1 /*
2 * Copyright (c) 1998, 2022, 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 // Add a node to the list of available nodes
174 void AddNodeToAvailableList(Node *n);
175
176 // Compute the local use count for the nodes in a block, and compute
177 // the list of instructions with no uses in the block as available
178 void ComputeUseCount(const Block *bb);
179
180 // Choose an instruction from the available list to add to the bundle
181 Node * ChooseNodeToBundle();
182
183 // See if this Node fits into the currently accumulating bundle
184 bool NodeFitsInBundle(Node *n);
185
186 // Decrement the use count for a node
187 void DecrementUseCounts(Node *n, const Block *bb);
188
189 // Garbage collect pinch nodes for reuse by other blocks.
190 void garbage_collect_pinch_nodes();
191 // Clean up a pinch node for reuse (helper for above).
192 void cleanup_pinch( Node *pinch );
193
194 // Information for statistics gathering
195 #ifndef PRODUCT
196 private:
197 // Gather information on size of nops relative to total
198 uint _branches, _unconditional_delays;
199
200 static uint _total_nop_size, _total_method_size;
201 static uint _total_branches, _total_unconditional_delays;
202 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
203
204 public:
205 static void print_statistics();
206
207 static void increment_instructions_per_bundle(uint i) {
208 _total_instructions_per_bundle[i]++;
209 }
210
211 static void increment_nop_size(uint s) {
212 _total_nop_size += s;
213 }
214
215 static void increment_method_size(uint s) {
216 _total_method_size += s;
217 }
218 #endif
219
220 };
221
222 volatile int C2SafepointPollStubTable::_stub_size = 0;
223
224 Label& C2SafepointPollStubTable::add_safepoint(uintptr_t safepoint_offset) {
225 C2SafepointPollStub* entry = new (Compile::current()->comp_arena()) C2SafepointPollStub(safepoint_offset);
226 _safepoints.append(entry);
227 return entry->_stub_label;
228 }
229
230 void C2SafepointPollStubTable::emit(CodeBuffer& cb) {
231 MacroAssembler masm(&cb);
232 for (int i = _safepoints.length() - 1; i >= 0; i--) {
233 // Make sure there is enough space in the code buffer
234 if (cb.insts()->maybe_expand_to_ensure_remaining(PhaseOutput::MAX_inst_size) && cb.blob() == NULL) {
235 ciEnv::current()->record_failure("CodeCache is full");
236 return;
237 }
238
239 C2SafepointPollStub* entry = _safepoints.at(i);
240 emit_stub(masm, entry);
241 }
242 }
243
244 int C2SafepointPollStubTable::stub_size_lazy() const {
245 int size = Atomic::load(&_stub_size);
246
247 if (size != 0) {
248 return size;
249 }
250
251 Compile* const C = Compile::current();
252 BufferBlob* const blob = C->output()->scratch_buffer_blob();
253 CodeBuffer cb(blob->content_begin(), C->output()->scratch_buffer_code_size());
254 MacroAssembler masm(&cb);
255 C2SafepointPollStub* entry = _safepoints.at(0);
256 emit_stub(masm, entry);
257 size += cb.insts_size();
258
259 Atomic::store(&_stub_size, size);
260
261 return size;
262 }
263
264 int C2SafepointPollStubTable::estimate_stub_size() const {
265 if (_safepoints.length() == 0) {
266 return 0;
267 }
268
269 int result = stub_size_lazy() * _safepoints.length();
270
271 #ifdef ASSERT
272 Compile* const C = Compile::current();
273 BufferBlob* const blob = C->output()->scratch_buffer_blob();
274 int size = 0;
275
276 for (int i = _safepoints.length() - 1; i >= 0; i--) {
277 CodeBuffer cb(blob->content_begin(), C->output()->scratch_buffer_code_size());
278 MacroAssembler masm(&cb);
279 C2SafepointPollStub* entry = _safepoints.at(i);
280 emit_stub(masm, entry);
281 size += cb.insts_size();
282 }
283 assert(size == result, "stubs should not have variable size");
284 #endif
285
286 return result;
287 }
288
289 // Nmethod entry barrier stubs
290 C2EntryBarrierStub* C2EntryBarrierStubTable::add_entry_barrier() {
291 C2EntryBarrierStub* stub = new (Compile::current()->comp_arena()) C2EntryBarrierStub();
292 _stubs.append(stub);
293 return stub;
294 }
295
296 void C2EntryBarrierStubTable::emit(CodeBuffer& cb) {
297 if (_stubs.is_empty()) {
298 // No stub - nothing to do
299 return;
300 }
301
302 C2_MacroAssembler masm(&cb);
303 for (C2EntryBarrierStub* stub : _stubs) {
304 // Make sure there is enough space in the code buffer
305 if (cb.insts()->maybe_expand_to_ensure_remaining(PhaseOutput::MAX_inst_size) && cb.blob() == NULL) {
306 ciEnv::current()->record_failure("CodeCache is full");
307 return;
308 }
309
310 intptr_t before = masm.offset();
311 masm.emit_entry_barrier_stub(stub);
312 intptr_t after = masm.offset();
313 int actual_size = (int)(after - before);
314 int expected_size = masm.entry_barrier_stub_size();
315 assert(actual_size == expected_size, "Estimated size is wrong, expected %d, was %d", expected_size, actual_size);
316 }
317 }
318
319 int C2EntryBarrierStubTable::estimate_stub_size() const {
320 if (BarrierSet::barrier_set()->barrier_set_nmethod() == NULL) {
321 // No nmethod entry barrier?
322 return 0;
323 }
324
325 return C2_MacroAssembler::entry_barrier_stub_size();
326 }
327
328 PhaseOutput::PhaseOutput()
329 : Phase(Phase::Output),
330 _code_buffer("Compile::Fill_buffer"),
331 _first_block_size(0),
332 _handler_table(),
333 _inc_table(),
334 _safepoint_poll_table(),
335 _entry_barrier_table(),
336 _oop_map_set(NULL),
337 _scratch_buffer_blob(NULL),
338 _scratch_locs_memory(NULL),
339 _scratch_const_size(-1),
340 _in_scratch_emit_size(false),
341 _frame_slots(0),
342 _code_offsets(),
343 _node_bundling_limit(0),
344 _node_bundling_base(NULL),
345 _orig_pc_slot(0),
346 _orig_pc_slot_offset_in_bytes(0),
347 _buf_sizes(),
348 _block(NULL),
349 _index(0) {
350 C->set_output(this);
351 if (C->stub_name() == NULL) {
352 int fixed_slots = C->fixed_slots();
353 if (C->needs_stack_repair()) {
354 fixed_slots -= 2;
355 }
356 // TODO 8284443 Only reserve extra slot if needed
357 if (InlineTypeReturnedAsFields) {
358 fixed_slots -= 2;
359 }
360 _orig_pc_slot = fixed_slots - (sizeof(address) / VMRegImpl::stack_slot_size);
361 }
362 }
363
364 PhaseOutput::~PhaseOutput() {
365 C->set_output(NULL);
366 if (_scratch_buffer_blob != NULL) {
367 BufferBlob::free(_scratch_buffer_blob);
368 }
369 }
370
371 void PhaseOutput::perform_mach_node_analysis() {
372 // Late barrier analysis must be done after schedule and bundle
373 // Otherwise liveness based spilling will fail
374 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
375 bs->late_barrier_analysis();
376
377 pd_perform_mach_node_analysis();
378
379 C->print_method(CompilerPhaseType::PHASE_MACHANALYSIS, 4);
380 }
381
382 // Convert Nodes to instruction bits and pass off to the VM
383 void PhaseOutput::Output() {
384 // RootNode goes
385 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
386
387 // The number of new nodes (mostly MachNop) is proportional to
388 // the number of java calls and inner loops which are aligned.
389 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
390 C->inner_loops()*(OptoLoopAlignment-1)),
391 "out of nodes before code generation" ) ) {
392 return;
393 }
394 // Make sure I can find the Start Node
395 Block *entry = C->cfg()->get_block(1);
396 Block *broot = C->cfg()->get_root_block();
397
398 const StartNode *start = entry->head()->as_Start();
399
400 // Replace StartNode with prolog
401 Label verified_entry;
402 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
403 entry->map_node(prolog, 0);
404 C->cfg()->map_node_to_block(prolog, entry);
405 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
406
407 // Virtual methods need an unverified entry point
408 if (C->is_osr_compilation()) {
409 if (PoisonOSREntry) {
410 // TODO: Should use a ShouldNotReachHereNode...
411 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
412 }
413 } else {
414 if (C->method()) {
415 if (C->method()->has_scalarized_args()) {
416 // Add entry point to unpack all inline type arguments
417 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false));
418 if (!C->method()->is_static()) {
419 // Add verified/unverified entry points to only unpack inline type receiver at interface calls
420 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ false));
421 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true));
422 C->cfg()->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true));
423 }
424 } else if (!C->method()->is_static()) {
425 // Insert unvalidated entry point
426 C->cfg()->insert(broot, 0, new MachUEPNode());
427 }
428 }
429 }
430
431 // Break before main entry point
432 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
433 (OptoBreakpoint && C->is_method_compilation()) ||
434 (OptoBreakpointOSR && C->is_osr_compilation()) ||
435 (OptoBreakpointC2R && !C->method()) ) {
436 // checking for C->method() means that OptoBreakpoint does not apply to
437 // runtime stubs or frame converters
438 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
439 }
440
441 // Insert epilogs before every return
442 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
443 Block* block = C->cfg()->get_block(i);
444 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
445 Node* m = block->end();
446 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
447 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
448 block->add_inst(epilog);
449 C->cfg()->map_node_to_block(epilog, block);
450 }
451 }
452 }
453
454 // Keeper of sizing aspects
455 _buf_sizes = BufferSizingData();
456
457 // Initialize code buffer
458 estimate_buffer_size(_buf_sizes._const);
459 if (C->failing()) return;
460
461 // Pre-compute the length of blocks and replace
462 // long branches with short if machine supports it.
463 // Must be done before ScheduleAndBundle due to SPARC delay slots
464 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
465 blk_starts[0] = 0;
466 shorten_branches(blk_starts);
467
468 if (!C->is_osr_compilation() && C->has_scalarized_args()) {
469 // Compute the offsets of the entry points required by the inline type calling convention
470 if (!C->method()->is_static()) {
471 // We have entries at the beginning of the method, implemented by the first 4 nodes.
472 // Entry (unverified) @ offset 0
473 // Verified_Inline_Entry_RO
474 // Inline_Entry (unverified)
475 // Verified_Inline_Entry
476 uint offset = 0;
477 _code_offsets.set_value(CodeOffsets::Entry, offset);
478
479 offset += ((MachVEPNode*)broot->get_node(0))->size(C->regalloc());
480 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, offset);
481
482 offset += ((MachVEPNode*)broot->get_node(1))->size(C->regalloc());
483 _code_offsets.set_value(CodeOffsets::Inline_Entry, offset);
484
485 offset += ((MachVEPNode*)broot->get_node(2))->size(C->regalloc());
486 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, offset);
487 } else {
488 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later
489 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, 0);
490 }
491 }
492
493 ScheduleAndBundle();
494 if (C->failing()) {
495 return;
496 }
497
498 perform_mach_node_analysis();
499
500 // Complete sizing of codebuffer
501 CodeBuffer* cb = init_buffer();
502 if (cb == NULL || C->failing()) {
503 return;
504 }
505
506 BuildOopMaps();
507
508 if (C->failing()) {
509 return;
510 }
511
512 fill_buffer(cb, blk_starts);
513 }
514
515 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
516 // Determine if we need to generate a stack overflow check.
517 // Do it if the method is not a stub function and
518 // has java calls or has frame size > vm_page_size/8.
519 // The debug VM checks that deoptimization doesn't trigger an
520 // unexpected stack overflow (compiled method stack banging should
521 // guarantee it doesn't happen) so we always need the stack bang in
522 // a debug VM.
523 return (C->stub_function() == NULL &&
524 (C->has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
525 DEBUG_ONLY(|| true)));
526 }
527
528 bool PhaseOutput::need_register_stack_bang() const {
529 // Determine if we need to generate a register stack overflow check.
530 // This is only used on architectures which have split register
531 // and memory stacks (ie. IA64).
532 // Bang if the method is not a stub function and has java calls
533 return (C->stub_function() == NULL && C->has_java_calls());
534 }
535
536
537 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
538 // of a loop. When aligning a loop we need to provide enough instructions
539 // in cpu's fetch buffer to feed decoders. The loop alignment could be
540 // avoided if we have enough instructions in fetch buffer at the head of a loop.
541 // By default, the size is set to 999999 by Block's constructor so that
542 // a loop will be aligned if the size is not reset here.
543 //
544 // Note: Mach instructions could contain several HW instructions
545 // so the size is estimated only.
546 //
547 void PhaseOutput::compute_loop_first_inst_sizes() {
548 // The next condition is used to gate the loop alignment optimization.
549 // Don't aligned a loop if there are enough instructions at the head of a loop
550 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
551 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
552 // equal to 11 bytes which is the largest address NOP instruction.
553 if (MaxLoopPad < OptoLoopAlignment - 1) {
554 uint last_block = C->cfg()->number_of_blocks() - 1;
555 for (uint i = 1; i <= last_block; i++) {
556 Block* block = C->cfg()->get_block(i);
557 // Check the first loop's block which requires an alignment.
558 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
559 uint sum_size = 0;
560 uint inst_cnt = NumberOfLoopInstrToAlign;
561 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
562
563 // Check subsequent fallthrough blocks if the loop's first
564 // block(s) does not have enough instructions.
565 Block *nb = block;
566 while(inst_cnt > 0 &&
567 i < last_block &&
568 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
569 !nb->has_successor(block)) {
570 i++;
571 nb = C->cfg()->get_block(i);
572 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
573 } // while( inst_cnt > 0 && i < last_block )
574
575 block->set_first_inst_size(sum_size);
576 } // f( b->head()->is_Loop() )
577 } // for( i <= last_block )
578 } // if( MaxLoopPad < OptoLoopAlignment-1 )
579 }
580
581 // The architecture description provides short branch variants for some long
582 // branch instructions. Replace eligible long branches with short branches.
583 void PhaseOutput::shorten_branches(uint* blk_starts) {
584
585 Compile::TracePhase tp("shorten branches", &timers[_t_shortenBranches]);
586
587 // Compute size of each block, method size, and relocation information size
588 uint nblocks = C->cfg()->number_of_blocks();
589
590 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
591 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
592 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
593
594 // Collect worst case block paddings
595 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
596 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
597
598 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
599 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
600
601 bool has_short_branch_candidate = false;
602
603 // Initialize the sizes to 0
604 int code_size = 0; // Size in bytes of generated code
605 int stub_size = 0; // Size in bytes of all stub entries
606 // Size in bytes of all relocation entries, including those in local stubs.
607 // Start with 2-bytes of reloc info for the unvalidated entry point
608 int reloc_size = 1; // Number of relocation entries
609
610 // Make three passes. The first computes pessimistic blk_starts,
611 // relative jmp_offset and reloc_size information. The second performs
612 // short branch substitution using the pessimistic sizing. The
613 // third inserts nops where needed.
614
615 // Step one, perform a pessimistic sizing pass.
616 uint last_call_adr = max_juint;
617 uint last_avoid_back_to_back_adr = max_juint;
618 uint nop_size = (new MachNopNode())->size(C->regalloc());
619 for (uint i = 0; i < nblocks; i++) { // For all blocks
620 Block* block = C->cfg()->get_block(i);
621 _block = block;
622
623 // During short branch replacement, we store the relative (to blk_starts)
624 // offset of jump in jmp_offset, rather than the absolute offset of jump.
625 // This is so that we do not need to recompute sizes of all nodes when
626 // we compute correct blk_starts in our next sizing pass.
627 jmp_offset[i] = 0;
628 jmp_size[i] = 0;
629 jmp_nidx[i] = -1;
630 DEBUG_ONLY( jmp_target[i] = 0; )
631 DEBUG_ONLY( jmp_rule[i] = 0; )
632
633 // Sum all instruction sizes to compute block size
634 uint last_inst = block->number_of_nodes();
635 uint blk_size = 0;
636 for (uint j = 0; j < last_inst; j++) {
637 _index = j;
638 Node* nj = block->get_node(_index);
639 // Handle machine instruction nodes
640 if (nj->is_Mach()) {
641 MachNode* mach = nj->as_Mach();
642 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
643 reloc_size += mach->reloc();
644 if (mach->is_MachCall()) {
645 // add size information for trampoline stub
646 // class CallStubImpl is platform-specific and defined in the *.ad files.
647 stub_size += CallStubImpl::size_call_trampoline();
648 reloc_size += CallStubImpl::reloc_call_trampoline();
649
650 MachCallNode *mcall = mach->as_MachCall();
651 // This destination address is NOT PC-relative
652
653 if (mcall->entry_point() != NULL) {
654 mcall->method_set((intptr_t)mcall->entry_point());
655 }
656
657 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
658 stub_size += CompiledStaticCall::to_interp_stub_size();
659 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
660 }
661 } else if (mach->is_MachSafePoint()) {
662 // If call/safepoint are adjacent, account for possible
663 // nop to disambiguate the two safepoints.
664 // ScheduleAndBundle() can rearrange nodes in a block,
665 // check for all offsets inside this block.
666 if (last_call_adr >= blk_starts[i]) {
667 blk_size += nop_size;
668 }
669 }
670 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
671 // Nop is inserted between "avoid back to back" instructions.
672 // ScheduleAndBundle() can rearrange nodes in a block,
673 // check for all offsets inside this block.
674 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
675 blk_size += nop_size;
676 }
677 }
678 if (mach->may_be_short_branch()) {
679 if (!nj->is_MachBranch()) {
680 #ifndef PRODUCT
681 nj->dump(3);
682 #endif
683 Unimplemented();
684 }
685 assert(jmp_nidx[i] == -1, "block should have only one branch");
686 jmp_offset[i] = blk_size;
687 jmp_size[i] = nj->size(C->regalloc());
688 jmp_nidx[i] = j;
689 has_short_branch_candidate = true;
690 }
691 }
692 blk_size += nj->size(C->regalloc());
693 // Remember end of call offset
694 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
695 last_call_adr = blk_starts[i]+blk_size;
696 }
697 // Remember end of avoid_back_to_back offset
698 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
699 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
700 }
701 }
702
703 // When the next block starts a loop, we may insert pad NOP
704 // instructions. Since we cannot know our future alignment,
705 // assume the worst.
706 if (i < nblocks - 1) {
707 Block* nb = C->cfg()->get_block(i + 1);
708 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
709 if (max_loop_pad > 0) {
710 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
711 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
712 // If either is the last instruction in this block, bump by
713 // max_loop_pad in lock-step with blk_size, so sizing
714 // calculations in subsequent blocks still can conservatively
715 // detect that it may the last instruction in this block.
716 if (last_call_adr == blk_starts[i]+blk_size) {
717 last_call_adr += max_loop_pad;
718 }
719 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
720 last_avoid_back_to_back_adr += max_loop_pad;
721 }
722 blk_size += max_loop_pad;
723 block_worst_case_pad[i + 1] = max_loop_pad;
724 }
725 }
726
727 // Save block size; update total method size
728 blk_starts[i+1] = blk_starts[i]+blk_size;
729 }
730
731 // Step two, replace eligible long jumps.
732 bool progress = true;
733 uint last_may_be_short_branch_adr = max_juint;
734 while (has_short_branch_candidate && progress) {
735 progress = false;
736 has_short_branch_candidate = false;
737 int adjust_block_start = 0;
738 for (uint i = 0; i < nblocks; i++) {
739 Block* block = C->cfg()->get_block(i);
740 int idx = jmp_nidx[i];
741 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
742 if (mach != NULL && mach->may_be_short_branch()) {
743 #ifdef ASSERT
744 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
745 int j;
746 // Find the branch; ignore trailing NOPs.
747 for (j = block->number_of_nodes()-1; j>=0; j--) {
748 Node* n = block->get_node(j);
749 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
750 break;
751 }
752 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
753 #endif
754 int br_size = jmp_size[i];
755 int br_offs = blk_starts[i] + jmp_offset[i];
756
757 // This requires the TRUE branch target be in succs[0]
758 uint bnum = block->non_connector_successor(0)->_pre_order;
759 int offset = blk_starts[bnum] - br_offs;
760 if (bnum > i) { // adjust following block's offset
761 offset -= adjust_block_start;
762 }
763
764 // This block can be a loop header, account for the padding
765 // in the previous block.
766 int block_padding = block_worst_case_pad[i];
767 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
768 // In the following code a nop could be inserted before
769 // the branch which will increase the backward distance.
770 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
771 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
772
773 if (needs_padding && offset <= 0)
774 offset -= nop_size;
775
776 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
777 // We've got a winner. Replace this branch.
778 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
779
780 // Update the jmp_size.
781 int new_size = replacement->size(C->regalloc());
782 int diff = br_size - new_size;
783 assert(diff >= (int)nop_size, "short_branch size should be smaller");
784 // Conservatively take into account padding between
785 // avoid_back_to_back branches. Previous branch could be
786 // converted into avoid_back_to_back branch during next
787 // rounds.
788 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
789 jmp_offset[i] += nop_size;
790 diff -= nop_size;
791 }
792 adjust_block_start += diff;
793 block->map_node(replacement, idx);
794 mach->subsume_by(replacement, C);
795 mach = replacement;
796 progress = true;
797
798 jmp_size[i] = new_size;
799 DEBUG_ONLY( jmp_target[i] = bnum; );
800 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
801 } else {
802 // The jump distance is not short, try again during next iteration.
803 has_short_branch_candidate = true;
804 }
805 } // (mach->may_be_short_branch())
806 if (mach != NULL && (mach->may_be_short_branch() ||
807 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
808 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
809 }
810 blk_starts[i+1] -= adjust_block_start;
811 }
812 }
813
814 #ifdef ASSERT
815 for (uint i = 0; i < nblocks; i++) { // For all blocks
816 if (jmp_target[i] != 0) {
817 int br_size = jmp_size[i];
818 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
819 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
820 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]);
821 }
822 assert(C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
823 }
824 }
825 #endif
826
827 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
828 // after ScheduleAndBundle().
829
830 // ------------------
831 // Compute size for code buffer
832 code_size = blk_starts[nblocks];
833
834 // Relocation records
835 reloc_size += 1; // Relo entry for exception handler
836
837 // Adjust reloc_size to number of record of relocation info
838 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
839 // a relocation index.
840 // The CodeBuffer will expand the locs array if this estimate is too low.
841 reloc_size *= 10 / sizeof(relocInfo);
842
843 _buf_sizes._reloc = reloc_size;
844 _buf_sizes._code = code_size;
845 _buf_sizes._stub = stub_size;
846 }
847
848 //------------------------------FillLocArray-----------------------------------
849 // Create a bit of debug info and append it to the array. The mapping is from
850 // Java local or expression stack to constant, register or stack-slot. For
851 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
852 // entry has been taken care of and caller should skip it).
853 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
854 // This should never have accepted Bad before
855 assert(OptoReg::is_valid(regnum), "location must be valid");
856 return (OptoReg::is_reg(regnum))
857 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
858 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
859 }
860
861
862 ObjectValue*
863 PhaseOutput::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
864 for (int i = 0; i < objs->length(); i++) {
865 assert(objs->at(i)->is_object(), "corrupt object cache");
866 ObjectValue* sv = (ObjectValue*) objs->at(i);
867 if (sv->id() == id) {
868 return sv;
869 }
870 }
871 // Otherwise..
872 return NULL;
873 }
874
875 void PhaseOutput::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
876 ObjectValue* sv ) {
877 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
878 objs->append(sv);
879 }
880
881
882 void PhaseOutput::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
883 GrowableArray<ScopeValue*> *array,
884 GrowableArray<ScopeValue*> *objs ) {
885 assert( local, "use _top instead of null" );
886 if (array->length() != idx) {
887 assert(array->length() == idx + 1, "Unexpected array count");
888 // Old functionality:
889 // return
890 // New functionality:
891 // Assert if the local is not top. In product mode let the new node
892 // override the old entry.
893 assert(local == C->top(), "LocArray collision");
894 if (local == C->top()) {
895 return;
896 }
897 array->pop();
898 }
899 const Type *t = local->bottom_type();
900
901 // Is it a safepoint scalar object node?
902 if (local->is_SafePointScalarObject()) {
903 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
904
905 ObjectValue* sv = sv_for_node_id(objs, spobj->_idx);
906 if (sv == NULL) {
907 ciKlass* cik = t->is_oopptr()->exact_klass();
908 assert(cik->is_instance_klass() ||
909 cik->is_array_klass(), "Not supported allocation.");
910 uint first_ind = spobj->first_index(sfpt->jvms());
911 // Nullable, scalarized inline types have an is_init input
912 // that needs to be checked before using the field values.
913 ScopeValue* is_init = NULL;
914 if (cik->is_inlinetype()) {
915 Node* init_node = sfpt->in(first_ind++);
916 assert(init_node != NULL, "is_init node not found");
917 if (!init_node->is_top()) {
918 const TypeInt* init_type = init_node->bottom_type()->is_int();
919 if (init_node->is_Con()) {
920 is_init = new ConstantIntValue(init_type->get_con());
921 } else {
922 OptoReg::Name init_reg = C->regalloc()->get_reg_first(init_node);
923 is_init = new_loc_value(C->regalloc(), init_reg, Location::normal);
924 }
925 }
926 }
927 sv = new ObjectValue(spobj->_idx,
928 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()), is_init);
929 set_sv_for_object_node(objs, sv);
930
931 for (uint i = 0; i < spobj->n_fields(); i++) {
932 Node* fld_node = sfpt->in(first_ind+i);
933 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
934 }
935 }
936 array->append(sv);
937 return;
938 }
939
940 // Grab the register number for the local
941 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
942 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
943 // Record the double as two float registers.
944 // The register mask for such a value always specifies two adjacent
945 // float registers, with the lower register number even.
946 // Normally, the allocation of high and low words to these registers
947 // is irrelevant, because nearly all operations on register pairs
948 // (e.g., StoreD) treat them as a single unit.
949 // Here, we assume in addition that the words in these two registers
950 // stored "naturally" (by operations like StoreD and double stores
951 // within the interpreter) such that the lower-numbered register
952 // is written to the lower memory address. This may seem like
953 // a machine dependency, but it is not--it is a requirement on
954 // the author of the <arch>.ad file to ensure that, for every
955 // even/odd double-register pair to which a double may be allocated,
956 // the word in the even single-register is stored to the first
957 // memory word. (Note that register numbers are completely
958 // arbitrary, and are not tied to any machine-level encodings.)
959 #ifdef _LP64
960 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
961 array->append(new ConstantIntValue((jint)0));
962 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
963 } else if ( t->base() == Type::Long ) {
964 array->append(new ConstantIntValue((jint)0));
965 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
966 } else if ( t->base() == Type::RawPtr ) {
967 // jsr/ret return address which must be restored into the full
968 // width 64-bit stack slot.
969 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
970 }
971 #else //_LP64
972 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
973 // Repack the double/long as two jints.
974 // The convention the interpreter uses is that the second local
975 // holds the first raw word of the native double representation.
976 // This is actually reasonable, since locals and stack arrays
977 // grow downwards in all implementations.
978 // (If, on some machine, the interpreter's Java locals or stack
979 // were to grow upwards, the embedded doubles would be word-swapped.)
980 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
981 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
982 }
983 #endif //_LP64
984 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
985 OptoReg::is_reg(regnum) ) {
986 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
987 ? Location::float_in_dbl : Location::normal ));
988 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
989 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
990 ? Location::int_in_long : Location::normal ));
991 } else if( t->base() == Type::NarrowOop ) {
992 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
993 } else if (t->base() == Type::VectorA || t->base() == Type::VectorS ||
994 t->base() == Type::VectorD || t->base() == Type::VectorX ||
995 t->base() == Type::VectorY || t->base() == Type::VectorZ) {
996 array->append(new_loc_value( C->regalloc(), regnum, Location::vector ));
997 } else if (C->regalloc()->is_oop(local)) {
998 assert(t->base() == Type::OopPtr || t->base() == Type::InstPtr ||
999 t->base() == Type::AryPtr,
1000 "Unexpected type: %s", t->msg());
1001 array->append(new_loc_value( C->regalloc(), regnum, Location::oop ));
1002 } else {
1003 assert(t->base() == Type::Int || t->base() == Type::Half ||
1004 t->base() == Type::FloatCon || t->base() == Type::FloatBot,
1005 "Unexpected type: %s", t->msg());
1006 array->append(new_loc_value( C->regalloc(), regnum, Location::normal ));
1007 }
1008 return;
1009 }
1010
1011 // No register. It must be constant data.
1012 switch (t->base()) {
1013 case Type::Half: // Second half of a double
1014 ShouldNotReachHere(); // Caller should skip 2nd halves
1015 break;
1016 case Type::AnyPtr:
1017 array->append(new ConstantOopWriteValue(NULL));
1018 break;
1019 case Type::AryPtr:
1020 case Type::InstPtr: // fall through
1021 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
1022 break;
1023 case Type::NarrowOop:
1024 if (t == TypeNarrowOop::NULL_PTR) {
1025 array->append(new ConstantOopWriteValue(NULL));
1026 } else {
1027 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
1028 }
1029 break;
1030 case Type::Int:
1031 array->append(new ConstantIntValue(t->is_int()->get_con()));
1032 break;
1033 case Type::RawPtr:
1034 // A return address (T_ADDRESS).
1035 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
1036 #ifdef _LP64
1037 // Must be restored to the full-width 64-bit stack slot.
1038 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
1039 #else
1040 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
1041 #endif
1042 break;
1043 case Type::FloatCon: {
1044 float f = t->is_float_constant()->getf();
1045 array->append(new ConstantIntValue(jint_cast(f)));
1046 break;
1047 }
1048 case Type::DoubleCon: {
1049 jdouble d = t->is_double_constant()->getd();
1050 #ifdef _LP64
1051 array->append(new ConstantIntValue((jint)0));
1052 array->append(new ConstantDoubleValue(d));
1053 #else
1054 // Repack the double as two jints.
1055 // The convention the interpreter uses is that the second local
1056 // holds the first raw word of the native double representation.
1057 // This is actually reasonable, since locals and stack arrays
1058 // grow downwards in all implementations.
1059 // (If, on some machine, the interpreter's Java locals or stack
1060 // were to grow upwards, the embedded doubles would be word-swapped.)
1061 jlong_accessor acc;
1062 acc.long_value = jlong_cast(d);
1063 array->append(new ConstantIntValue(acc.words[1]));
1064 array->append(new ConstantIntValue(acc.words[0]));
1065 #endif
1066 break;
1067 }
1068 case Type::Long: {
1069 jlong d = t->is_long()->get_con();
1070 #ifdef _LP64
1071 array->append(new ConstantIntValue((jint)0));
1072 array->append(new ConstantLongValue(d));
1073 #else
1074 // Repack the long as two jints.
1075 // The convention the interpreter uses is that the second local
1076 // holds the first raw word of the native double representation.
1077 // This is actually reasonable, since locals and stack arrays
1078 // grow downwards in all implementations.
1079 // (If, on some machine, the interpreter's Java locals or stack
1080 // were to grow upwards, the embedded doubles would be word-swapped.)
1081 jlong_accessor acc;
1082 acc.long_value = d;
1083 array->append(new ConstantIntValue(acc.words[1]));
1084 array->append(new ConstantIntValue(acc.words[0]));
1085 #endif
1086 break;
1087 }
1088 case Type::Top: // Add an illegal value here
1089 array->append(new LocationValue(Location()));
1090 break;
1091 default:
1092 ShouldNotReachHere();
1093 break;
1094 }
1095 }
1096
1097 // Determine if this node starts a bundle
1098 bool PhaseOutput::starts_bundle(const Node *n) const {
1099 return (_node_bundling_limit > n->_idx &&
1100 _node_bundling_base[n->_idx].starts_bundle());
1101 }
1102
1103 //--------------------------Process_OopMap_Node--------------------------------
1104 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
1105 // Handle special safepoint nodes for synchronization
1106 MachSafePointNode *sfn = mach->as_MachSafePoint();
1107 MachCallNode *mcall;
1108
1109 int safepoint_pc_offset = current_offset;
1110 bool is_method_handle_invoke = false;
1111 bool return_oop = false;
1112 bool return_scalarized = false;
1113 bool has_ea_local_in_scope = sfn->_has_ea_local_in_scope;
1114 bool arg_escape = false;
1115
1116 // Add the safepoint in the DebugInfoRecorder
1117 if( !mach->is_MachCall() ) {
1118 mcall = NULL;
1119 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
1120 } else {
1121 mcall = mach->as_MachCall();
1122
1123 // Is the call a MethodHandle call?
1124 if (mcall->is_MachCallJava()) {
1125 if (mcall->as_MachCallJava()->_method_handle_invoke) {
1126 assert(C->has_method_handle_invokes(), "must have been set during call generation");
1127 is_method_handle_invoke = true;
1128 }
1129 arg_escape = mcall->as_MachCallJava()->_arg_escape;
1130 }
1131
1132 // Check if a call returns an object.
1133 if (mcall->returns_pointer() || mcall->returns_scalarized()) {
1134 return_oop = true;
1135 }
1136 if (mcall->returns_scalarized()) {
1137 return_scalarized = true;
1138 }
1139 safepoint_pc_offset += mcall->ret_addr_offset();
1140 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
1141 }
1142
1143 // Loop over the JVMState list to add scope information
1144 // Do not skip safepoints with a NULL method, they need monitor info
1145 JVMState* youngest_jvms = sfn->jvms();
1146 int max_depth = youngest_jvms->depth();
1147
1148 // Allocate the object pool for scalar-replaced objects -- the map from
1149 // small-integer keys (which can be recorded in the local and ostack
1150 // arrays) to descriptions of the object state.
1151 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
1152
1153 // Visit scopes from oldest to youngest.
1154 for (int depth = 1; depth <= max_depth; depth++) {
1155 JVMState* jvms = youngest_jvms->of_depth(depth);
1156 int idx;
1157 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1158 // Safepoints that do not have method() set only provide oop-map and monitor info
1159 // to support GC; these do not support deoptimization.
1160 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
1161 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
1162 int num_mon = jvms->nof_monitors();
1163 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
1164 "JVMS local count must match that of the method");
1165
1166 // Add Local and Expression Stack Information
1167
1168 // Insert locals into the locarray
1169 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
1170 for( idx = 0; idx < num_locs; idx++ ) {
1171 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
1172 }
1173
1174 // Insert expression stack entries into the exparray
1175 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1176 for( idx = 0; idx < num_exps; idx++ ) {
1177 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1178 }
1179
1180 // Add in mappings of the monitors
1181 assert( !method ||
1182 !method->is_synchronized() ||
1183 method->is_native() ||
1184 num_mon > 0 ||
1185 !GenerateSynchronizationCode,
1186 "monitors must always exist for synchronized methods");
1187
1188 // Build the growable array of ScopeValues for exp stack
1189 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1190
1191 // Loop over monitors and insert into array
1192 for (idx = 0; idx < num_mon; idx++) {
1193 // Grab the node that defines this monitor
1194 Node* box_node = sfn->monitor_box(jvms, idx);
1195 Node* obj_node = sfn->monitor_obj(jvms, idx);
1196
1197 // Create ScopeValue for object
1198 ScopeValue *scval = NULL;
1199
1200 if (obj_node->is_SafePointScalarObject()) {
1201 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1202 scval = PhaseOutput::sv_for_node_id(objs, spobj->_idx);
1203 if (scval == NULL) {
1204 const Type *t = spobj->bottom_type();
1205 ciKlass* cik = t->is_oopptr()->exact_klass();
1206 assert(cik->is_instance_klass() ||
1207 cik->is_array_klass(), "Not supported allocation.");
1208 ObjectValue* sv = new ObjectValue(spobj->_idx,
1209 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1210 PhaseOutput::set_sv_for_object_node(objs, sv);
1211
1212 uint first_ind = spobj->first_index(youngest_jvms);
1213 for (uint i = 0; i < spobj->n_fields(); i++) {
1214 Node* fld_node = sfn->in(first_ind+i);
1215 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1216 }
1217 scval = sv;
1218 }
1219 } else if (!obj_node->is_Con()) {
1220 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1221 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1222 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1223 } else {
1224 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1225 }
1226 } else {
1227 const TypePtr *tp = obj_node->get_ptr_type();
1228 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1229 }
1230
1231 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1232 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1233 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1234 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1235 }
1236
1237 // We dump the object pool first, since deoptimization reads it in first.
1238 C->debug_info()->dump_object_pool(objs);
1239
1240 // Build first class objects to pass to scope
1241 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1242 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1243 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1244
1245 // Make method available for all Safepoints
1246 ciMethod* scope_method = method ? method : C->method();
1247 // Describe the scope here
1248 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1249 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1250 // Now we can describe the scope.
1251 methodHandle null_mh;
1252 bool rethrow_exception = false;
1253 C->debug_info()->describe_scope(
1254 safepoint_pc_offset,
1255 null_mh,
1256 scope_method,
1257 jvms->bci(),
1258 jvms->should_reexecute(),
1259 rethrow_exception,
1260 is_method_handle_invoke,
1261 return_oop,
1262 return_scalarized,
1263 has_ea_local_in_scope,
1264 arg_escape,
1265 locvals,
1266 expvals,
1267 monvals
1268 );
1269 } // End jvms loop
1270
1271 // Mark the end of the scope set.
1272 C->debug_info()->end_safepoint(safepoint_pc_offset);
1273 }
1274
1275
1276
1277 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1278 class NonSafepointEmitter {
1279 Compile* C;
1280 JVMState* _pending_jvms;
1281 int _pending_offset;
1282
1283 void emit_non_safepoint();
1284
1285 public:
1286 NonSafepointEmitter(Compile* compile) {
1287 this->C = compile;
1288 _pending_jvms = NULL;
1289 _pending_offset = 0;
1290 }
1291
1292 void observe_instruction(Node* n, int pc_offset) {
1293 if (!C->debug_info()->recording_non_safepoints()) return;
1294
1295 Node_Notes* nn = C->node_notes_at(n->_idx);
1296 if (nn == NULL || nn->jvms() == NULL) return;
1297 if (_pending_jvms != NULL &&
1298 _pending_jvms->same_calls_as(nn->jvms())) {
1299 // Repeated JVMS? Stretch it up here.
1300 _pending_offset = pc_offset;
1301 } else {
1302 if (_pending_jvms != NULL &&
1303 _pending_offset < pc_offset) {
1304 emit_non_safepoint();
1305 }
1306 _pending_jvms = NULL;
1307 if (pc_offset > C->debug_info()->last_pc_offset()) {
1308 // This is the only way _pending_jvms can become non-NULL:
1309 _pending_jvms = nn->jvms();
1310 _pending_offset = pc_offset;
1311 }
1312 }
1313 }
1314
1315 // Stay out of the way of real safepoints:
1316 void observe_safepoint(JVMState* jvms, int pc_offset) {
1317 if (_pending_jvms != NULL &&
1318 !_pending_jvms->same_calls_as(jvms) &&
1319 _pending_offset < pc_offset) {
1320 emit_non_safepoint();
1321 }
1322 _pending_jvms = NULL;
1323 }
1324
1325 void flush_at_end() {
1326 if (_pending_jvms != NULL) {
1327 emit_non_safepoint();
1328 }
1329 _pending_jvms = NULL;
1330 }
1331 };
1332
1333 void NonSafepointEmitter::emit_non_safepoint() {
1334 JVMState* youngest_jvms = _pending_jvms;
1335 int pc_offset = _pending_offset;
1336
1337 // Clear it now:
1338 _pending_jvms = NULL;
1339
1340 DebugInformationRecorder* debug_info = C->debug_info();
1341 assert(debug_info->recording_non_safepoints(), "sanity");
1342
1343 debug_info->add_non_safepoint(pc_offset);
1344 int max_depth = youngest_jvms->depth();
1345
1346 // Visit scopes from oldest to youngest.
1347 for (int depth = 1; depth <= max_depth; depth++) {
1348 JVMState* jvms = youngest_jvms->of_depth(depth);
1349 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1350 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1351 methodHandle null_mh;
1352 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1353 }
1354
1355 // Mark the end of the scope set.
1356 debug_info->end_non_safepoint(pc_offset);
1357 }
1358
1359 //------------------------------init_buffer------------------------------------
1360 void PhaseOutput::estimate_buffer_size(int& const_req) {
1361
1362 // Set the initially allocated size
1363 const_req = initial_const_capacity;
1364
1365 // The extra spacing after the code is necessary on some platforms.
1366 // Sometimes we need to patch in a jump after the last instruction,
1367 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1368
1369 // Compute the byte offset where we can store the deopt pc.
1370 if (C->fixed_slots() != 0) {
1371 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1372 }
1373
1374 // Compute prolog code size
1375 _method_size = 0;
1376 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1377 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1378
1379 if (C->has_mach_constant_base_node()) {
1380 uint add_size = 0;
1381 // Fill the constant table.
1382 // Note: This must happen before shorten_branches.
1383 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1384 Block* b = C->cfg()->get_block(i);
1385
1386 for (uint j = 0; j < b->number_of_nodes(); j++) {
1387 Node* n = b->get_node(j);
1388
1389 // If the node is a MachConstantNode evaluate the constant
1390 // value section.
1391 if (n->is_MachConstant()) {
1392 MachConstantNode* machcon = n->as_MachConstant();
1393 machcon->eval_constant(C);
1394 } else if (n->is_Mach()) {
1395 // On Power there are more nodes that issue constants.
1396 add_size += (n->as_Mach()->ins_num_consts() * 8);
1397 }
1398 }
1399 }
1400
1401 // Calculate the offsets of the constants and the size of the
1402 // constant table (including the padding to the next section).
1403 constant_table().calculate_offsets_and_size();
1404 const_req = constant_table().size() + add_size;
1405 }
1406
1407 // Initialize the space for the BufferBlob used to find and verify
1408 // instruction size in MachNode::emit_size()
1409 init_scratch_buffer_blob(const_req);
1410 }
1411
1412 CodeBuffer* PhaseOutput::init_buffer() {
1413 int stub_req = _buf_sizes._stub;
1414 int code_req = _buf_sizes._code;
1415 int const_req = _buf_sizes._const;
1416
1417 int pad_req = NativeCall::instruction_size;
1418
1419 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1420 stub_req += bs->estimate_stub_size();
1421 stub_req += safepoint_poll_table()->estimate_stub_size();
1422 stub_req += entry_barrier_table()->estimate_stub_size();
1423
1424 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1425 // class HandlerImpl is platform-specific and defined in the *.ad files.
1426 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1427 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1428 stub_req += MAX_stubs_size; // ensure per-stub margin
1429 code_req += MAX_inst_size; // ensure per-instruction margin
1430
1431 if (StressCodeBuffers)
1432 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1433
1434 int total_req =
1435 const_req +
1436 code_req +
1437 pad_req +
1438 stub_req +
1439 exception_handler_req +
1440 deopt_handler_req; // deopt handler
1441
1442 if (C->has_method_handle_invokes())
1443 total_req += deopt_handler_req; // deopt MH handler
1444
1445 CodeBuffer* cb = code_buffer();
1446 cb->initialize(total_req, _buf_sizes._reloc);
1447
1448 // Have we run out of code space?
1449 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1450 C->record_failure("CodeCache is full");
1451 return NULL;
1452 }
1453 // Configure the code buffer.
1454 cb->initialize_consts_size(const_req);
1455 cb->initialize_stubs_size(stub_req);
1456 cb->initialize_oop_recorder(C->env()->oop_recorder());
1457
1458 // fill in the nop array for bundling computations
1459 MachNode *_nop_list[Bundle::_nop_count];
1460 Bundle::initialize_nops(_nop_list);
1461
1462 return cb;
1463 }
1464
1465 //------------------------------fill_buffer------------------------------------
1466 void PhaseOutput::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1467 // blk_starts[] contains offsets calculated during short branches processing,
1468 // offsets should not be increased during following steps.
1469
1470 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1471 // of a loop. It is used to determine the padding for loop alignment.
1472 Compile::TracePhase tp("fill buffer", &timers[_t_fillBuffer]);
1473
1474 compute_loop_first_inst_sizes();
1475
1476 // Create oopmap set.
1477 _oop_map_set = new OopMapSet();
1478
1479 // !!!!! This preserves old handling of oopmaps for now
1480 C->debug_info()->set_oopmaps(_oop_map_set);
1481
1482 uint nblocks = C->cfg()->number_of_blocks();
1483 // Count and start of implicit null check instructions
1484 uint inct_cnt = 0;
1485 uint* inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1486
1487 // Count and start of calls
1488 uint* call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1489
1490 uint return_offset = 0;
1491 int nop_size = (new MachNopNode())->size(C->regalloc());
1492
1493 int previous_offset = 0;
1494 int current_offset = 0;
1495 int last_call_offset = -1;
1496 int last_avoid_back_to_back_offset = -1;
1497 #ifdef ASSERT
1498 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1499 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1500 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1501 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1502 #endif
1503
1504 // Create an array of unused labels, one for each basic block, if printing is enabled
1505 #if defined(SUPPORT_OPTO_ASSEMBLY)
1506 int* node_offsets = NULL;
1507 uint node_offset_limit = C->unique();
1508
1509 if (C->print_assembly()) {
1510 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1511 }
1512 if (node_offsets != NULL) {
1513 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1514 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1515 }
1516 #endif
1517
1518 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1519
1520 // Emit the constant table.
1521 if (C->has_mach_constant_base_node()) {
1522 if (!constant_table().emit(*cb)) {
1523 C->record_failure("consts section overflow");
1524 return;
1525 }
1526 }
1527
1528 // Create an array of labels, one for each basic block
1529 Label* blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1530 for (uint i = 0; i <= nblocks; i++) {
1531 blk_labels[i].init();
1532 }
1533
1534 // Now fill in the code buffer
1535 Node* delay_slot = NULL;
1536 for (uint i = 0; i < nblocks; i++) {
1537 Block* block = C->cfg()->get_block(i);
1538 _block = block;
1539 Node* head = block->head();
1540
1541 // If this block needs to start aligned (i.e, can be reached other
1542 // than by falling-thru from the previous block), then force the
1543 // start of a new bundle.
1544 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1545 cb->flush_bundle(true);
1546 }
1547
1548 #ifdef ASSERT
1549 if (!block->is_connector()) {
1550 stringStream st;
1551 block->dump_head(C->cfg(), &st);
1552 MacroAssembler(cb).block_comment(st.freeze());
1553 }
1554 jmp_target[i] = 0;
1555 jmp_offset[i] = 0;
1556 jmp_size[i] = 0;
1557 jmp_rule[i] = 0;
1558 #endif
1559 int blk_offset = current_offset;
1560
1561 // Define the label at the beginning of the basic block
1562 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1563
1564 uint last_inst = block->number_of_nodes();
1565
1566 // Emit block normally, except for last instruction.
1567 // Emit means "dump code bits into code buffer".
1568 for (uint j = 0; j<last_inst; j++) {
1569 _index = j;
1570
1571 // Get the node
1572 Node* n = block->get_node(j);
1573
1574 // See if delay slots are supported
1575 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1576 assert(delay_slot == NULL, "no use of delay slot node");
1577 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1578
1579 delay_slot = n;
1580 continue;
1581 }
1582
1583 // If this starts a new instruction group, then flush the current one
1584 // (but allow split bundles)
1585 if (Pipeline::requires_bundling() && starts_bundle(n))
1586 cb->flush_bundle(false);
1587
1588 // Special handling for SafePoint/Call Nodes
1589 bool is_mcall = false;
1590 if (n->is_Mach()) {
1591 MachNode *mach = n->as_Mach();
1592 is_mcall = n->is_MachCall();
1593 bool is_sfn = n->is_MachSafePoint();
1594
1595 // If this requires all previous instructions be flushed, then do so
1596 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1597 cb->flush_bundle(true);
1598 current_offset = cb->insts_size();
1599 }
1600
1601 // A padding may be needed again since a previous instruction
1602 // could be moved to delay slot.
1603
1604 // align the instruction if necessary
1605 int padding = mach->compute_padding(current_offset);
1606 // Make sure safepoint node for polling is distinct from a call's
1607 // return by adding a nop if needed.
1608 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1609 padding = nop_size;
1610 }
1611 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1612 current_offset == last_avoid_back_to_back_offset) {
1613 // Avoid back to back some instructions.
1614 padding = nop_size;
1615 }
1616
1617 if (padding > 0) {
1618 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1619 int nops_cnt = padding / nop_size;
1620 MachNode *nop = new MachNopNode(nops_cnt);
1621 block->insert_node(nop, j++);
1622 last_inst++;
1623 C->cfg()->map_node_to_block(nop, block);
1624 // Ensure enough space.
1625 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1626 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1627 C->record_failure("CodeCache is full");
1628 return;
1629 }
1630 nop->emit(*cb, C->regalloc());
1631 cb->flush_bundle(true);
1632 current_offset = cb->insts_size();
1633 }
1634
1635 bool observe_safepoint = is_sfn;
1636 // Remember the start of the last call in a basic block
1637 if (is_mcall) {
1638 MachCallNode *mcall = mach->as_MachCall();
1639
1640 if (mcall->entry_point() != NULL) {
1641 // This destination address is NOT PC-relative
1642 mcall->method_set((intptr_t)mcall->entry_point());
1643 }
1644
1645 // Save the return address
1646 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1647
1648 observe_safepoint = mcall->guaranteed_safepoint();
1649 }
1650
1651 // sfn will be valid whenever mcall is valid now because of inheritance
1652 if (observe_safepoint) {
1653 // Handle special safepoint nodes for synchronization
1654 if (!is_mcall) {
1655 MachSafePointNode *sfn = mach->as_MachSafePoint();
1656 // !!!!! Stubs only need an oopmap right now, so bail out
1657 if (sfn->jvms()->method() == NULL) {
1658 // Write the oopmap directly to the code blob??!!
1659 continue;
1660 }
1661 } // End synchronization
1662
1663 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1664 current_offset);
1665 Process_OopMap_Node(mach, current_offset);
1666 } // End if safepoint
1667
1668 // If this is a null check, then add the start of the previous instruction to the list
1669 else if( mach->is_MachNullCheck() ) {
1670 inct_starts[inct_cnt++] = previous_offset;
1671 }
1672
1673 // If this is a branch, then fill in the label with the target BB's label
1674 else if (mach->is_MachBranch()) {
1675 // This requires the TRUE branch target be in succs[0]
1676 uint block_num = block->non_connector_successor(0)->_pre_order;
1677
1678 // Try to replace long branch if delay slot is not used,
1679 // it is mostly for back branches since forward branch's
1680 // distance is not updated yet.
1681 bool delay_slot_is_used = valid_bundle_info(n) &&
1682 C->output()->node_bundling(n)->use_unconditional_delay();
1683 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1684 assert(delay_slot == NULL, "not expecting delay slot node");
1685 int br_size = n->size(C->regalloc());
1686 int offset = blk_starts[block_num] - current_offset;
1687 if (block_num >= i) {
1688 // Current and following block's offset are not
1689 // finalized yet, adjust distance by the difference
1690 // between calculated and final offsets of current block.
1691 offset -= (blk_starts[i] - blk_offset);
1692 }
1693 // In the following code a nop could be inserted before
1694 // the branch which will increase the backward distance.
1695 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1696 if (needs_padding && offset <= 0)
1697 offset -= nop_size;
1698
1699 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1700 // We've got a winner. Replace this branch.
1701 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1702
1703 // Update the jmp_size.
1704 int new_size = replacement->size(C->regalloc());
1705 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1706 // Insert padding between avoid_back_to_back branches.
1707 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1708 MachNode *nop = new MachNopNode();
1709 block->insert_node(nop, j++);
1710 C->cfg()->map_node_to_block(nop, block);
1711 last_inst++;
1712 nop->emit(*cb, C->regalloc());
1713 cb->flush_bundle(true);
1714 current_offset = cb->insts_size();
1715 }
1716 #ifdef ASSERT
1717 jmp_target[i] = block_num;
1718 jmp_offset[i] = current_offset - blk_offset;
1719 jmp_size[i] = new_size;
1720 jmp_rule[i] = mach->rule();
1721 #endif
1722 block->map_node(replacement, j);
1723 mach->subsume_by(replacement, C);
1724 n = replacement;
1725 mach = replacement;
1726 }
1727 }
1728 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1729 } else if (mach->ideal_Opcode() == Op_Jump) {
1730 for (uint h = 0; h < block->_num_succs; h++) {
1731 Block* succs_block = block->_succs[h];
1732 for (uint j = 1; j < succs_block->num_preds(); j++) {
1733 Node* jpn = succs_block->pred(j);
1734 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1735 uint block_num = succs_block->non_connector()->_pre_order;
1736 Label *blkLabel = &blk_labels[block_num];
1737 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1738 }
1739 }
1740 }
1741 }
1742 #ifdef ASSERT
1743 // Check that oop-store precedes the card-mark
1744 else if (mach->ideal_Opcode() == Op_StoreCM) {
1745 uint storeCM_idx = j;
1746 int count = 0;
1747 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1748 Node *oop_store = mach->in(prec); // Precedence edge
1749 if (oop_store == NULL) continue;
1750 count++;
1751 uint i4;
1752 for (i4 = 0; i4 < last_inst; ++i4) {
1753 if (block->get_node(i4) == oop_store) {
1754 break;
1755 }
1756 }
1757 // Note: This test can provide a false failure if other precedence
1758 // edges have been added to the storeCMNode.
1759 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1760 }
1761 assert(count > 0, "storeCM expects at least one precedence edge");
1762 }
1763 #endif
1764 else if (!n->is_Proj()) {
1765 // Remember the beginning of the previous instruction, in case
1766 // it's followed by a flag-kill and a null-check. Happens on
1767 // Intel all the time, with add-to-memory kind of opcodes.
1768 previous_offset = current_offset;
1769 }
1770
1771 // Not an else-if!
1772 // If this is a trap based cmp then add its offset to the list.
1773 if (mach->is_TrapBasedCheckNode()) {
1774 inct_starts[inct_cnt++] = current_offset;
1775 }
1776 }
1777
1778 // Verify that there is sufficient space remaining
1779 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1780 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1781 C->record_failure("CodeCache is full");
1782 return;
1783 }
1784
1785 // Save the offset for the listing
1786 #if defined(SUPPORT_OPTO_ASSEMBLY)
1787 if ((node_offsets != NULL) && (n->_idx < node_offset_limit)) {
1788 node_offsets[n->_idx] = cb->insts_size();
1789 }
1790 #endif
1791 assert(!C->failing(), "Should not reach here if failing.");
1792
1793 // "Normal" instruction case
1794 DEBUG_ONLY(uint instr_offset = cb->insts_size());
1795 n->emit(*cb, C->regalloc());
1796 current_offset = cb->insts_size();
1797
1798 // Above we only verified that there is enough space in the instruction section.
1799 // However, the instruction may emit stubs that cause code buffer expansion.
1800 // Bail out here if expansion failed due to a lack of code cache space.
1801 if (C->failing()) {
1802 return;
1803 }
1804
1805 assert(!is_mcall || (call_returns[block->_pre_order] <= (uint)current_offset),
1806 "ret_addr_offset() not within emitted code");
1807 #ifdef ASSERT
1808 uint n_size = n->size(C->regalloc());
1809 if (n_size < (current_offset-instr_offset)) {
1810 MachNode* mach = n->as_Mach();
1811 n->dump();
1812 mach->dump_format(C->regalloc(), tty);
1813 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1814 Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty);
1815 tty->print_cr(" ------------------- ");
1816 BufferBlob* blob = this->scratch_buffer_blob();
1817 address blob_begin = blob->content_begin();
1818 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1819 assert(false, "wrong size of mach node");
1820 }
1821 #endif
1822 non_safepoints.observe_instruction(n, current_offset);
1823
1824 // mcall is last "call" that can be a safepoint
1825 // record it so we can see if a poll will directly follow it
1826 // in which case we'll need a pad to make the PcDesc sites unique
1827 // see 5010568. This can be slightly inaccurate but conservative
1828 // in the case that return address is not actually at current_offset.
1829 // This is a small price to pay.
1830
1831 if (is_mcall) {
1832 last_call_offset = current_offset;
1833 }
1834
1835 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1836 // Avoid back to back some instructions.
1837 last_avoid_back_to_back_offset = current_offset;
1838 }
1839
1840 // See if this instruction has a delay slot
1841 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1842 guarantee(delay_slot != NULL, "expecting delay slot node");
1843
1844 // Back up 1 instruction
1845 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1846
1847 // Save the offset for the listing
1848 #if defined(SUPPORT_OPTO_ASSEMBLY)
1849 if ((node_offsets != NULL) && (delay_slot->_idx < node_offset_limit)) {
1850 node_offsets[delay_slot->_idx] = cb->insts_size();
1851 }
1852 #endif
1853
1854 // Support a SafePoint in the delay slot
1855 if (delay_slot->is_MachSafePoint()) {
1856 MachNode *mach = delay_slot->as_Mach();
1857 // !!!!! Stubs only need an oopmap right now, so bail out
1858 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1859 // Write the oopmap directly to the code blob??!!
1860 delay_slot = NULL;
1861 continue;
1862 }
1863
1864 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1865 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1866 adjusted_offset);
1867 // Generate an OopMap entry
1868 Process_OopMap_Node(mach, adjusted_offset);
1869 }
1870
1871 // Insert the delay slot instruction
1872 delay_slot->emit(*cb, C->regalloc());
1873
1874 // Don't reuse it
1875 delay_slot = NULL;
1876 }
1877
1878 } // End for all instructions in block
1879
1880 // If the next block is the top of a loop, pad this block out to align
1881 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1882 if (i < nblocks-1) {
1883 Block *nb = C->cfg()->get_block(i + 1);
1884 int padding = nb->alignment_padding(current_offset);
1885 if( padding > 0 ) {
1886 MachNode *nop = new MachNopNode(padding / nop_size);
1887 block->insert_node(nop, block->number_of_nodes());
1888 C->cfg()->map_node_to_block(nop, block);
1889 nop->emit(*cb, C->regalloc());
1890 current_offset = cb->insts_size();
1891 }
1892 }
1893 // Verify that the distance for generated before forward
1894 // short branches is still valid.
1895 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1896
1897 // Save new block start offset
1898 blk_starts[i] = blk_offset;
1899 } // End of for all blocks
1900 blk_starts[nblocks] = current_offset;
1901
1902 non_safepoints.flush_at_end();
1903
1904 // Offset too large?
1905 if (C->failing()) return;
1906
1907 // Define a pseudo-label at the end of the code
1908 MacroAssembler(cb).bind( blk_labels[nblocks] );
1909
1910 // Compute the size of the first block
1911 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1912
1913 #ifdef ASSERT
1914 for (uint i = 0; i < nblocks; i++) { // For all blocks
1915 if (jmp_target[i] != 0) {
1916 int br_size = jmp_size[i];
1917 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1918 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1919 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]);
1920 assert(false, "Displacement too large for short jmp");
1921 }
1922 }
1923 }
1924 #endif
1925
1926 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1927 bs->emit_stubs(*cb);
1928 if (C->failing()) return;
1929
1930 // Fill in stubs for calling the runtime from safepoint polls.
1931 safepoint_poll_table()->emit(*cb);
1932 if (C->failing()) return;
1933
1934 // Fill in stubs for calling the runtime from nmethod entries.
1935 entry_barrier_table()->emit(*cb);
1936 if (C->failing()) return;
1937
1938 #ifndef PRODUCT
1939 // Information on the size of the method, without the extraneous code
1940 Scheduling::increment_method_size(cb->insts_size());
1941 #endif
1942
1943 // ------------------
1944 // Fill in exception table entries.
1945 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1946
1947 // Only java methods have exception handlers and deopt handlers
1948 // class HandlerImpl is platform-specific and defined in the *.ad files.
1949 if (C->method()) {
1950 // Emit the exception handler code.
1951 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1952 if (C->failing()) {
1953 return; // CodeBuffer::expand failed
1954 }
1955 // Emit the deopt handler code.
1956 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1957
1958 // Emit the MethodHandle deopt handler code (if required).
1959 if (C->has_method_handle_invokes() && !C->failing()) {
1960 // We can use the same code as for the normal deopt handler, we
1961 // just need a different entry point address.
1962 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1963 }
1964 }
1965
1966 // One last check for failed CodeBuffer::expand:
1967 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1968 C->record_failure("CodeCache is full");
1969 return;
1970 }
1971
1972 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1973 if (C->print_assembly()) {
1974 tty->cr();
1975 tty->print_cr("============================= C2-compiled nmethod ==============================");
1976 }
1977 #endif
1978
1979 #if defined(SUPPORT_OPTO_ASSEMBLY)
1980 // Dump the assembly code, including basic-block numbers
1981 if (C->print_assembly()) {
1982 ttyLocker ttyl; // keep the following output all in one block
1983 if (!VMThread::should_terminate()) { // test this under the tty lock
1984 // print_metadata and dump_asm may safepoint which makes us loose the ttylock.
1985 // We call them first and write to a stringStream, then we retake the lock to
1986 // make sure the end tag is coherent, and that xmlStream->pop_tag is done thread safe.
1987 ResourceMark rm;
1988 stringStream method_metadata_str;
1989 if (C->method() != NULL) {
1990 C->method()->print_metadata(&method_metadata_str);
1991 }
1992 stringStream dump_asm_str;
1993 dump_asm_on(&dump_asm_str, node_offsets, node_offset_limit);
1994
1995 NoSafepointVerifier nsv;
1996 ttyLocker ttyl2;
1997 // This output goes directly to the tty, not the compiler log.
1998 // To enable tools to match it up with the compilation activity,
1999 // be sure to tag this tty output with the compile ID.
2000 if (xtty != NULL) {
2001 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
2002 C->is_osr_compilation() ? " compile_kind='osr'" : "");
2003 }
2004 if (C->method() != NULL) {
2005 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
2006 tty->print_raw(method_metadata_str.freeze());
2007 } else if (C->stub_name() != NULL) {
2008 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
2009 }
2010 tty->cr();
2011 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
2012 tty->print_raw(dump_asm_str.freeze());
2013 tty->print_cr("--------------------------------------------------------------------------------");
2014 if (xtty != NULL) {
2015 xtty->tail("opto_assembly");
2016 }
2017 }
2018 }
2019 #endif
2020 }
2021
2022 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
2023 _inc_table.set_size(cnt);
2024
2025 uint inct_cnt = 0;
2026 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2027 Block* block = C->cfg()->get_block(i);
2028 Node *n = NULL;
2029 int j;
2030
2031 // Find the branch; ignore trailing NOPs.
2032 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
2033 n = block->get_node(j);
2034 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
2035 break;
2036 }
2037 }
2038
2039 // If we didn't find anything, continue
2040 if (j < 0) {
2041 continue;
2042 }
2043
2044 // Compute ExceptionHandlerTable subtable entry and add it
2045 // (skip empty blocks)
2046 if (n->is_Catch()) {
2047
2048 // Get the offset of the return from the call
2049 uint call_return = call_returns[block->_pre_order];
2050 #ifdef ASSERT
2051 assert( call_return > 0, "no call seen for this basic block" );
2052 while (block->get_node(--j)->is_MachProj()) ;
2053 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
2054 #endif
2055 // last instruction is a CatchNode, find it's CatchProjNodes
2056 int nof_succs = block->_num_succs;
2057 // allocate space
2058 GrowableArray<intptr_t> handler_bcis(nof_succs);
2059 GrowableArray<intptr_t> handler_pcos(nof_succs);
2060 // iterate through all successors
2061 for (int j = 0; j < nof_succs; j++) {
2062 Block* s = block->_succs[j];
2063 bool found_p = false;
2064 for (uint k = 1; k < s->num_preds(); k++) {
2065 Node* pk = s->pred(k);
2066 if (pk->is_CatchProj() && pk->in(0) == n) {
2067 const CatchProjNode* p = pk->as_CatchProj();
2068 found_p = true;
2069 // add the corresponding handler bci & pco information
2070 if (p->_con != CatchProjNode::fall_through_index) {
2071 // p leads to an exception handler (and is not fall through)
2072 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
2073 // no duplicates, please
2074 if (!handler_bcis.contains(p->handler_bci())) {
2075 uint block_num = s->non_connector()->_pre_order;
2076 handler_bcis.append(p->handler_bci());
2077 handler_pcos.append(blk_labels[block_num].loc_pos());
2078 }
2079 }
2080 }
2081 }
2082 assert(found_p, "no matching predecessor found");
2083 // Note: Due to empty block removal, one block may have
2084 // several CatchProj inputs, from the same Catch.
2085 }
2086
2087 // Set the offset of the return from the call
2088 assert(handler_bcis.find(-1) != -1, "must have default handler");
2089 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
2090 continue;
2091 }
2092
2093 // Handle implicit null exception table updates
2094 if (n->is_MachNullCheck()) {
2095 uint block_num = block->non_connector_successor(0)->_pre_order;
2096 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2097 continue;
2098 }
2099 // Handle implicit exception table updates: trap instructions.
2100 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
2101 uint block_num = block->non_connector_successor(0)->_pre_order;
2102 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
2103 continue;
2104 }
2105 } // End of for all blocks fill in exception table entries
2106 }
2107
2108 // Static Variables
2109 #ifndef PRODUCT
2110 uint Scheduling::_total_nop_size = 0;
2111 uint Scheduling::_total_method_size = 0;
2112 uint Scheduling::_total_branches = 0;
2113 uint Scheduling::_total_unconditional_delays = 0;
2114 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
2115 #endif
2116
2117 // Initializer for class Scheduling
2118
2119 Scheduling::Scheduling(Arena *arena, Compile &compile)
2120 : _arena(arena),
2121 _cfg(compile.cfg()),
2122 _regalloc(compile.regalloc()),
2123 _scheduled(arena),
2124 _available(arena),
2125 _reg_node(arena),
2126 _pinch_free_list(arena),
2127 _next_node(NULL),
2128 _bundle_instr_count(0),
2129 _bundle_cycle_number(0),
2130 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
2131 #ifndef PRODUCT
2132 , _branches(0)
2133 , _unconditional_delays(0)
2134 #endif
2135 {
2136 // Create a MachNopNode
2137 _nop = new MachNopNode();
2138
2139 // Now that the nops are in the array, save the count
2140 // (but allow entries for the nops)
2141 _node_bundling_limit = compile.unique();
2142 uint node_max = _regalloc->node_regs_max_index();
2143
2144 compile.output()->set_node_bundling_limit(_node_bundling_limit);
2145
2146 // This one is persistent within the Compile class
2147 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
2148
2149 // Allocate space for fixed-size arrays
2150 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
2151 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
2152
2153 // Clear the arrays
2154 for (uint i = 0; i < node_max; i++) {
2155 ::new (&_node_bundling_base[i]) Bundle();
2156 }
2157 memset(_uses, 0, node_max * sizeof(short));
2158 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2159
2160 // Clear the bundling information
2161 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2162
2163 // Get the last node
2164 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2165
2166 _next_node = block->get_node(block->number_of_nodes() - 1);
2167 }
2168
2169 #ifndef PRODUCT
2170 // Scheduling destructor
2171 Scheduling::~Scheduling() {
2172 _total_branches += _branches;
2173 _total_unconditional_delays += _unconditional_delays;
2174 }
2175 #endif
2176
2177 // Step ahead "i" cycles
2178 void Scheduling::step(uint i) {
2179
2180 Bundle *bundle = node_bundling(_next_node);
2181 bundle->set_starts_bundle();
2182
2183 // Update the bundle record, but leave the flags information alone
2184 if (_bundle_instr_count > 0) {
2185 bundle->set_instr_count(_bundle_instr_count);
2186 bundle->set_resources_used(_bundle_use.resourcesUsed());
2187 }
2188
2189 // Update the state information
2190 _bundle_instr_count = 0;
2191 _bundle_cycle_number += i;
2192 _bundle_use.step(i);
2193 }
2194
2195 void Scheduling::step_and_clear() {
2196 Bundle *bundle = node_bundling(_next_node);
2197 bundle->set_starts_bundle();
2198
2199 // Update the bundle record
2200 if (_bundle_instr_count > 0) {
2201 bundle->set_instr_count(_bundle_instr_count);
2202 bundle->set_resources_used(_bundle_use.resourcesUsed());
2203
2204 _bundle_cycle_number += 1;
2205 }
2206
2207 // Clear the bundling information
2208 _bundle_instr_count = 0;
2209 _bundle_use.reset();
2210
2211 memcpy(_bundle_use_elements,
2212 Pipeline_Use::elaborated_elements,
2213 sizeof(Pipeline_Use::elaborated_elements));
2214 }
2215
2216 // Perform instruction scheduling and bundling over the sequence of
2217 // instructions in backwards order.
2218 void PhaseOutput::ScheduleAndBundle() {
2219
2220 // Don't optimize this if it isn't a method
2221 if (!C->method())
2222 return;
2223
2224 // Don't optimize this if scheduling is disabled
2225 if (!C->do_scheduling())
2226 return;
2227
2228 // Scheduling code works only with pairs (8 bytes) maximum.
2229 // And when the scalable vector register is used, we may spill/unspill
2230 // the whole reg regardless of the max vector size.
2231 if (C->max_vector_size() > 8 ||
2232 (C->max_vector_size() > 0 && Matcher::supports_scalable_vector())) {
2233 return;
2234 }
2235
2236 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2237
2238 // Create a data structure for all the scheduling information
2239 Scheduling scheduling(Thread::current()->resource_area(), *C);
2240
2241 // Walk backwards over each basic block, computing the needed alignment
2242 // Walk over all the basic blocks
2243 scheduling.DoScheduling();
2244
2245 #ifndef PRODUCT
2246 if (C->trace_opto_output()) {
2247 tty->print("\n---- After ScheduleAndBundle ----\n");
2248 print_scheduling();
2249 }
2250 #endif
2251 }
2252
2253 #ifndef PRODUCT
2254 // Separated out so that it can be called directly from debugger
2255 void PhaseOutput::print_scheduling() {
2256 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2257 tty->print("\nBB#%03d:\n", i);
2258 Block* block = C->cfg()->get_block(i);
2259 for (uint j = 0; j < block->number_of_nodes(); j++) {
2260 Node* n = block->get_node(j);
2261 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2262 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2263 n->dump();
2264 }
2265 }
2266 }
2267 #endif
2268
2269 // See if this node fits into the present instruction bundle
2270 bool Scheduling::NodeFitsInBundle(Node *n) {
2271 uint n_idx = n->_idx;
2272
2273 // If this is the unconditional delay instruction, then it fits
2274 if (n == _unconditional_delay_slot) {
2275 #ifndef PRODUCT
2276 if (_cfg->C->trace_opto_output())
2277 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2278 #endif
2279 return (true);
2280 }
2281
2282 // If the node cannot be scheduled this cycle, skip it
2283 if (_current_latency[n_idx] > _bundle_cycle_number) {
2284 #ifndef PRODUCT
2285 if (_cfg->C->trace_opto_output())
2286 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2287 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2288 #endif
2289 return (false);
2290 }
2291
2292 const Pipeline *node_pipeline = n->pipeline();
2293
2294 uint instruction_count = node_pipeline->instructionCount();
2295 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2296 instruction_count = 0;
2297 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2298 instruction_count++;
2299
2300 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2301 #ifndef PRODUCT
2302 if (_cfg->C->trace_opto_output())
2303 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2304 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2305 #endif
2306 return (false);
2307 }
2308
2309 // Don't allow non-machine nodes to be handled this way
2310 if (!n->is_Mach() && instruction_count == 0)
2311 return (false);
2312
2313 // See if there is any overlap
2314 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2315
2316 if (delay > 0) {
2317 #ifndef PRODUCT
2318 if (_cfg->C->trace_opto_output())
2319 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2320 #endif
2321 return false;
2322 }
2323
2324 #ifndef PRODUCT
2325 if (_cfg->C->trace_opto_output())
2326 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2327 #endif
2328
2329 return true;
2330 }
2331
2332 Node * Scheduling::ChooseNodeToBundle() {
2333 uint siz = _available.size();
2334
2335 if (siz == 0) {
2336
2337 #ifndef PRODUCT
2338 if (_cfg->C->trace_opto_output())
2339 tty->print("# ChooseNodeToBundle: NULL\n");
2340 #endif
2341 return (NULL);
2342 }
2343
2344 // Fast path, if only 1 instruction in the bundle
2345 if (siz == 1) {
2346 #ifndef PRODUCT
2347 if (_cfg->C->trace_opto_output()) {
2348 tty->print("# ChooseNodeToBundle (only 1): ");
2349 _available[0]->dump();
2350 }
2351 #endif
2352 return (_available[0]);
2353 }
2354
2355 // Don't bother, if the bundle is already full
2356 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2357 for ( uint i = 0; i < siz; i++ ) {
2358 Node *n = _available[i];
2359
2360 // Skip projections, we'll handle them another way
2361 if (n->is_Proj())
2362 continue;
2363
2364 // This presupposed that instructions are inserted into the
2365 // available list in a legality order; i.e. instructions that
2366 // must be inserted first are at the head of the list
2367 if (NodeFitsInBundle(n)) {
2368 #ifndef PRODUCT
2369 if (_cfg->C->trace_opto_output()) {
2370 tty->print("# ChooseNodeToBundle: ");
2371 n->dump();
2372 }
2373 #endif
2374 return (n);
2375 }
2376 }
2377 }
2378
2379 // Nothing fits in this bundle, choose the highest priority
2380 #ifndef PRODUCT
2381 if (_cfg->C->trace_opto_output()) {
2382 tty->print("# ChooseNodeToBundle: ");
2383 _available[0]->dump();
2384 }
2385 #endif
2386
2387 return _available[0];
2388 }
2389
2390 void Scheduling::AddNodeToAvailableList(Node *n) {
2391 assert( !n->is_Proj(), "projections never directly made available" );
2392 #ifndef PRODUCT
2393 if (_cfg->C->trace_opto_output()) {
2394 tty->print("# AddNodeToAvailableList: ");
2395 n->dump();
2396 }
2397 #endif
2398
2399 int latency = _current_latency[n->_idx];
2400
2401 // Insert in latency order (insertion sort)
2402 uint i;
2403 for ( i=0; i < _available.size(); i++ )
2404 if (_current_latency[_available[i]->_idx] > latency)
2405 break;
2406
2407 // Special Check for compares following branches
2408 if( n->is_Mach() && _scheduled.size() > 0 ) {
2409 int op = n->as_Mach()->ideal_Opcode();
2410 Node *last = _scheduled[0];
2411 if( last->is_MachIf() && last->in(1) == n &&
2412 ( op == Op_CmpI ||
2413 op == Op_CmpU ||
2414 op == Op_CmpUL ||
2415 op == Op_CmpP ||
2416 op == Op_CmpF ||
2417 op == Op_CmpD ||
2418 op == Op_CmpL ) ) {
2419
2420 // Recalculate position, moving to front of same latency
2421 for ( i=0 ; i < _available.size(); i++ )
2422 if (_current_latency[_available[i]->_idx] >= latency)
2423 break;
2424 }
2425 }
2426
2427 // Insert the node in the available list
2428 _available.insert(i, n);
2429
2430 #ifndef PRODUCT
2431 if (_cfg->C->trace_opto_output())
2432 dump_available();
2433 #endif
2434 }
2435
2436 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2437 for ( uint i=0; i < n->len(); i++ ) {
2438 Node *def = n->in(i);
2439 if (!def) continue;
2440 if( def->is_Proj() ) // If this is a machine projection, then
2441 def = def->in(0); // propagate usage thru to the base instruction
2442
2443 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2444 continue;
2445 }
2446
2447 // Compute the latency
2448 uint l = _bundle_cycle_number + n->latency(i);
2449 if (_current_latency[def->_idx] < l)
2450 _current_latency[def->_idx] = l;
2451
2452 // If this does not have uses then schedule it
2453 if ((--_uses[def->_idx]) == 0)
2454 AddNodeToAvailableList(def);
2455 }
2456 }
2457
2458 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2459 #ifndef PRODUCT
2460 if (_cfg->C->trace_opto_output()) {
2461 tty->print("# AddNodeToBundle: ");
2462 n->dump();
2463 }
2464 #endif
2465
2466 // Remove this from the available list
2467 uint i;
2468 for (i = 0; i < _available.size(); i++)
2469 if (_available[i] == n)
2470 break;
2471 assert(i < _available.size(), "entry in _available list not found");
2472 _available.remove(i);
2473
2474 // See if this fits in the current bundle
2475 const Pipeline *node_pipeline = n->pipeline();
2476 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2477
2478 // Check for instructions to be placed in the delay slot. We
2479 // do this before we actually schedule the current instruction,
2480 // because the delay slot follows the current instruction.
2481 if (Pipeline::_branch_has_delay_slot &&
2482 node_pipeline->hasBranchDelay() &&
2483 !_unconditional_delay_slot) {
2484
2485 uint siz = _available.size();
2486
2487 // Conditional branches can support an instruction that
2488 // is unconditionally executed and not dependent by the
2489 // branch, OR a conditionally executed instruction if
2490 // the branch is taken. In practice, this means that
2491 // the first instruction at the branch target is
2492 // copied to the delay slot, and the branch goes to
2493 // the instruction after that at the branch target
2494 if ( n->is_MachBranch() ) {
2495
2496 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2497 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2498
2499 #ifndef PRODUCT
2500 _branches++;
2501 #endif
2502
2503 // At least 1 instruction is on the available list
2504 // that is not dependent on the branch
2505 for (uint i = 0; i < siz; i++) {
2506 Node *d = _available[i];
2507 const Pipeline *avail_pipeline = d->pipeline();
2508
2509 // Don't allow safepoints in the branch shadow, that will
2510 // cause a number of difficulties
2511 if ( avail_pipeline->instructionCount() == 1 &&
2512 !avail_pipeline->hasMultipleBundles() &&
2513 !avail_pipeline->hasBranchDelay() &&
2514 Pipeline::instr_has_unit_size() &&
2515 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2516 NodeFitsInBundle(d) &&
2517 !node_bundling(d)->used_in_delay()) {
2518
2519 if (d->is_Mach() && !d->is_MachSafePoint()) {
2520 // A node that fits in the delay slot was found, so we need to
2521 // set the appropriate bits in the bundle pipeline information so
2522 // that it correctly indicates resource usage. Later, when we
2523 // attempt to add this instruction to the bundle, we will skip
2524 // setting the resource usage.
2525 _unconditional_delay_slot = d;
2526 node_bundling(n)->set_use_unconditional_delay();
2527 node_bundling(d)->set_used_in_unconditional_delay();
2528 _bundle_use.add_usage(avail_pipeline->resourceUse());
2529 _current_latency[d->_idx] = _bundle_cycle_number;
2530 _next_node = d;
2531 ++_bundle_instr_count;
2532 #ifndef PRODUCT
2533 _unconditional_delays++;
2534 #endif
2535 break;
2536 }
2537 }
2538 }
2539 }
2540
2541 // No delay slot, add a nop to the usage
2542 if (!_unconditional_delay_slot) {
2543 // See if adding an instruction in the delay slot will overflow
2544 // the bundle.
2545 if (!NodeFitsInBundle(_nop)) {
2546 #ifndef PRODUCT
2547 if (_cfg->C->trace_opto_output())
2548 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2549 #endif
2550 step(1);
2551 }
2552
2553 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2554 _next_node = _nop;
2555 ++_bundle_instr_count;
2556 }
2557
2558 // See if the instruction in the delay slot requires a
2559 // step of the bundles
2560 if (!NodeFitsInBundle(n)) {
2561 #ifndef PRODUCT
2562 if (_cfg->C->trace_opto_output())
2563 tty->print("# *** STEP(branch won't fit) ***\n");
2564 #endif
2565 // Update the state information
2566 _bundle_instr_count = 0;
2567 _bundle_cycle_number += 1;
2568 _bundle_use.step(1);
2569 }
2570 }
2571
2572 // Get the number of instructions
2573 uint instruction_count = node_pipeline->instructionCount();
2574 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2575 instruction_count = 0;
2576
2577 // Compute the latency information
2578 uint delay = 0;
2579
2580 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2581 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2582 if (relative_latency < 0)
2583 relative_latency = 0;
2584
2585 delay = _bundle_use.full_latency(relative_latency, node_usage);
2586
2587 // Does not fit in this bundle, start a new one
2588 if (delay > 0) {
2589 step(delay);
2590
2591 #ifndef PRODUCT
2592 if (_cfg->C->trace_opto_output())
2593 tty->print("# *** STEP(%d) ***\n", delay);
2594 #endif
2595 }
2596 }
2597
2598 // If this was placed in the delay slot, ignore it
2599 if (n != _unconditional_delay_slot) {
2600
2601 if (delay == 0) {
2602 if (node_pipeline->hasMultipleBundles()) {
2603 #ifndef PRODUCT
2604 if (_cfg->C->trace_opto_output())
2605 tty->print("# *** STEP(multiple instructions) ***\n");
2606 #endif
2607 step(1);
2608 }
2609
2610 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2611 #ifndef PRODUCT
2612 if (_cfg->C->trace_opto_output())
2613 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2614 instruction_count + _bundle_instr_count,
2615 Pipeline::_max_instrs_per_cycle);
2616 #endif
2617 step(1);
2618 }
2619 }
2620
2621 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2622 _bundle_instr_count++;
2623
2624 // Set the node's latency
2625 _current_latency[n->_idx] = _bundle_cycle_number;
2626
2627 // Now merge the functional unit information
2628 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2629 _bundle_use.add_usage(node_usage);
2630
2631 // Increment the number of instructions in this bundle
2632 _bundle_instr_count += instruction_count;
2633
2634 // Remember this node for later
2635 if (n->is_Mach())
2636 _next_node = n;
2637 }
2638
2639 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2640 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2641 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2642 // into the block. All other scheduled nodes get put in the schedule here.
2643 int op = n->Opcode();
2644 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2645 (op != Op_Node && // Not an unused antidepedence node and
2646 // not an unallocated boxlock
2647 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2648
2649 // Push any trailing projections
2650 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2651 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2652 Node *foi = n->fast_out(i);
2653 if( foi->is_Proj() )
2654 _scheduled.push(foi);
2655 }
2656 }
2657
2658 // Put the instruction in the schedule list
2659 _scheduled.push(n);
2660 }
2661
2662 #ifndef PRODUCT
2663 if (_cfg->C->trace_opto_output())
2664 dump_available();
2665 #endif
2666
2667 // Walk all the definitions, decrementing use counts, and
2668 // if a definition has a 0 use count, place it in the available list.
2669 DecrementUseCounts(n,bb);
2670 }
2671
2672 // This method sets the use count within a basic block. We will ignore all
2673 // uses outside the current basic block. As we are doing a backwards walk,
2674 // any node we reach that has a use count of 0 may be scheduled. This also
2675 // avoids the problem of cyclic references from phi nodes, as long as phi
2676 // nodes are at the front of the basic block. This method also initializes
2677 // the available list to the set of instructions that have no uses within this
2678 // basic block.
2679 void Scheduling::ComputeUseCount(const Block *bb) {
2680 #ifndef PRODUCT
2681 if (_cfg->C->trace_opto_output())
2682 tty->print("# -> ComputeUseCount\n");
2683 #endif
2684
2685 // Clear the list of available and scheduled instructions, just in case
2686 _available.clear();
2687 _scheduled.clear();
2688
2689 // No delay slot specified
2690 _unconditional_delay_slot = NULL;
2691
2692 #ifdef ASSERT
2693 for( uint i=0; i < bb->number_of_nodes(); i++ )
2694 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2695 #endif
2696
2697 // Force the _uses count to never go to zero for unscheduable pieces
2698 // of the block
2699 for( uint k = 0; k < _bb_start; k++ )
2700 _uses[bb->get_node(k)->_idx] = 1;
2701 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2702 _uses[bb->get_node(l)->_idx] = 1;
2703
2704 // Iterate backwards over the instructions in the block. Don't count the
2705 // branch projections at end or the block header instructions.
2706 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2707 Node *n = bb->get_node(j);
2708 if( n->is_Proj() ) continue; // Projections handled another way
2709
2710 // Account for all uses
2711 for ( uint k = 0; k < n->len(); k++ ) {
2712 Node *inp = n->in(k);
2713 if (!inp) continue;
2714 assert(inp != n, "no cycles allowed" );
2715 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2716 if (inp->is_Proj()) { // Skip through Proj's
2717 inp = inp->in(0);
2718 }
2719 ++_uses[inp->_idx]; // Count 1 block-local use
2720 }
2721 }
2722
2723 // If this instruction has a 0 use count, then it is available
2724 if (!_uses[n->_idx]) {
2725 _current_latency[n->_idx] = _bundle_cycle_number;
2726 AddNodeToAvailableList(n);
2727 }
2728
2729 #ifndef PRODUCT
2730 if (_cfg->C->trace_opto_output()) {
2731 tty->print("# uses: %3d: ", _uses[n->_idx]);
2732 n->dump();
2733 }
2734 #endif
2735 }
2736
2737 #ifndef PRODUCT
2738 if (_cfg->C->trace_opto_output())
2739 tty->print("# <- ComputeUseCount\n");
2740 #endif
2741 }
2742
2743 // This routine performs scheduling on each basic block in reverse order,
2744 // using instruction latencies and taking into account function unit
2745 // availability.
2746 void Scheduling::DoScheduling() {
2747 #ifndef PRODUCT
2748 if (_cfg->C->trace_opto_output())
2749 tty->print("# -> DoScheduling\n");
2750 #endif
2751
2752 Block *succ_bb = NULL;
2753 Block *bb;
2754 Compile* C = Compile::current();
2755
2756 // Walk over all the basic blocks in reverse order
2757 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2758 bb = _cfg->get_block(i);
2759
2760 #ifndef PRODUCT
2761 if (_cfg->C->trace_opto_output()) {
2762 tty->print("# Schedule BB#%03d (initial)\n", i);
2763 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2764 bb->get_node(j)->dump();
2765 }
2766 }
2767 #endif
2768
2769 // On the head node, skip processing
2770 if (bb == _cfg->get_root_block()) {
2771 continue;
2772 }
2773
2774 // Skip empty, connector blocks
2775 if (bb->is_connector())
2776 continue;
2777
2778 // If the following block is not the sole successor of
2779 // this one, then reset the pipeline information
2780 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2781 #ifndef PRODUCT
2782 if (_cfg->C->trace_opto_output()) {
2783 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2784 _next_node->_idx, _bundle_instr_count);
2785 }
2786 #endif
2787 step_and_clear();
2788 }
2789
2790 // Leave untouched the starting instruction, any Phis, a CreateEx node
2791 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2792 _bb_end = bb->number_of_nodes()-1;
2793 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2794 Node *n = bb->get_node(_bb_start);
2795 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2796 // Also, MachIdealNodes do not get scheduled
2797 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2798 MachNode *mach = n->as_Mach();
2799 int iop = mach->ideal_Opcode();
2800 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2801 if( iop == Op_Con ) continue; // Do not schedule Top
2802 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2803 mach->pipeline() == MachNode::pipeline_class() &&
2804 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2805 continue;
2806 break; // Funny loop structure to be sure...
2807 }
2808 // Compute last "interesting" instruction in block - last instruction we
2809 // might schedule. _bb_end points just after last schedulable inst. We
2810 // normally schedule conditional branches (despite them being forced last
2811 // in the block), because they have delay slots we can fill. Calls all
2812 // have their delay slots filled in the template expansions, so we don't
2813 // bother scheduling them.
2814 Node *last = bb->get_node(_bb_end);
2815 // Ignore trailing NOPs.
2816 while (_bb_end > 0 && last->is_Mach() &&
2817 last->as_Mach()->ideal_Opcode() == Op_Con) {
2818 last = bb->get_node(--_bb_end);
2819 }
2820 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2821 if( last->is_Catch() ||
2822 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2823 // There might be a prior call. Skip it.
2824 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2825 } else if( last->is_MachNullCheck() ) {
2826 // Backup so the last null-checked memory instruction is
2827 // outside the schedulable range. Skip over the nullcheck,
2828 // projection, and the memory nodes.
2829 Node *mem = last->in(1);
2830 do {
2831 _bb_end--;
2832 } while (mem != bb->get_node(_bb_end));
2833 } else {
2834 // Set _bb_end to point after last schedulable inst.
2835 _bb_end++;
2836 }
2837
2838 assert( _bb_start <= _bb_end, "inverted block ends" );
2839
2840 // Compute the register antidependencies for the basic block
2841 ComputeRegisterAntidependencies(bb);
2842 if (C->failing()) return; // too many D-U pinch points
2843
2844 // Compute the usage within the block, and set the list of all nodes
2845 // in the block that have no uses within the block.
2846 ComputeUseCount(bb);
2847
2848 // Schedule the remaining instructions in the block
2849 while ( _available.size() > 0 ) {
2850 Node *n = ChooseNodeToBundle();
2851 guarantee(n != NULL, "no nodes available");
2852 AddNodeToBundle(n,bb);
2853 }
2854
2855 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2856 #ifdef ASSERT
2857 for( uint l = _bb_start; l < _bb_end; l++ ) {
2858 Node *n = bb->get_node(l);
2859 uint m;
2860 for( m = 0; m < _bb_end-_bb_start; m++ )
2861 if( _scheduled[m] == n )
2862 break;
2863 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2864 }
2865 #endif
2866
2867 // Now copy the instructions (in reverse order) back to the block
2868 for ( uint k = _bb_start; k < _bb_end; k++ )
2869 bb->map_node(_scheduled[_bb_end-k-1], k);
2870
2871 #ifndef PRODUCT
2872 if (_cfg->C->trace_opto_output()) {
2873 tty->print("# Schedule BB#%03d (final)\n", i);
2874 uint current = 0;
2875 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2876 Node *n = bb->get_node(j);
2877 if( valid_bundle_info(n) ) {
2878 Bundle *bundle = node_bundling(n);
2879 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2880 tty->print("*** Bundle: ");
2881 bundle->dump();
2882 }
2883 n->dump();
2884 }
2885 }
2886 }
2887 #endif
2888 #ifdef ASSERT
2889 verify_good_schedule(bb,"after block local scheduling");
2890 #endif
2891 }
2892
2893 #ifndef PRODUCT
2894 if (_cfg->C->trace_opto_output())
2895 tty->print("# <- DoScheduling\n");
2896 #endif
2897
2898 // Record final node-bundling array location
2899 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2900
2901 } // end DoScheduling
2902
2903 // Verify that no live-range used in the block is killed in the block by a
2904 // wrong DEF. This doesn't verify live-ranges that span blocks.
2905
2906 // Check for edge existence. Used to avoid adding redundant precedence edges.
2907 static bool edge_from_to( Node *from, Node *to ) {
2908 for( uint i=0; i<from->len(); i++ )
2909 if( from->in(i) == to )
2910 return true;
2911 return false;
2912 }
2913
2914 #ifdef ASSERT
2915 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2916 // Check for bad kills
2917 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2918 Node *prior_use = _reg_node[def];
2919 if( prior_use && !edge_from_to(prior_use,n) ) {
2920 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2921 n->dump();
2922 tty->print_cr("...");
2923 prior_use->dump();
2924 assert(edge_from_to(prior_use,n), "%s", msg);
2925 }
2926 _reg_node.map(def,NULL); // Kill live USEs
2927 }
2928 }
2929
2930 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2931
2932 // Zap to something reasonable for the verify code
2933 _reg_node.clear();
2934
2935 // Walk over the block backwards. Check to make sure each DEF doesn't
2936 // kill a live value (other than the one it's supposed to). Add each
2937 // USE to the live set.
2938 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2939 Node *n = b->get_node(i);
2940 int n_op = n->Opcode();
2941 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2942 // Fat-proj kills a slew of registers
2943 RegMaskIterator rmi(n->out_RegMask());
2944 while (rmi.has_next()) {
2945 OptoReg::Name kill = rmi.next();
2946 verify_do_def(n, kill, msg);
2947 }
2948 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2949 // Get DEF'd registers the normal way
2950 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2951 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2952 }
2953
2954 // Now make all USEs live
2955 for( uint i=1; i<n->req(); i++ ) {
2956 Node *def = n->in(i);
2957 assert(def != 0, "input edge required");
2958 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2959 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2960 if( OptoReg::is_valid(reg_lo) ) {
2961 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2962 _reg_node.map(reg_lo,n);
2963 }
2964 if( OptoReg::is_valid(reg_hi) ) {
2965 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2966 _reg_node.map(reg_hi,n);
2967 }
2968 }
2969
2970 }
2971
2972 // Zap to something reasonable for the Antidependence code
2973 _reg_node.clear();
2974 }
2975 #endif
2976
2977 // Conditionally add precedence edges. Avoid putting edges on Projs.
2978 static void add_prec_edge_from_to( Node *from, Node *to ) {
2979 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2980 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2981 from = from->in(0);
2982 }
2983 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2984 !edge_from_to( from, to ) ) // Avoid duplicate edge
2985 from->add_prec(to);
2986 }
2987
2988 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2989 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2990 return;
2991
2992 if (OptoReg::is_reg(def_reg)) {
2993 VMReg vmreg = OptoReg::as_VMReg(def_reg);
2994 if (vmreg->is_reg() && !vmreg->is_concrete() && !vmreg->prev()->is_concrete()) {
2995 // This is one of the high slots of a vector register.
2996 // ScheduleAndBundle already checked there are no live wide
2997 // vectors in this method so it can be safely ignored.
2998 return;
2999 }
3000 }
3001
3002 Node *pinch = _reg_node[def_reg]; // Get pinch point
3003 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
3004 is_def ) { // Check for a true def (not a kill)
3005 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
3006 return;
3007 }
3008
3009 Node *kill = def; // Rename 'def' to more descriptive 'kill'
3010 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
3011
3012 // After some number of kills there _may_ be a later def
3013 Node *later_def = NULL;
3014
3015 Compile* C = Compile::current();
3016
3017 // Finding a kill requires a real pinch-point.
3018 // Check for not already having a pinch-point.
3019 // Pinch points are Op_Node's.
3020 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
3021 later_def = pinch; // Must be def/kill as optimistic pinch-point
3022 if ( _pinch_free_list.size() > 0) {
3023 pinch = _pinch_free_list.pop();
3024 } else {
3025 pinch = new Node(1); // Pinch point to-be
3026 }
3027 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
3028 _cfg->C->record_method_not_compilable("too many D-U pinch points");
3029 return;
3030 }
3031 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
3032 _reg_node.map(def_reg,pinch); // Record pinch-point
3033 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
3034 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
3035 pinch->init_req(0, C->top()); // set not NULL for the next call
3036 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
3037 later_def = NULL; // and no later def
3038 }
3039 pinch->set_req(0,later_def); // Hook later def so we can find it
3040 } else { // Else have valid pinch point
3041 if( pinch->in(0) ) // If there is a later-def
3042 later_def = pinch->in(0); // Get it
3043 }
3044
3045 // Add output-dependence edge from later def to kill
3046 if( later_def ) // If there is some original def
3047 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
3048
3049 // See if current kill is also a use, and so is forced to be the pinch-point.
3050 if( pinch->Opcode() == Op_Node ) {
3051 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
3052 for( uint i=1; i<uses->req(); i++ ) {
3053 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
3054 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
3055 // Yes, found a use/kill pinch-point
3056 pinch->set_req(0,NULL); //
3057 pinch->replace_by(kill); // Move anti-dep edges up
3058 pinch = kill;
3059 _reg_node.map(def_reg,pinch);
3060 return;
3061 }
3062 }
3063 }
3064
3065 // Add edge from kill to pinch-point
3066 add_prec_edge_from_to(kill,pinch);
3067 }
3068
3069 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
3070 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
3071 return;
3072 Node *pinch = _reg_node[use_reg]; // Get pinch point
3073 // Check for no later def_reg/kill in block
3074 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
3075 // Use has to be block-local as well
3076 _cfg->get_block_for_node(use) == b) {
3077 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
3078 pinch->req() == 1 ) { // pinch not yet in block?
3079 pinch->del_req(0); // yank pointer to later-def, also set flag
3080 // Insert the pinch-point in the block just after the last use
3081 b->insert_node(pinch, b->find_node(use) + 1);
3082 _bb_end++; // Increase size scheduled region in block
3083 }
3084
3085 add_prec_edge_from_to(pinch,use);
3086 }
3087 }
3088
3089 // We insert antidependences between the reads and following write of
3090 // allocated registers to prevent illegal code motion. Hopefully, the
3091 // number of added references should be fairly small, especially as we
3092 // are only adding references within the current basic block.
3093 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
3094
3095 #ifdef ASSERT
3096 verify_good_schedule(b,"before block local scheduling");
3097 #endif
3098
3099 // A valid schedule, for each register independently, is an endless cycle
3100 // of: a def, then some uses (connected to the def by true dependencies),
3101 // then some kills (defs with no uses), finally the cycle repeats with a new
3102 // def. The uses are allowed to float relative to each other, as are the
3103 // kills. No use is allowed to slide past a kill (or def). This requires
3104 // antidependencies between all uses of a single def and all kills that
3105 // follow, up to the next def. More edges are redundant, because later defs
3106 // & kills are already serialized with true or antidependencies. To keep
3107 // the edge count down, we add a 'pinch point' node if there's more than
3108 // one use or more than one kill/def.
3109
3110 // We add dependencies in one bottom-up pass.
3111
3112 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
3113
3114 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
3115 // register. If not, we record the DEF/KILL in _reg_node, the
3116 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
3117 // "pinch point", a new Node that's in the graph but not in the block.
3118 // We put edges from the prior and current DEF/KILLs to the pinch point.
3119 // We put the pinch point in _reg_node. If there's already a pinch point
3120 // we merely add an edge from the current DEF/KILL to the pinch point.
3121
3122 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3123 // put an edge from the pinch point to the USE.
3124
3125 // To be expedient, the _reg_node array is pre-allocated for the whole
3126 // compilation. _reg_node is lazily initialized; it either contains a NULL,
3127 // or a valid def/kill/pinch-point, or a leftover node from some prior
3128 // block. Leftover node from some prior block is treated like a NULL (no
3129 // prior def, so no anti-dependence needed). Valid def is distinguished by
3130 // it being in the current block.
3131 bool fat_proj_seen = false;
3132 uint last_safept = _bb_end-1;
3133 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
3134 Node* last_safept_node = end_node;
3135 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3136 Node *n = b->get_node(i);
3137 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3138 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3139 // Fat-proj kills a slew of registers
3140 // This can add edges to 'n' and obscure whether or not it was a def,
3141 // hence the is_def flag.
3142 fat_proj_seen = true;
3143 RegMaskIterator rmi(n->out_RegMask());
3144 while (rmi.has_next()) {
3145 OptoReg::Name kill = rmi.next();
3146 anti_do_def(b, n, kill, is_def);
3147 }
3148 } else {
3149 // Get DEF'd registers the normal way
3150 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3151 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3152 }
3153
3154 // Kill projections on a branch should appear to occur on the
3155 // branch, not afterwards, so grab the masks from the projections
3156 // and process them.
3157 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3158 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3159 Node* use = n->fast_out(i);
3160 if (use->is_Proj()) {
3161 RegMaskIterator rmi(use->out_RegMask());
3162 while (rmi.has_next()) {
3163 OptoReg::Name kill = rmi.next();
3164 anti_do_def(b, n, kill, false);
3165 }
3166 }
3167 }
3168 }
3169
3170 // Check each register used by this instruction for a following DEF/KILL
3171 // that must occur afterward and requires an anti-dependence edge.
3172 for( uint j=0; j<n->req(); j++ ) {
3173 Node *def = n->in(j);
3174 if( def ) {
3175 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3176 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3177 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3178 }
3179 }
3180 // Do not allow defs of new derived values to float above GC
3181 // points unless the base is definitely available at the GC point.
3182
3183 Node *m = b->get_node(i);
3184
3185 // Add precedence edge from following safepoint to use of derived pointer
3186 if( last_safept_node != end_node &&
3187 m != last_safept_node) {
3188 for (uint k = 1; k < m->req(); k++) {
3189 const Type *t = m->in(k)->bottom_type();
3190 if( t->isa_oop_ptr() &&
3191 t->is_ptr()->offset() != 0 ) {
3192 last_safept_node->add_prec( m );
3193 break;
3194 }
3195 }
3196
3197 // Do not allow a CheckCastPP node whose input is a raw pointer to
3198 // float past a safepoint. This can occur when a buffered inline
3199 // type is allocated in a loop and the CheckCastPP from that
3200 // allocation is reused outside the loop. If the use inside the
3201 // loop is scalarized the CheckCastPP will no longer be connected
3202 // to the loop safepoint. See JDK-8264340.
3203 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CheckCastPP) {
3204 Node *def = m->in(1);
3205 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
3206 last_safept_node->add_prec(m);
3207 }
3208 }
3209 }
3210
3211 if( n->jvms() ) { // Precedence edge from derived to safept
3212 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3213 if( b->get_node(last_safept) != last_safept_node ) {
3214 last_safept = b->find_node(last_safept_node);
3215 }
3216 for( uint j=last_safept; j > i; j-- ) {
3217 Node *mach = b->get_node(j);
3218 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3219 mach->add_prec( n );
3220 }
3221 last_safept = i;
3222 last_safept_node = m;
3223 }
3224 }
3225
3226 if (fat_proj_seen) {
3227 // Garbage collect pinch nodes that were not consumed.
3228 // They are usually created by a fat kill MachProj for a call.
3229 garbage_collect_pinch_nodes();
3230 }
3231 }
3232
3233 // Garbage collect pinch nodes for reuse by other blocks.
3234 //
3235 // The block scheduler's insertion of anti-dependence
3236 // edges creates many pinch nodes when the block contains
3237 // 2 or more Calls. A pinch node is used to prevent a
3238 // combinatorial explosion of edges. If a set of kills for a
3239 // register is anti-dependent on a set of uses (or defs), rather
3240 // than adding an edge in the graph between each pair of kill
3241 // and use (or def), a pinch is inserted between them:
3242 //
3243 // use1 use2 use3
3244 // \ | /
3245 // \ | /
3246 // pinch
3247 // / | \
3248 // / | \
3249 // kill1 kill2 kill3
3250 //
3251 // One pinch node is created per register killed when
3252 // the second call is encountered during a backwards pass
3253 // over the block. Most of these pinch nodes are never
3254 // wired into the graph because the register is never
3255 // used or def'ed in the block.
3256 //
3257 void Scheduling::garbage_collect_pinch_nodes() {
3258 #ifndef PRODUCT
3259 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3260 #endif
3261 int trace_cnt = 0;
3262 for (uint k = 0; k < _reg_node.Size(); k++) {
3263 Node* pinch = _reg_node[k];
3264 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
3265 // no predecence input edges
3266 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
3267 cleanup_pinch(pinch);
3268 _pinch_free_list.push(pinch);
3269 _reg_node.map(k, NULL);
3270 #ifndef PRODUCT
3271 if (_cfg->C->trace_opto_output()) {
3272 trace_cnt++;
3273 if (trace_cnt > 40) {
3274 tty->print("\n");
3275 trace_cnt = 0;
3276 }
3277 tty->print(" %d", pinch->_idx);
3278 }
3279 #endif
3280 }
3281 }
3282 #ifndef PRODUCT
3283 if (_cfg->C->trace_opto_output()) tty->print("\n");
3284 #endif
3285 }
3286
3287 // Clean up a pinch node for reuse.
3288 void Scheduling::cleanup_pinch( Node *pinch ) {
3289 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3290
3291 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3292 Node* use = pinch->last_out(i);
3293 uint uses_found = 0;
3294 for (uint j = use->req(); j < use->len(); j++) {
3295 if (use->in(j) == pinch) {
3296 use->rm_prec(j);
3297 uses_found++;
3298 }
3299 }
3300 assert(uses_found > 0, "must be a precedence edge");
3301 i -= uses_found; // we deleted 1 or more copies of this edge
3302 }
3303 // May have a later_def entry
3304 pinch->set_req(0, NULL);
3305 }
3306
3307 #ifndef PRODUCT
3308
3309 void Scheduling::dump_available() const {
3310 tty->print("#Availist ");
3311 for (uint i = 0; i < _available.size(); i++)
3312 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3313 tty->cr();
3314 }
3315
3316 // Print Scheduling Statistics
3317 void Scheduling::print_statistics() {
3318 // Print the size added by nops for bundling
3319 tty->print("Nops added %d bytes to total of %d bytes",
3320 _total_nop_size, _total_method_size);
3321 if (_total_method_size > 0)
3322 tty->print(", for %.2f%%",
3323 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3324 tty->print("\n");
3325
3326 // Print the number of branch shadows filled
3327 if (Pipeline::_branch_has_delay_slot) {
3328 tty->print("Of %d branches, %d had unconditional delay slots filled",
3329 _total_branches, _total_unconditional_delays);
3330 if (_total_branches > 0)
3331 tty->print(", for %.2f%%",
3332 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3333 tty->print("\n");
3334 }
3335
3336 uint total_instructions = 0, total_bundles = 0;
3337
3338 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3339 uint bundle_count = _total_instructions_per_bundle[i];
3340 total_instructions += bundle_count * i;
3341 total_bundles += bundle_count;
3342 }
3343
3344 if (total_bundles > 0)
3345 tty->print("Average ILP (excluding nops) is %.2f\n",
3346 ((double)total_instructions) / ((double)total_bundles));
3347 }
3348 #endif
3349
3350 //-----------------------init_scratch_buffer_blob------------------------------
3351 // Construct a temporary BufferBlob and cache it for this compile.
3352 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3353 // If there is already a scratch buffer blob allocated and the
3354 // constant section is big enough, use it. Otherwise free the
3355 // current and allocate a new one.
3356 BufferBlob* blob = scratch_buffer_blob();
3357 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
3358 // Use the current blob.
3359 } else {
3360 if (blob != NULL) {
3361 BufferBlob::free(blob);
3362 }
3363
3364 ResourceMark rm;
3365 _scratch_const_size = const_size;
3366 int size = C2Compiler::initial_code_buffer_size(const_size);
3367 if (C->has_scalarized_args()) {
3368 // Inline type entry points (MachVEPNodes) require lots of space for GC barriers and oop verification
3369 // when loading object fields from the buffered argument. Increase scratch buffer size accordingly.
3370 ciMethod* method = C->method();
3371 int barrier_size = UseZGC ? 200 : (7 DEBUG_ONLY(+ 37));
3372 int arg_num = 0;
3373 if (!method->is_static()) {
3374 if (method->is_scalarized_arg(arg_num)) {
3375 size += method->holder()->as_inline_klass()->oop_count() * barrier_size;
3376 }
3377 arg_num++;
3378 }
3379 for (ciSignatureStream str(method->signature()); !str.at_return_type(); str.next()) {
3380 if (method->is_scalarized_arg(arg_num)) {
3381 size += str.type()->as_inline_klass()->oop_count() * barrier_size;
3382 }
3383 arg_num++;
3384 }
3385 }
3386 blob = BufferBlob::create("Compile::scratch_buffer", size);
3387 // Record the buffer blob for next time.
3388 set_scratch_buffer_blob(blob);
3389 // Have we run out of code space?
3390 if (scratch_buffer_blob() == NULL) {
3391 // Let CompilerBroker disable further compilations.
3392 C->record_failure("Not enough space for scratch buffer in CodeCache");
3393 return;
3394 }
3395 }
3396
3397 // Initialize the relocation buffers
3398 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3399 set_scratch_locs_memory(locs_buf);
3400 }
3401
3402
3403 //-----------------------scratch_emit_size-------------------------------------
3404 // Helper function that computes size by emitting code
3405 uint PhaseOutput::scratch_emit_size(const Node* n) {
3406 // Start scratch_emit_size section.
3407 set_in_scratch_emit_size(true);
3408
3409 // Emit into a trash buffer and count bytes emitted.
3410 // This is a pretty expensive way to compute a size,
3411 // but it works well enough if seldom used.
3412 // All common fixed-size instructions are given a size
3413 // method by the AD file.
3414 // Note that the scratch buffer blob and locs memory are
3415 // allocated at the beginning of the compile task, and
3416 // may be shared by several calls to scratch_emit_size.
3417 // The allocation of the scratch buffer blob is particularly
3418 // expensive, since it has to grab the code cache lock.
3419 BufferBlob* blob = this->scratch_buffer_blob();
3420 assert(blob != NULL, "Initialize BufferBlob at start");
3421 assert(blob->size() > MAX_inst_size, "sanity");
3422 relocInfo* locs_buf = scratch_locs_memory();
3423 address blob_begin = blob->content_begin();
3424 address blob_end = (address)locs_buf;
3425 assert(blob->contains(blob_end), "sanity");
3426 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3427 buf.initialize_consts_size(_scratch_const_size);
3428 buf.initialize_stubs_size(MAX_stubs_size);
3429 assert(locs_buf != NULL, "sanity");
3430 int lsize = MAX_locs_size / 3;
3431 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3432 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3433 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3434 // Mark as scratch buffer.
3435 buf.consts()->set_scratch_emit();
3436 buf.insts()->set_scratch_emit();
3437 buf.stubs()->set_scratch_emit();
3438
3439 // Do the emission.
3440
3441 Label fakeL; // Fake label for branch instructions.
3442 Label* saveL = NULL;
3443 uint save_bnum = 0;
3444 bool is_branch = n->is_MachBranch();
3445 if (is_branch) {
3446 MacroAssembler masm(&buf);
3447 masm.bind(fakeL);
3448 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3449 n->as_MachBranch()->label_set(&fakeL, 0);
3450 }
3451 n->emit(buf, C->regalloc());
3452
3453 // Emitting into the scratch buffer should not fail
3454 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3455
3456 // Restore label.
3457 if (is_branch) {
3458 n->as_MachBranch()->label_set(saveL, save_bnum);
3459 }
3460
3461 // End scratch_emit_size section.
3462 set_in_scratch_emit_size(false);
3463
3464 return buf.insts_size();
3465 }
3466
3467 void PhaseOutput::install() {
3468 if (!C->should_install_code()) {
3469 return;
3470 } else if (C->stub_function() != NULL) {
3471 install_stub(C->stub_name());
3472 } else {
3473 install_code(C->method(),
3474 C->entry_bci(),
3475 CompileBroker::compiler2(),
3476 C->has_unsafe_access(),
3477 SharedRuntime::is_wide_vector(C->max_vector_size()),
3478 C->rtm_state());
3479 }
3480 }
3481
3482 void PhaseOutput::install_code(ciMethod* target,
3483 int entry_bci,
3484 AbstractCompiler* compiler,
3485 bool has_unsafe_access,
3486 bool has_wide_vectors,
3487 RTMState rtm_state) {
3488 // Check if we want to skip execution of all compiled code.
3489 {
3490 #ifndef PRODUCT
3491 if (OptoNoExecute) {
3492 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3493 return;
3494 }
3495 #endif
3496 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3497
3498 if (C->is_osr_compilation()) {
3499 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3500 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3501 } else {
3502 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3503 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry) == -1) {
3504 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry, _first_block_size);
3505 }
3506 if (_code_offsets.value(CodeOffsets::Verified_Inline_Entry_RO) == -1) {
3507 _code_offsets.set_value(CodeOffsets::Verified_Inline_Entry_RO, _first_block_size);
3508 }
3509 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
3510 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
3511 }
3512 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3513 }
3514
3515 C->env()->register_method(target,
3516 entry_bci,
3517 &_code_offsets,
3518 _orig_pc_slot_offset_in_bytes,
3519 code_buffer(),
3520 frame_size_in_words(),
3521 _oop_map_set,
3522 &_handler_table,
3523 inc_table(),
3524 compiler,
3525 has_unsafe_access,
3526 SharedRuntime::is_wide_vector(C->max_vector_size()),
3527 C->has_monitors(),
3528 0,
3529 C->rtm_state());
3530
3531 if (C->log() != NULL) { // Print code cache state into compiler log
3532 C->log()->code_cache_state();
3533 }
3534 }
3535 }
3536 void PhaseOutput::install_stub(const char* stub_name) {
3537 // Entry point will be accessed using stub_entry_point();
3538 if (code_buffer() == NULL) {
3539 Matcher::soft_match_failure();
3540 } else {
3541 if (PrintAssembly && (WizardMode || Verbose))
3542 tty->print_cr("### Stub::%s", stub_name);
3543
3544 if (!C->failing()) {
3545 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3546
3547 // Make the NMethod
3548 // For now we mark the frame as never safe for profile stackwalking
3549 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3550 code_buffer(),
3551 CodeOffsets::frame_never_safe,
3552 // _code_offsets.value(CodeOffsets::Frame_Complete),
3553 frame_size_in_words(),
3554 oop_map_set(),
3555 false);
3556 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
3557
3558 C->set_stub_entry_point(rs->entry_point());
3559 }
3560 }
3561 }
3562
3563 // Support for bundling info
3564 Bundle* PhaseOutput::node_bundling(const Node *n) {
3565 assert(valid_bundle_info(n), "oob");
3566 return &_node_bundling_base[n->_idx];
3567 }
3568
3569 bool PhaseOutput::valid_bundle_info(const Node *n) {
3570 return (_node_bundling_limit > n->_idx);
3571 }
3572
3573 //------------------------------frame_size_in_words-----------------------------
3574 // frame_slots in units of words
3575 int PhaseOutput::frame_size_in_words() const {
3576 // shift is 0 in LP32 and 1 in LP64
3577 const int shift = (LogBytesPerWord - LogBytesPerInt);
3578 int words = _frame_slots >> shift;
3579 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3580 return words;
3581 }
3582
3583 // To bang the stack of this compiled method we use the stack size
3584 // that the interpreter would need in case of a deoptimization. This
3585 // removes the need to bang the stack in the deoptimization blob which
3586 // in turn simplifies stack overflow handling.
3587 int PhaseOutput::bang_size_in_bytes() const {
3588 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3589 }
3590
3591 //------------------------------dump_asm---------------------------------------
3592 // Dump formatted assembly
3593 #if defined(SUPPORT_OPTO_ASSEMBLY)
3594 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3595
3596 int pc_digits = 3; // #chars required for pc
3597 int sb_chars = 3; // #chars for "start bundle" indicator
3598 int tab_size = 8;
3599 if (pcs != NULL) {
3600 int max_pc = 0;
3601 for (uint i = 0; i < pc_limit; i++) {
3602 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3603 }
3604 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3605 }
3606 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3607
3608 bool cut_short = false;
3609 st->print_cr("#");
3610 st->print("# "); C->tf()->dump_on(st); st->cr();
3611 st->print_cr("#");
3612
3613 // For all blocks
3614 int pc = 0x0; // Program counter
3615 char starts_bundle = ' ';
3616 C->regalloc()->dump_frame();
3617
3618 Node *n = NULL;
3619 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3620 if (VMThread::should_terminate()) {
3621 cut_short = true;
3622 break;
3623 }
3624 Block* block = C->cfg()->get_block(i);
3625 if (block->is_connector() && !Verbose) {
3626 continue;
3627 }
3628 n = block->head();
3629 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3630 pc = pcs[n->_idx];
3631 st->print("%*.*x", pc_digits, pc_digits, pc);
3632 }
3633 st->fill_to(prefix_len);
3634 block->dump_head(C->cfg(), st);
3635 if (block->is_connector()) {
3636 st->fill_to(prefix_len);
3637 st->print_cr("# Empty connector block");
3638 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3639 st->fill_to(prefix_len);
3640 st->print_cr("# Block is sole successor of call");
3641 }
3642
3643 // For all instructions
3644 Node *delay = NULL;
3645 for (uint j = 0; j < block->number_of_nodes(); j++) {
3646 if (VMThread::should_terminate()) {
3647 cut_short = true;
3648 break;
3649 }
3650 n = block->get_node(j);
3651 if (valid_bundle_info(n)) {
3652 Bundle* bundle = node_bundling(n);
3653 if (bundle->used_in_unconditional_delay()) {
3654 delay = n;
3655 continue;
3656 }
3657 if (bundle->starts_bundle()) {
3658 starts_bundle = '+';
3659 }
3660 }
3661
3662 if (WizardMode) {
3663 n->dump();
3664 }
3665
3666 if( !n->is_Region() && // Dont print in the Assembly
3667 !n->is_Phi() && // a few noisely useless nodes
3668 !n->is_Proj() &&
3669 !n->is_MachTemp() &&
3670 !n->is_SafePointScalarObject() &&
3671 !n->is_Catch() && // Would be nice to print exception table targets
3672 !n->is_MergeMem() && // Not very interesting
3673 !n->is_top() && // Debug info table constants
3674 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3675 ) {
3676 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3677 pc = pcs[n->_idx];
3678 st->print("%*.*x", pc_digits, pc_digits, pc);
3679 } else {
3680 st->fill_to(pc_digits);
3681 }
3682 st->print(" %c ", starts_bundle);
3683 starts_bundle = ' ';
3684 st->fill_to(prefix_len);
3685 n->format(C->regalloc(), st);
3686 st->cr();
3687 }
3688
3689 // If we have an instruction with a delay slot, and have seen a delay,
3690 // then back up and print it
3691 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3692 // Coverity finding - Explicit null dereferenced.
3693 guarantee(delay != NULL, "no unconditional delay instruction");
3694 if (WizardMode) delay->dump();
3695
3696 if (node_bundling(delay)->starts_bundle())
3697 starts_bundle = '+';
3698 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3699 pc = pcs[n->_idx];
3700 st->print("%*.*x", pc_digits, pc_digits, pc);
3701 } else {
3702 st->fill_to(pc_digits);
3703 }
3704 st->print(" %c ", starts_bundle);
3705 starts_bundle = ' ';
3706 st->fill_to(prefix_len);
3707 delay->format(C->regalloc(), st);
3708 st->cr();
3709 delay = NULL;
3710 }
3711
3712 // Dump the exception table as well
3713 if( n->is_Catch() && (Verbose || WizardMode) ) {
3714 // Print the exception table for this offset
3715 _handler_table.print_subtable_for(pc);
3716 }
3717 st->bol(); // Make sure we start on a new line
3718 }
3719 st->cr(); // one empty line between blocks
3720 assert(cut_short || delay == NULL, "no unconditional delay branch");
3721 } // End of per-block dump
3722
3723 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3724 }
3725 #endif
3726
3727 #ifndef PRODUCT
3728 void PhaseOutput::print_statistics() {
3729 Scheduling::print_statistics();
3730 }
3731 #endif