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