); i++) {
2436 if (_current_latency[_available[i]->_idx] > latency) {
2437 break;
2438 } else if (_current_latency[_available[i]->_idx] == latency &&
2439 n->is_MachSpillCopy() && _available[i]->is_MachSpillCopy() &&
2440 compare_two_spill_nodes(n, _available[i]) > 0) {
2441 break;
2442 }
2443 }
2444
2445 // Special Check for compares following branches
2446 if( n->is_Mach() && _scheduled.size() > 0 ) {
2447 int op = n->as_Mach()->ideal_Opcode();
2448 Node *last = _scheduled[0];
2449 if( last->is_MachIf() && last->in(1) == n &&
2450 ( op == Op_CmpI ||
2451 op == Op_CmpU ||
2452 op == Op_CmpUL ||
2453 op == Op_CmpP ||
2454 op == Op_CmpF ||
2455 op == Op_CmpD ||
2456 op == Op_CmpL ) ) {
2457
2458 // Recalculate position, moving to front of same latency
2459 for ( i=0 ; i < _available.size(); i++ )
2460 if (_current_latency[_available[i]->_idx] >= latency)
2461 break;
2462 }
2463 }
2464
2465 // Insert the node in the available list
2466 _available.insert(i, n);
2467
2468 #ifndef PRODUCT
2469 if (_cfg->C->trace_opto_output())
2470 dump_available();
2471 #endif
2472 }
2473
2474 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2475 for ( uint i=0; i < n->len(); i++ ) {
2476 Node *def = n->in(i);
2477 if (!def) continue;
2478 if( def->is_Proj() ) // If this is a machine projection, then
2479 def = def->in(0); // propagate usage thru to the base instruction
2480
2481 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2482 continue;
2483 }
2484
2485 // Compute the latency
2486 uint l = _bundle_cycle_number + n->latency(i);
2487 if (_current_latency[def->_idx] < l)
2488 _current_latency[def->_idx] = l;
2489
2490 // If this does not have uses then schedule it
2491 if ((--_uses[def->_idx]) == 0)
2492 AddNodeToAvailableList(def);
2493 }
2494 }
2495
2496 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2497 #ifndef PRODUCT
2498 if (_cfg->C->trace_opto_output()) {
2499 tty->print("# AddNodeToBundle: ");
2500 n->dump();
2501 }
2502 #endif
2503
2504 // Remove this from the available list
2505 uint i;
2506 for (i = 0; i < _available.size(); i++)
2507 if (_available[i] == n)
2508 break;
2509 assert(i < _available.size(), "entry in _available list not found");
2510 _available.remove(i);
2511
2512 // See if this fits in the current bundle
2513 const Pipeline *node_pipeline = n->pipeline();
2514 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2515
2516 // Check for instructions to be placed in the delay slot. We
2517 // do this before we actually schedule the current instruction,
2518 // because the delay slot follows the current instruction.
2519 if (Pipeline::_branch_has_delay_slot &&
2520 node_pipeline->hasBranchDelay() &&
2521 !_unconditional_delay_slot) {
2522
2523 uint siz = _available.size();
2524
2525 // Conditional branches can support an instruction that
2526 // is unconditionally executed and not dependent by the
2527 // branch, OR a conditionally executed instruction if
2528 // the branch is taken. In practice, this means that
2529 // the first instruction at the branch target is
2530 // copied to the delay slot, and the branch goes to
2531 // the instruction after that at the branch target
2532 if ( n->is_MachBranch() ) {
2533
2534 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2535 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2536
2537 #ifndef PRODUCT
2538 _branches++;
2539 #endif
2540
2541 // At least 1 instruction is on the available list
2542 // that is not dependent on the branch
2543 for (uint i = 0; i < siz; i++) {
2544 Node *d = _available[i];
2545 const Pipeline *avail_pipeline = d->pipeline();
2546
2547 // Don't allow safepoints in the branch shadow, that will
2548 // cause a number of difficulties
2549 if ( avail_pipeline->instructionCount() == 1 &&
2550 !avail_pipeline->hasMultipleBundles() &&
2551 !avail_pipeline->hasBranchDelay() &&
2552 Pipeline::instr_has_unit_size() &&
2553 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2554 NodeFitsInBundle(d) &&
2555 !node_bundling(d)->used_in_delay()) {
2556
2557 if (d->is_Mach() && !d->is_MachSafePoint()) {
2558 // A node that fits in the delay slot was found, so we need to
2559 // set the appropriate bits in the bundle pipeline information so
2560 // that it correctly indicates resource usage. Later, when we
2561 // attempt to add this instruction to the bundle, we will skip
2562 // setting the resource usage.
2563 _unconditional_delay_slot = d;
2564 node_bundling(n)->set_use_unconditional_delay();
2565 node_bundling(d)->set_used_in_unconditional_delay();
2566 _bundle_use.add_usage(avail_pipeline->resourceUse());
2567 _current_latency[d->_idx] = _bundle_cycle_number;
2568 _next_node = d;
2569 ++_bundle_instr_count;
2570 #ifndef PRODUCT
2571 _unconditional_delays++;
2572 #endif
2573 break;
2574 }
2575 }
2576 }
2577 }
2578
2579 // No delay slot, add a nop to the usage
2580 if (!_unconditional_delay_slot) {
2581 // See if adding an instruction in the delay slot will overflow
2582 // the bundle.
2583 if (!NodeFitsInBundle(_nop)) {
2584 #ifndef PRODUCT
2585 if (_cfg->C->trace_opto_output())
2586 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2587 #endif
2588 step(1);
2589 }
2590
2591 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2592 _next_node = _nop;
2593 ++_bundle_instr_count;
2594 }
2595
2596 // See if the instruction in the delay slot requires a
2597 // step of the bundles
2598 if (!NodeFitsInBundle(n)) {
2599 #ifndef PRODUCT
2600 if (_cfg->C->trace_opto_output())
2601 tty->print("# *** STEP(branch won't fit) ***\n");
2602 #endif
2603 // Update the state information
2604 _bundle_instr_count = 0;
2605 _bundle_cycle_number += 1;
2606 _bundle_use.step(1);
2607 }
2608 }
2609
2610 // Get the number of instructions
2611 uint instruction_count = node_pipeline->instructionCount();
2612 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2613 instruction_count = 0;
2614
2615 // Compute the latency information
2616 uint delay = 0;
2617
2618 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2619 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2620 if (relative_latency < 0)
2621 relative_latency = 0;
2622
2623 delay = _bundle_use.full_latency(relative_latency, node_usage);
2624
2625 // Does not fit in this bundle, start a new one
2626 if (delay > 0) {
2627 step(delay);
2628
2629 #ifndef PRODUCT
2630 if (_cfg->C->trace_opto_output())
2631 tty->print("# *** STEP(%d) ***\n", delay);
2632 #endif
2633 }
2634 }
2635
2636 // If this was placed in the delay slot, ignore it
2637 if (n != _unconditional_delay_slot) {
2638
2639 if (delay == 0) {
2640 if (node_pipeline->hasMultipleBundles()) {
2641 #ifndef PRODUCT
2642 if (_cfg->C->trace_opto_output())
2643 tty->print("# *** STEP(multiple instructions) ***\n");
2644 #endif
2645 step(1);
2646 }
2647
2648 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2649 #ifndef PRODUCT
2650 if (_cfg->C->trace_opto_output())
2651 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2652 instruction_count + _bundle_instr_count,
2653 Pipeline::_max_instrs_per_cycle);
2654 #endif
2655 step(1);
2656 }
2657 }
2658
2659 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2660 _bundle_instr_count++;
2661
2662 // Set the node's latency
2663 _current_latency[n->_idx] = _bundle_cycle_number;
2664
2665 // Now merge the functional unit information
2666 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2667 _bundle_use.add_usage(node_usage);
2668
2669 // Increment the number of instructions in this bundle
2670 _bundle_instr_count += instruction_count;
2671
2672 // Remember this node for later
2673 if (n->is_Mach())
2674 _next_node = n;
2675 }
2676
2677 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2678 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2679 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2680 // into the block. All other scheduled nodes get put in the schedule here.
2681 int op = n->Opcode();
2682 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2683 (op != Op_Node && // Not an unused antidepedence node and
2684 // not an unallocated boxlock
2685 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2686
2687 // Push any trailing projections
2688 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2689 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2690 Node *foi = n->fast_out(i);
2691 if( foi->is_Proj() )
2692 _scheduled.push(foi);
2693 }
2694 }
2695
2696 // Put the instruction in the schedule list
2697 _scheduled.push(n);
2698 }
2699
2700 #ifndef PRODUCT
2701 if (_cfg->C->trace_opto_output())
2702 dump_available();
2703 #endif
2704
2705 // Walk all the definitions, decrementing use counts, and
2706 // if a definition has a 0 use count, place it in the available list.
2707 DecrementUseCounts(n,bb);
2708 }
2709
2710 // This method sets the use count within a basic block. We will ignore all
2711 // uses outside the current basic block. As we are doing a backwards walk,
2712 // any node we reach that has a use count of 0 may be scheduled. This also
2713 // avoids the problem of cyclic references from phi nodes, as long as phi
2714 // nodes are at the front of the basic block. This method also initializes
2715 // the available list to the set of instructions that have no uses within this
2716 // basic block.
2717 void Scheduling::ComputeUseCount(const Block *bb) {
2718 #ifndef PRODUCT
2719 if (_cfg->C->trace_opto_output())
2720 tty->print("# -> ComputeUseCount\n");
2721 #endif
2722
2723 // Clear the list of available and scheduled instructions, just in case
2724 _available.clear();
2725 _scheduled.clear();
2726
2727 // No delay slot specified
2728 _unconditional_delay_slot = nullptr;
2729
2730 #ifdef ASSERT
2731 for( uint i=0; i < bb->number_of_nodes(); i++ )
2732 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2733 #endif
2734
2735 // Force the _uses count to never go to zero for unscheduable pieces
2736 // of the block
2737 for( uint k = 0; k < _bb_start; k++ )
2738 _uses[bb->get_node(k)->_idx] = 1;
2739 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2740 _uses[bb->get_node(l)->_idx] = 1;
2741
2742 // Iterate backwards over the instructions in the block. Don't count the
2743 // branch projections at end or the block header instructions.
2744 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2745 Node *n = bb->get_node(j);
2746 if( n->is_Proj() ) continue; // Projections handled another way
2747
2748 // Account for all uses
2749 for ( uint k = 0; k < n->len(); k++ ) {
2750 Node *inp = n->in(k);
2751 if (!inp) continue;
2752 assert(inp != n, "no cycles allowed" );
2753 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2754 if (inp->is_Proj()) { // Skip through Proj's
2755 inp = inp->in(0);
2756 }
2757 ++_uses[inp->_idx]; // Count 1 block-local use
2758 }
2759 }
2760
2761 // If this instruction has a 0 use count, then it is available
2762 if (!_uses[n->_idx]) {
2763 _current_latency[n->_idx] = _bundle_cycle_number;
2764 AddNodeToAvailableList(n);
2765 }
2766
2767 #ifndef PRODUCT
2768 if (_cfg->C->trace_opto_output()) {
2769 tty->print("# uses: %3d: ", _uses[n->_idx]);
2770 n->dump();
2771 }
2772 #endif
2773 }
2774
2775 #ifndef PRODUCT
2776 if (_cfg->C->trace_opto_output())
2777 tty->print("# <- ComputeUseCount\n");
2778 #endif
2779 }
2780
2781 // This routine performs scheduling on each basic block in reverse order,
2782 // using instruction latencies and taking into account function unit
2783 // availability.
2784 void Scheduling::DoScheduling() {
2785 #ifndef PRODUCT
2786 if (_cfg->C->trace_opto_output())
2787 tty->print("# -> DoScheduling\n");
2788 #endif
2789
2790 Block *succ_bb = nullptr;
2791 Block *bb;
2792 Compile* C = Compile::current();
2793
2794 // Walk over all the basic blocks in reverse order
2795 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2796 bb = _cfg->get_block(i);
2797
2798 #ifndef PRODUCT
2799 if (_cfg->C->trace_opto_output()) {
2800 tty->print("# Schedule BB#%03d (initial)\n", i);
2801 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2802 bb->get_node(j)->dump();
2803 }
2804 }
2805 #endif
2806
2807 // On the head node, skip processing
2808 if (bb == _cfg->get_root_block()) {
2809 continue;
2810 }
2811
2812 // Skip empty, connector blocks
2813 if (bb->is_connector())
2814 continue;
2815
2816 // If the following block is not the sole successor of
2817 // this one, then reset the pipeline information
2818 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2819 #ifndef PRODUCT
2820 if (_cfg->C->trace_opto_output()) {
2821 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2822 _next_node->_idx, _bundle_instr_count);
2823 }
2824 #endif
2825 step_and_clear();
2826 }
2827
2828 // Leave untouched the starting instruction, any Phis, a CreateEx node
2829 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2830 _bb_end = bb->number_of_nodes()-1;
2831 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2832 Node *n = bb->get_node(_bb_start);
2833 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2834 // Also, MachIdealNodes do not get scheduled
2835 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2836 MachNode *mach = n->as_Mach();
2837 int iop = mach->ideal_Opcode();
2838 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2839 if( iop == Op_Con ) continue; // Do not schedule Top
2840 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2841 mach->pipeline() == MachNode::pipeline_class() &&
2842 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2843 continue;
2844 break; // Funny loop structure to be sure...
2845 }
2846 // Compute last "interesting" instruction in block - last instruction we
2847 // might schedule. _bb_end points just after last schedulable inst. We
2848 // normally schedule conditional branches (despite them being forced last
2849 // in the block), because they have delay slots we can fill. Calls all
2850 // have their delay slots filled in the template expansions, so we don't
2851 // bother scheduling them.
2852 Node *last = bb->get_node(_bb_end);
2853 // Ignore trailing NOPs.
2854 while (_bb_end > 0 && last->is_Mach() &&
2855 last->as_Mach()->ideal_Opcode() == Op_Con) {
2856 last = bb->get_node(--_bb_end);
2857 }
2858 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2859 if( last->is_Catch() ||
2860 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2861 // There might be a prior call. Skip it.
2862 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2863 } else if( last->is_MachNullCheck() ) {
2864 // Backup so the last null-checked memory instruction is
2865 // outside the schedulable range. Skip over the nullcheck,
2866 // projection, and the memory nodes.
2867 Node *mem = last->in(1);
2868 do {
2869 _bb_end--;
2870 } while (mem != bb->get_node(_bb_end));
2871 } else {
2872 // Set _bb_end to point after last schedulable inst.
2873 _bb_end++;
2874 }
2875
2876 assert( _bb_start <= _bb_end, "inverted block ends" );
2877
2878 // Compute the register antidependencies for the basic block
2879 ComputeRegisterAntidependencies(bb);
2880 if (C->failing()) return; // too many D-U pinch points
2881
2882 // Compute the usage within the block, and set the list of all nodes
2883 // in the block that have no uses within the block.
2884 ComputeUseCount(bb);
2885
2886 // Schedule the remaining instructions in the block
2887 while ( _available.size() > 0 ) {
2888 Node *n = ChooseNodeToBundle();
2889 guarantee(n != nullptr, "no nodes available");
2890 AddNodeToBundle(n,bb);
2891 }
2892
2893 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2894 #ifdef ASSERT
2895 for( uint l = _bb_start; l < _bb_end; l++ ) {
2896 Node *n = bb->get_node(l);
2897 uint m;
2898 for( m = 0; m < _bb_end-_bb_start; m++ )
2899 if( _scheduled[m] == n )
2900 break;
2901 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2902 }
2903 #endif
2904
2905 // Now copy the instructions (in reverse order) back to the block
2906 for ( uint k = _bb_start; k < _bb_end; k++ )
2907 bb->map_node(_scheduled[_bb_end-k-1], k);
2908
2909 #ifndef PRODUCT
2910 if (_cfg->C->trace_opto_output()) {
2911 tty->print("# Schedule BB#%03d (final)\n", i);
2912 uint current = 0;
2913 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2914 Node *n = bb->get_node(j);
2915 if( valid_bundle_info(n) ) {
2916 Bundle *bundle = node_bundling(n);
2917 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2918 tty->print("*** Bundle: ");
2919 bundle->dump();
2920 }
2921 n->dump();
2922 }
2923 }
2924 }
2925 #endif
2926 #ifdef ASSERT
2927 verify_good_schedule(bb,"after block local scheduling");
2928 #endif
2929 }
2930
2931 #ifndef PRODUCT
2932 if (_cfg->C->trace_opto_output())
2933 tty->print("# <- DoScheduling\n");
2934 #endif
2935
2936 // Record final node-bundling array location
2937 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2938
2939 } // end DoScheduling
2940
2941 // Verify that no live-range used in the block is killed in the block by a
2942 // wrong DEF. This doesn't verify live-ranges that span blocks.
2943
2944 // Check for edge existence. Used to avoid adding redundant precedence edges.
2945 static bool edge_from_to( Node *from, Node *to ) {
2946 for( uint i=0; i<from->len(); i++ )
2947 if( from->in(i) == to )
2948 return true;
2949 return false;
2950 }
2951
2952 #ifdef ASSERT
2953 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2954 // Check for bad kills
2955 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2956 Node *prior_use = _reg_node[def];
2957 if( prior_use && !edge_from_to(prior_use,n) ) {
2958 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2959 n->dump();
2960 tty->print_cr("...");
2961 prior_use->dump();
2962 assert(edge_from_to(prior_use,n), "%s", msg);
2963 }
2964 _reg_node.map(def,nullptr); // Kill live USEs
2965 }
2966 }
2967
2968 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2969
2970 // Zap to something reasonable for the verify code
2971 _reg_node.clear();
2972
2973 // Walk over the block backwards. Check to make sure each DEF doesn't
2974 // kill a live value (other than the one it's supposed to). Add each
2975 // USE to the live set.
2976 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2977 Node *n = b->get_node(i);
2978 int n_op = n->Opcode();
2979 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2980 // Fat-proj kills a slew of registers
2981 RegMaskIterator rmi(n->out_RegMask());
2982 while (rmi.has_next()) {
2983 OptoReg::Name kill = rmi.next();
2984 verify_do_def(n, kill, msg);
2985 }
2986 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2987 // Get DEF'd registers the normal way
2988 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2989 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2990 }
2991
2992 // Now make all USEs live
2993 for( uint i=1; i<n->req(); i++ ) {
2994 Node *def = n->in(i);
2995 assert(def != nullptr, "input edge required");
2996 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2997 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2998 if( OptoReg::is_valid(reg_lo) ) {
2999 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
3000 _reg_node.map(reg_lo,n);
3001 }
3002 if( OptoReg::is_valid(reg_hi) ) {
3003 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
3004 _reg_node.map(reg_hi,n);
3005 }
3006 }
3007
3008 }
3009
3010 // Zap to something reasonable for the Antidependence code
3011 _reg_node.clear();
3012 }
3013 #endif
3014
3015 // Conditionally add precedence edges. Avoid putting edges on Projs.
3016 static void add_prec_edge_from_to( Node *from, Node *to ) {
3017 if( from->is_Proj() ) { // Put precedence edge on Proj's input
3018 assert( from->req() == 1 && (from->len() == 1 || from->in(1) == nullptr), "no precedence edges on projections" );
3019 from = from->in(0);
3020 }
3021 if( from != to && // No cycles (for things like LD L0,[L0+4] )
3022 !edge_from_to( from, to ) ) // Avoid duplicate edge
3023 from->add_prec(to);
3024 }
3025
3026 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
3027 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
3028 return;
3029
3030 if (OptoReg::is_reg(def_reg)) {
3031 VMReg vmreg = OptoReg::as_VMReg(def_reg);
3032 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
3033 // This is one of the high slots of a vector register.
3034 // ScheduleAndBundle already checked there are no live wide
3035 // vectors in this method so it can be safely ignored.
3036 return;
3037 }
3038 }
3039
3040 Node *pinch = _reg_node[def_reg]; // Get pinch point
3041 if ((pinch == nullptr) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
3042 is_def ) { // Check for a true def (not a kill)
3043 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
3044 return;
3045 }
3046
3047 Node *kill = def; // Rename 'def' to more descriptive 'kill'
3048 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
3049
3050 // After some number of kills there _may_ be a later def
3051 Node *later_def = nullptr;
3052
3053 Compile* C = Compile::current();
3054
3055 // Finding a kill requires a real pinch-point.
3056 // Check for not already having a pinch-point.
3057 // Pinch points are Op_Node's.
3058 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
3059 later_def = pinch; // Must be def/kill as optimistic pinch-point
3060 if ( _pinch_free_list.size() > 0) {
3061 pinch = _pinch_free_list.pop();
3062 } else {
3063 pinch = new Node(1); // Pinch point to-be
3064 }
3065 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
3066 DEBUG_ONLY( pinch->dump(); );
3067 assert(false, "too many D-U pinch points: %d >= %d", pinch->_idx, _regalloc->node_regs_max_index());
3068 _cfg->C->record_method_not_compilable("too many D-U pinch points");
3069 return;
3070 }
3071 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3072 _reg_node.map(def_reg,pinch); // Record pinch-point
3073 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3074 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3075 pinch->init_req(0, C->top()); // set not null for the next call
3076 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3077 later_def = nullptr; // and no later def
3078 }
3079 pinch->set_req(0,later_def); // Hook later def so we can find it
3080 } else { // Else have valid pinch point
3081 if( pinch->in(0) ) // If there is a later-def
3082 later_def = pinch->in(0); // Get it
3083 }
3084
3085 // Add output-dependence edge from later def to kill
3086 if( later_def ) // If there is some original def
3087 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3088
3089 // See if current kill is also a use, and so is forced to be the pinch-point.
3090 if( pinch->Opcode() == Op_Node ) {
3091 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3092 for( uint i=1; i<uses->req(); i++ ) {
3093 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3094 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3095 // Yes, found a use/kill pinch-point
3096 pinch->set_req(0,nullptr); //
3097 pinch->replace_by(kill); // Move anti-dep edges up
3098 pinch = kill;
3099 _reg_node.map(def_reg,pinch);
3100 return;
3101 }
3102 }
3103 }
3104
3105 // Add edge from kill to pinch-point
3106 add_prec_edge_from_to(kill,pinch);
3107 }
3108
3109 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3110 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3111 return;
3112 Node *pinch = _reg_node[use_reg]; // Get pinch point
3113 // Check for no later def_reg/kill in block
3114 if ((pinch != nullptr) && _cfg->get_block_for_node(pinch) == b &&
3115 // Use has to be block-local as well
3116 _cfg->get_block_for_node(use) == b) {
3117 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3118 pinch->req() == 1 ) { // pinch not yet in block?
3119 pinch->del_req(0); // yank pointer to later-def, also set flag
3120 // Insert the pinch-point in the block just after the last use
3121 b->insert_node(pinch, b->find_node(use) + 1);
3122 _bb_end++; // Increase size scheduled region in block
3123 }
3124
3125 add_prec_edge_from_to(pinch,use);
3126 }
3127 }
3128
3129 // We insert antidependences between the reads and following write of
3130 // allocated registers to prevent illegal code motion. Hopefully, the
3131 // number of added references should be fairly small, especially as we
3132 // are only adding references within the current basic block.
3133 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3134
3135 #ifdef ASSERT
3136 verify_good_schedule(b,"before block local scheduling");
3137 #endif
3138
3139 // A valid schedule, for each register independently, is an endless cycle
3140 // of: a def, then some uses (connected to the def by true dependencies),
3141 // then some kills (defs with no uses), finally the cycle repeats with a new
3142 // def. The uses are allowed to float relative to each other, as are the
3143 // kills. No use is allowed to slide past a kill (or def). This requires
3144 // antidependencies between all uses of a single def and all kills that
3145 // follow, up to the next def. More edges are redundant, because later defs
3146 // & kills are already serialized with true or antidependencies. To keep
3147 // the edge count down, we add a 'pinch point' node if there's more than
3148 // one use or more than one kill/def.
3149
3150 // We add dependencies in one bottom-up pass.
3151
3152 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3153
3154 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3155 // register. If not, we record the DEF/KILL in _reg_node, the
3156 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3157 // "pinch point", a new Node that's in the graph but not in the block.
3158 // We put edges from the prior and current DEF/KILLs to the pinch point.
3159 // We put the pinch point in _reg_node. If there's already a pinch point
3160 // we merely add an edge from the current DEF/KILL to the pinch point.
3161
3162 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3163 // put an edge from the pinch point to the USE.
3164
3165 // To be expedient, the _reg_node array is pre-allocated for the whole
3166 // compilation. _reg_node is lazily initialized; it either contains a null,
3167 // or a valid def/kill/pinch-point, or a leftover node from some prior
3168 // block. Leftover node from some prior block is treated like a null (no
3169 // prior def, so no anti-dependence needed). Valid def is distinguished by
3170 // it being in the current block.
3171 bool fat_proj_seen = false;
3172 uint last_safept = _bb_end-1;
3173 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : nullptr;
3174 Node* last_safept_node = end_node;
3175 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3176 Node *n = b->get_node(i);
3177 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3178 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3179 // Fat-proj kills a slew of registers
3180 // This can add edges to 'n' and obscure whether or not it was a def,
3181 // hence the is_def flag.
3182 fat_proj_seen = true;
3183 RegMaskIterator rmi(n->out_RegMask());
3184 while (rmi.has_next()) {
3185 OptoReg::Name kill = rmi.next();
3186 anti_do_def(b, n, kill, is_def);
3187 }
3188 } else {
3189 // Get DEF'd registers the normal way
3190 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3191 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3192 }
3193
3194 // Kill projections on a branch should appear to occur on the
3195 // branch, not afterwards, so grab the masks from the projections
3196 // and process them.
3197 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3198 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3199 Node* use = n->fast_out(i);
3200 if (use->is_Proj()) {
3201 RegMaskIterator rmi(use->out_RegMask());
3202 while (rmi.has_next()) {
3203 OptoReg::Name kill = rmi.next();
3204 anti_do_def(b, n, kill, false);
3205 }
3206 }
3207 }
3208 }
3209
3210 // Check each register used by this instruction for a following DEF/KILL
3211 // that must occur afterward and requires an anti-dependence edge.
3212 for( uint j=0; j<n->req(); j++ ) {
3213 Node *def = n->in(j);
3214 if( def ) {
3215 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3216 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3217 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3218 }
3219 }
3220 // Do not allow defs of new derived values to float above GC
3221 // points unless the base is definitely available at the GC point.
3222
3223 Node *m = b->get_node(i);
3224
3225 // Add precedence edge from following safepoint to use of derived pointer
3226 if( last_safept_node != end_node &&
3227 m != last_safept_node) {
3228 for (uint k = 1; k < m->req(); k++) {
3229 const Type *t = m->in(k)->bottom_type();
3230 if( t->isa_oop_ptr() &&
3231 t->is_ptr()->offset() != 0 ) {
3232 last_safept_node->add_prec( m );
3233 break;
3234 }
3235 }
3236
3237 // Do not allow a CheckCastPP node whose input is a raw pointer to
3238 // float past a safepoint. This can occur when a buffered inline
3239 // type is allocated in a loop and the CheckCastPP from that
3240 // allocation is reused outside the loop. If the use inside the
3241 // loop is scalarized the CheckCastPP will no longer be connected
3242 // to the loop safepoint. See JDK-8264340.
3243 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3244 Node *def = m->in(1);
3245 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
3246 last_safept_node->add_prec(m);
3247 }
3248 }
3249 }
3250
3251 if( n->jvms() ) { // Precedence edge from derived to safept
3252 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3253 if( b->get_node(last_safept) != last_safept_node ) {
3254 last_safept = b->find_node(last_safept_node);
3255 }
3256 for( uint j=last_safept; j > i; j-- ) {
3257 Node *mach = b->get_node(j);
3258 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3259 mach->add_prec( n );
3260 }
3261 last_safept = i;
3262 last_safept_node = m;
3263 }
3264 }
3265
3266 if (fat_proj_seen) {
3267 // Garbage collect pinch nodes that were not consumed.
3268 // They are usually created by a fat kill MachProj for a call.
3269 garbage_collect_pinch_nodes();
3270 }
3271 }
3272
3273 // Garbage collect pinch nodes for reuse by other blocks.
3274 //
3275 // The block scheduler's insertion of anti-dependence
3276 // edges creates many pinch nodes when the block contains
3277 // 2 or more Calls. A pinch node is used to prevent a
3278 // combinatorial explosion of edges. If a set of kills for a
3279 // register is anti-dependent on a set of uses (or defs), rather
3280 // than adding an edge in the graph between each pair of kill
3281 // and use (or def), a pinch is inserted between them:
3282 //
3283 // use1 use2 use3
3284 // \ | /
3285 // \ | /
3286 // pinch
3287 // / | \
3288 // / | \
3289 // kill1 kill2 kill3
3290 //
3291 // One pinch node is created per register killed when
3292 // the second call is encountered during a backwards pass
3293 // over the block. Most of these pinch nodes are never
3294 // wired into the graph because the register is never
3295 // used or def'ed in the block.
3296 //
3297 void Scheduling::garbage_collect_pinch_nodes() {
3298 #ifndef PRODUCT
3299 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3300 #endif
3301 int trace_cnt = 0;
3302 for (uint k = 0; k < _reg_node.max(); k++) {
3303 Node* pinch = _reg_node[k];
3304 if ((pinch != nullptr) && pinch->Opcode() == Op_Node &&
3305 // no predecence input edges
3306 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == nullptr) ) {
3307 cleanup_pinch(pinch);
3308 _pinch_free_list.push(pinch);
3309 _reg_node.map(k, nullptr);
3310 #ifndef PRODUCT
3311 if (_cfg->C->trace_opto_output()) {
3312 trace_cnt++;
3313 if (trace_cnt > 40) {
3314 tty->print("\n");
3315 trace_cnt = 0;
3316 }
3317 tty->print(" %d", pinch->_idx);
3318 }
3319 #endif
3320 }
3321 }
3322 #ifndef PRODUCT
3323 if (_cfg->C->trace_opto_output()) tty->print("\n");
3324 #endif
3325 }
3326
3327 // Clean up a pinch node for reuse.
3328 void Scheduling::cleanup_pinch( Node *pinch ) {
3329 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3330
3331 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3332 Node* use = pinch->last_out(i);
3333 uint uses_found = 0;
3334 for (uint j = use->req(); j < use->len(); j++) {
3335 if (use->in(j) == pinch) {
3336 use->rm_prec(j);
3337 uses_found++;
3338 }
3339 }
3340 assert(uses_found > 0, "must be a precedence edge");
3341 i -= uses_found; // we deleted 1 or more copies of this edge
3342 }
3343 // May have a later_def entry
3344 pinch->set_req(0, nullptr);
3345 }
3346
3347 #ifndef PRODUCT
3348
3349 void Scheduling::dump_available() const {
3350 tty->print("#Availist ");
3351 for (uint i = 0; i < _available.size(); i++)
3352 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3353 tty->cr();
3354 }
3355
3356 // Print Scheduling Statistics
3357 void Scheduling::print_statistics() {
3358 // Print the size added by nops for bundling
3359 tty->print("Nops added %d bytes to total of %d bytes",
3360 _total_nop_size, _total_method_size);
3361 if (_total_method_size > 0)
3362 tty->print(", for %.2f%%",
3363 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3364 tty->print("\n");
3365
3366 // Print the number of branch shadows filled
3367 if (Pipeline::_branch_has_delay_slot) {
3368 tty->print("Of %d branches, %d had unconditional delay slots filled",
3369 _total_branches, _total_unconditional_delays);
3370 if (_total_branches > 0)
3371 tty->print(", for %.2f%%",
3372 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3373 tty->print("\n");
3374 }
3375
3376 uint total_instructions = 0, total_bundles = 0;
3377
3378 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3379 uint bundle_count = _total_instructions_per_bundle[i];
3380 total_instructions += bundle_count * i;
3381 total_bundles += bundle_count;
3382 }
3383
3384 if (total_bundles > 0)
3385 tty->print("Average ILP (excluding nops) is %.2f\n",
3386 ((double)total_instructions) / ((double)total_bundles));
3387 }
3388 #endif
3389
3390 //-----------------------init_scratch_buffer_blob------------------------------
3391 // Construct a temporary BufferBlob and cache it for this compile.
3392 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3393 // If there is already a scratch buffer blob allocated and the
3394 // constant section is big enough, use it. Otherwise free the
3395 // current and allocate a new one.
3396 BufferBlob* blob = scratch_buffer_blob();
3397 if ((blob != nullptr) && (const_size <= _scratch_const_size)) {
3398 // Use the current blob.
3399 } else {
3400 if (blob != nullptr) {
3401 BufferBlob::free(blob);
3402 }
3403
3404 ResourceMark rm;
3405 _scratch_const_size = const_size;
3406 int size = C2Compiler::initial_code_buffer_size(const_size);
3407 if (C->has_scalarized_args()) {
3408 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3409 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3410 ciMethod* method = C->method();
3411 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3412 int arg_num = 0;
3413 if (!method->is_static()) {
3414 if (method->is_scalarized_arg(arg_num)) {
3415 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3416 }
3417 arg_num++;
3418 }
3419 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3420 if (method->is_scalarized_arg(arg_num)) {
3421 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3422 }
3423 arg_num++;
3424 }
3425 }
3426 blob = BufferBlob::create("Compile::scratch_buffer", size);
3427 // Record the buffer blob for next time.
3428 set_scratch_buffer_blob(blob);
3429 // Have we run out of code space?
3430 if (scratch_buffer_blob() == nullptr) {
3431 // Let CompilerBroker disable further compilations.
3432 C->record_failure("Not enough space for scratch buffer in CodeCache");
3433 return;
3434 }
3435 }
3436
3437 // Initialize the relocation buffers
3438 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3439 set_scratch_locs_memory(locs_buf);
3440 }
3441
3442
3443 //-----------------------scratch_emit_size-------------------------------------
3444 // Helper function that computes size by emitting code
3445 uint PhaseOutput::scratch_emit_size(const Node* n) {
3446 // Start scratch_emit_size section.
3447 set_in_scratch_emit_size(true);
3448
3449 // Emit into a trash buffer and count bytes emitted.
3450 // This is a pretty expensive way to compute a size,
3451 // but it works well enough if seldom used.
3452 // All common fixed-size instructions are given a size
3453 // method by the AD file.
3454 // Note that the scratch buffer blob and locs memory are
3455 // allocated at the beginning of the compile task, and
3456 // may be shared by several calls to scratch_emit_size.
3457 // The allocation of the scratch buffer blob is particularly
3458 // expensive, since it has to grab the code cache lock.
3459 BufferBlob* blob = this->scratch_buffer_blob();
3460 assert(blob != nullptr, "Initialize BufferBlob at start");
3461 assert(blob->size() > MAX_inst_size, "sanity");
3462 relocInfo* locs_buf = scratch_locs_memory();
3463 address blob_begin = blob->content_begin();
3464 address blob_end = (address)locs_buf;
3465 assert(blob->contains(blob_end), "sanity");
3466 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3467 buf.initialize_consts_size(_scratch_const_size);
3468 buf.initialize_stubs_size(MAX_stubs_size);
3469 assert(locs_buf != nullptr, "sanity");
3470 int lsize = MAX_locs_size / 3;
3471 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3472 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3473 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3474 // Mark as scratch buffer.
3475 buf.consts()->set_scratch_emit();
3476 buf.insts()->set_scratch_emit();
3477 buf.stubs()->set_scratch_emit();
3478
3479 // Do the emission.
3480
3481 Label fakeL; // Fake label for branch instructions.
3482 Label* saveL = nullptr;
3483 uint save_bnum = 0;
3484 bool is_branch = n->is_MachBranch();
3485 C2_MacroAssembler masm(&buf);
3486 masm.bind(fakeL);
3487 if (is_branch) {
3488 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3489 n->as_MachBranch()->label_set(&fakeL, 0);
3490 }
3491 n->emit(&masm, C->regalloc());
3492
3493 // Emitting into the scratch buffer should not fail
3494 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3495
3496 // Restore label.
3497 if (is_branch) {
3498 n->as_MachBranch()->label_set(saveL, save_bnum);
3499 }
3500
3501 // End scratch_emit_size section.
3502 set_in_scratch_emit_size(false);
3503
3504 return buf.insts_size();
3505 }
3506
3507 void PhaseOutput::install() {
3508 if (!C->should_install_code()) {
3509 return;
3510 } else if (C->stub_function() != nullptr) {
3511 install_stub(C->stub_name());
3512 } else {
3513 install_code(C->method(),
3514 C->entry_bci(),
3515 CompileBroker::compiler2(),
3516 C->has_unsafe_access(),
3517 SharedRuntime::is_wide_vector(C->max_vector_size()));
3518 }
3519 }
3520
3521 void PhaseOutput::install_code(ciMethod* target,
3522 int entry_bci,
3523 AbstractCompiler* compiler,
3524 bool has_unsafe_access,
3525 bool has_wide_vectors) {
3526 // Check if we want to skip execution of all compiled code.
3527 {
3528 #ifndef PRODUCT
3529 if (OptoNoExecute) {
3530 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3531 return;
3532 }
3533 #endif
3534 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3535
3536 if (C->is_osr_compilation()) {
3537 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3538 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3539 } else {
3540 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3541 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3542 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3543 }
3544 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3545 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3546 }
3547 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3548 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3549 }
3550 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3551 }
3552
3553 C->env()->register_method(target,
3554 entry_bci,
3555 &_code_offsets,
3556 _orig_pc_slot_offset_in_bytes,
3557 code_buffer(),
3558 frame_size_in_words(),
3559 _oop_map_set,
3560 &_handler_table,
3561 inc_table(),
3562 compiler,
3563 has_unsafe_access,
3564 SharedRuntime::is_wide_vector(C->max_vector_size()),
3565 C->has_monitors(),
3566 C->has_scoped_access(),
3567 0);
3568
3569 if (C->log() != nullptr) { // Print code cache state into compiler log
3570 C->log()->code_cache_state();
3571 }
3572 }
3573 }
3574 void PhaseOutput::install_stub(const char* stub_name) {
3575 // Entry point will be accessed using stub_entry_point();
3576 if (code_buffer() == nullptr) {
3577 Matcher::soft_match_failure();
3578 } else {
3579 if (PrintAssembly && (WizardMode || Verbose))
3580 tty->print_cr("### Stub::%s", stub_name);
3581
3582 if (!C->failing()) {
3583 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3584
3585 // Make the NMethod
3586 // For now we mark the frame as never safe for profile stackwalking
3587 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3588 code_buffer(),
3589 CodeOffsets::frame_never_safe,
3590 // _code_offsets.value(CodeOffsets::Frame_Complete),
3591 frame_size_in_words(),
3592 oop_map_set(),
3593 false);
3594 assert(rs != nullptr && rs->is_runtime_stub(), "sanity check");
3595
3596 C->set_stub_entry_point(rs->entry_point());
3597 }
3598 }
3599 }
3600
3601 // Support for bundling info
3602 Bundle* PhaseOutput::node_bundling(const Node *n) {
3603 assert(valid_bundle_info(n), "oob");
3604 return &_node_bundling_base[n->_idx];
3605 }
3606
3607 bool PhaseOutput::valid_bundle_info(const Node *n) {
3608 return (_node_bundling_limit > n->_idx);
3609 }
3610
3611 //------------------------------frame_size_in_words-----------------------------
3612 // frame_slots in units of words
3613 int PhaseOutput::frame_size_in_words() const {
3614 // shift is 0 in LP32 and 1 in LP64
3615 const int shift = (LogBytesPerWord - LogBytesPerInt);
3616 int words = _frame_slots >> shift;
3617 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3618 return words;
3619 }
3620
3621 // To bang the stack of this compiled method we use the stack size
3622 // that the interpreter would need in case of a deoptimization. This
3623 // removes the need to bang the stack in the deoptimization blob which
3624 // in turn simplifies stack overflow handling.
3625 int PhaseOutput::bang_size_in_bytes() const {
3626 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3627 }
3628
3629 //------------------------------dump_asm---------------------------------------
3630 // Dump formatted assembly
3631 #if defined(SUPPORT_OPTO_ASSEMBLY)
3632 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3633
3634 int pc_digits = 3; // #chars required for pc
3635 int sb_chars = 3; // #chars for "start bundle" indicator
3636 int tab_size = 8;
3637 if (pcs != nullptr) {
3638 int max_pc = 0;
3639 for (uint i = 0; i < pc_limit; i++) {
3640 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3641 }
3642 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3643 }
3644 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3645
3646 bool cut_short = false;
3647 st->print_cr("#");
3648 st->print("# "); C->tf()->dump_on(st); st->cr();
3649 st->print_cr("#");
3650
3651 // For all blocks
3652 int pc = 0x0; // Program counter
3653 char starts_bundle = ' ';
3654 C->regalloc()->dump_frame();
3655
3656 Node *n = nullptr;
3657 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3658 if (VMThread::should_terminate()) {
3659 cut_short = true;
3660 break;
3661 }
3662 Block* block = C->cfg()->get_block(i);
3663 if (block->is_connector() && !Verbose) {
3664 continue;
3665 }
3666 n = block->head();
3667 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3668 pc = pcs[n->_idx];
3669 st->print("%*.*x", pc_digits, pc_digits, pc);
3670 }
3671 st->fill_to(prefix_len);
3672 block->dump_head(C->cfg(), st);
3673 if (block->is_connector()) {
3674 st->fill_to(prefix_len);
3675 st->print_cr("# Empty connector block");
3676 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3677 st->fill_to(prefix_len);
3678 st->print_cr("# Block is sole successor of call");
3679 }
3680
3681 // For all instructions
3682 Node *delay = nullptr;
3683 for (uint j = 0; j < block->number_of_nodes(); j++) {
3684 if (VMThread::should_terminate()) {
3685 cut_short = true;
3686 break;
3687 }
3688 n = block->get_node(j);
3689 if (valid_bundle_info(n)) {
3690 Bundle* bundle = node_bundling(n);
3691 if (bundle->used_in_unconditional_delay()) {
3692 delay = n;
3693 continue;
3694 }
3695 if (bundle->starts_bundle()) {
3696 starts_bundle = '+';
3697 }
3698 }
3699
3700 if (WizardMode) {
3701 n->dump();
3702 }
3703
3704 if( !n->is_Region() && // Dont print in the Assembly
3705 !n->is_Phi() && // a few noisely useless nodes
3706 !n->is_Proj() &&
3707 !n->is_MachTemp() &&
3708 !n->is_SafePointScalarObject() &&
3709 !n->is_Catch() && // Would be nice to print exception table targets
3710 !n->is_MergeMem() && // Not very interesting
3711 !n->is_top() && // Debug info table constants
3712 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3713 ) {
3714 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3715 pc = pcs[n->_idx];
3716 st->print("%*.*x", pc_digits, pc_digits, pc);
3717 } else {
3718 st->fill_to(pc_digits);
3719 }
3720 st->print(" %c ", starts_bundle);
3721 starts_bundle = ' ';
3722 st->fill_to(prefix_len);
3723 n->format(C->regalloc(), st);
3724 st->cr();
3725 }
3726
3727 // If we have an instruction with a delay slot, and have seen a delay,
3728 // then back up and print it
3729 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3730 // Coverity finding - Explicit null dereferenced.
3731 guarantee(delay != nullptr, "no unconditional delay instruction");
3732 if (WizardMode) delay->dump();
3733
3734 if (node_bundling(delay)->starts_bundle())
3735 starts_bundle = '+';
3736 if ((pcs != nullptr) && (n->_idx < pc_limit)) {
3737 pc = pcs[n->_idx];
3738 st->print("%*.*x", pc_digits, pc_digits, pc);
3739 } else {
3740 st->fill_to(pc_digits);
3741 }
3742 st->print(" %c ", starts_bundle);
3743 starts_bundle = ' ';
3744 st->fill_to(prefix_len);
3745 delay->format(C->regalloc(), st);
3746 st->cr();
3747 delay = nullptr;
3748 }
3749
3750 // Dump the exception table as well
3751 if( n->is_Catch() && (Verbose || WizardMode) ) {
3752 // Print the exception table for this offset
3753 _handler_table.print_subtable_for(pc);
3754 }
3755 st->bol(); // Make sure we start on a new line
3756 }
3757 st->cr(); // one empty line between blocks
3758 assert(cut_short || delay == nullptr, "no unconditional delay branch");
3759 } // End of per-block dump
3760
3761 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3762 }
3763 #endif
3764
3765 #ifndef PRODUCT
3766 void PhaseOutput::print_statistics() {
3767 Scheduling::print_statistics();
3768 }
3769 #endif