1 /*
2 * Copyright (c) 1998, 2020, 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 "opto/ad.hpp"
39 #include "opto/block.hpp"
40 #include "opto/c2compiler.hpp"
41 #include "opto/callnode.hpp"
42 #include "opto/cfgnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/machnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/optoreg.hpp"
47 #include "opto/output.hpp"
48 #include "opto/regalloc.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/type.hpp"
52 #include "runtime/handles.inline.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/macros.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/xmlstream.hpp"
57
58 #ifndef PRODUCT
59 #define DEBUG_ARG(x) , x
60 #else
61 #define DEBUG_ARG(x)
62 #endif
63
64 //------------------------------Scheduling----------------------------------
65 // This class contains all the information necessary to implement instruction
66 // scheduling and bundling.
67 class Scheduling {
68
69 private:
70 // Arena to use
71 Arena *_arena;
72
73 // Control-Flow Graph info
74 PhaseCFG *_cfg;
75
76 // Register Allocation info
77 PhaseRegAlloc *_regalloc;
78
79 // Number of nodes in the method
80 uint _node_bundling_limit;
81
82 // List of scheduled nodes. Generated in reverse order
83 Node_List _scheduled;
84
85 // List of nodes currently available for choosing for scheduling
86 Node_List _available;
87
88 // For each instruction beginning a bundle, the number of following
89 // nodes to be bundled with it.
90 Bundle *_node_bundling_base;
91
92 // Mapping from register to Node
93 Node_List _reg_node;
94
95 // Free list for pinch nodes.
96 Node_List _pinch_free_list;
97
98 // Latency from the beginning of the containing basic block (base 1)
99 // for each node.
100 unsigned short *_node_latency;
101
102 // Number of uses of this node within the containing basic block.
103 short *_uses;
104
105 // Schedulable portion of current block. Skips Region/Phi/CreateEx up
106 // front, branch+proj at end. Also skips Catch/CProj (same as
107 // branch-at-end), plus just-prior exception-throwing call.
108 uint _bb_start, _bb_end;
109
110 // Latency from the end of the basic block as scheduled
111 unsigned short *_current_latency;
112
113 // Remember the next node
114 Node *_next_node;
115
116 // Use this for an unconditional branch delay slot
117 Node *_unconditional_delay_slot;
118
119 // Pointer to a Nop
120 MachNopNode *_nop;
121
122 // Length of the current bundle, in instructions
123 uint _bundle_instr_count;
124
125 // Current Cycle number, for computing latencies and bundling
126 uint _bundle_cycle_number;
127
128 // Bundle information
129 Pipeline_Use_Element _bundle_use_elements[resource_count];
130 Pipeline_Use _bundle_use;
131
132 // Dump the available list
133 void dump_available() const;
134
135 public:
136 Scheduling(Arena *arena, Compile &compile);
137
138 // Destructor
139 NOT_PRODUCT( ~Scheduling(); )
140
141 // Step ahead "i" cycles
142 void step(uint i);
143
144 // Step ahead 1 cycle, and clear the bundle state (for example,
145 // at a branch target)
146 void step_and_clear();
147
148 Bundle* node_bundling(const Node *n) {
149 assert(valid_bundle_info(n), "oob");
150 return (&_node_bundling_base[n->_idx]);
151 }
152
153 bool valid_bundle_info(const Node *n) const {
154 return (_node_bundling_limit > n->_idx);
155 }
156
157 bool starts_bundle(const Node *n) const {
158 return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
159 }
160
161 // Do the scheduling
162 void DoScheduling();
163
164 // Compute the local latencies walking forward over the list of
165 // nodes for a basic block
166 void ComputeLocalLatenciesForward(const Block *bb);
167
168 // Compute the register antidependencies within a basic block
169 void ComputeRegisterAntidependencies(Block *bb);
170 void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
171 void verify_good_schedule( Block *b, const char *msg );
172 void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
173 void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
174
175 // Add a node to the current bundle
176 void AddNodeToBundle(Node *n, const Block *bb);
177
178 // Add a node to the list of available nodes
179 void AddNodeToAvailableList(Node *n);
180
181 // Compute the local use count for the nodes in a block, and compute
182 // the list of instructions with no uses in the block as available
183 void ComputeUseCount(const Block *bb);
184
185 // Choose an instruction from the available list to add to the bundle
186 Node * ChooseNodeToBundle();
187
188 // See if this Node fits into the currently accumulating bundle
189 bool NodeFitsInBundle(Node *n);
190
191 // Decrement the use count for a node
192 void DecrementUseCounts(Node *n, const Block *bb);
193
194 // Garbage collect pinch nodes for reuse by other blocks.
195 void garbage_collect_pinch_nodes();
196 // Clean up a pinch node for reuse (helper for above).
197 void cleanup_pinch( Node *pinch );
198
199 // Information for statistics gathering
200 #ifndef PRODUCT
201 private:
202 // Gather information on size of nops relative to total
203 uint _branches, _unconditional_delays;
204
205 static uint _total_nop_size, _total_method_size;
206 static uint _total_branches, _total_unconditional_delays;
207 static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
208
209 public:
210 static void print_statistics();
211
212 static void increment_instructions_per_bundle(uint i) {
213 _total_instructions_per_bundle[i]++;
214 }
215
216 static void increment_nop_size(uint s) {
217 _total_nop_size += s;
218 }
219
220 static void increment_method_size(uint s) {
221 _total_method_size += s;
222 }
223 #endif
224
225 };
226
227
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 _oop_map_set(NULL),
235 _scratch_buffer_blob(NULL),
236 _scratch_locs_memory(NULL),
237 _scratch_const_size(-1),
238 _in_scratch_emit_size(false),
239 _frame_slots(0),
240 _code_offsets(),
241 _node_bundling_limit(0),
242 _node_bundling_base(NULL),
243 _orig_pc_slot(0),
244 _orig_pc_slot_offset_in_bytes(0),
245 _buf_sizes(),
246 _block(NULL),
247 _index(0) {
248 C->set_output(this);
249 if (C->stub_name() == NULL) {
250 _orig_pc_slot = C->fixed_slots() - (sizeof(address) / VMRegImpl::stack_slot_size);
251 }
252 }
253
254 PhaseOutput::~PhaseOutput() {
255 C->set_output(NULL);
256 if (_scratch_buffer_blob != NULL) {
257 BufferBlob::free(_scratch_buffer_blob);
258 }
259 }
260
261 void PhaseOutput::perform_mach_node_analysis() {
262 // Late barrier analysis must be done after schedule and bundle
263 // Otherwise liveness based spilling will fail
264 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
265 bs->late_barrier_analysis();
266
267 pd_perform_mach_node_analysis();
268 }
269
270 // Convert Nodes to instruction bits and pass off to the VM
271 void PhaseOutput::Output() {
272 // RootNode goes
273 assert( C->cfg()->get_root_block()->number_of_nodes() == 0, "" );
274
275 // The number of new nodes (mostly MachNop) is proportional to
276 // the number of java calls and inner loops which are aligned.
277 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
278 C->inner_loops()*(OptoLoopAlignment-1)),
279 "out of nodes before code generation" ) ) {
280 return;
281 }
282 // Make sure I can find the Start Node
283 Block *entry = C->cfg()->get_block(1);
284 Block *broot = C->cfg()->get_root_block();
285
286 const StartNode *start = entry->head()->as_Start();
287
288 // Replace StartNode with prolog
289 MachPrologNode *prolog = new MachPrologNode();
290 entry->map_node(prolog, 0);
291 C->cfg()->map_node_to_block(prolog, entry);
292 C->cfg()->unmap_node_from_block(start); // start is no longer in any block
293
294 // Virtual methods need an unverified entry point
295
296 if( C->is_osr_compilation() ) {
297 if( PoisonOSREntry ) {
298 // TODO: Should use a ShouldNotReachHereNode...
299 C->cfg()->insert( broot, 0, new MachBreakpointNode() );
300 }
301 } else {
302 if( C->method() && !C->method()->flags().is_static() ) {
303 // Insert unvalidated entry point
304 C->cfg()->insert( broot, 0, new MachUEPNode() );
305 }
306
307 }
308
309 // Break before main entry point
310 if ((C->method() && C->directive()->BreakAtExecuteOption) ||
311 (OptoBreakpoint && C->is_method_compilation()) ||
312 (OptoBreakpointOSR && C->is_osr_compilation()) ||
313 (OptoBreakpointC2R && !C->method()) ) {
314 // checking for C->method() means that OptoBreakpoint does not apply to
315 // runtime stubs or frame converters
316 C->cfg()->insert( entry, 1, new MachBreakpointNode() );
317 }
318
319 // Insert epilogs before every return
320 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
321 Block* block = C->cfg()->get_block(i);
322 if (!block->is_connector() && block->non_connector_successor(0) == C->cfg()->get_root_block()) { // Found a program exit point?
323 Node* m = block->end();
324 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
325 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
326 block->add_inst(epilog);
327 C->cfg()->map_node_to_block(epilog, block);
328 }
329 }
330 }
331
332 // Keeper of sizing aspects
333 _buf_sizes = BufferSizingData();
334
335 // Initialize code buffer
336 estimate_buffer_size(_buf_sizes._const);
337 if (C->failing()) return;
338
339 // Pre-compute the length of blocks and replace
340 // long branches with short if machine supports it.
341 // Must be done before ScheduleAndBundle due to SPARC delay slots
342 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, C->cfg()->number_of_blocks() + 1);
343 blk_starts[0] = 0;
344 shorten_branches(blk_starts);
345
346 ScheduleAndBundle();
347 if (C->failing()) {
348 return;
349 }
350
351 perform_mach_node_analysis();
352
353 // Complete sizing of codebuffer
354 CodeBuffer* cb = init_buffer();
355 if (cb == NULL || C->failing()) {
356 return;
357 }
358
359 BuildOopMaps();
360
361 if (C->failing()) {
362 return;
363 }
364
365 fill_buffer(cb, blk_starts);
366 }
367
368 bool PhaseOutput::need_stack_bang(int frame_size_in_bytes) const {
369 // Determine if we need to generate a stack overflow check.
370 // Do it if the method is not a stub function and
371 // has java calls or has frame size > vm_page_size/8.
372 // The debug VM checks that deoptimization doesn't trigger an
373 // unexpected stack overflow (compiled method stack banging should
374 // guarantee it doesn't happen) so we always need the stack bang in
375 // a debug VM.
376 return (UseStackBanging && C->stub_function() == NULL &&
377 (C->has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
378 DEBUG_ONLY(|| true)));
379 }
380
381 bool PhaseOutput::need_register_stack_bang() const {
382 // Determine if we need to generate a register stack overflow check.
383 // This is only used on architectures which have split register
384 // and memory stacks (ie. IA64).
385 // Bang if the method is not a stub function and has java calls
386 return (C->stub_function() == NULL && C->has_java_calls());
387 }
388
389
390 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
391 // of a loop. When aligning a loop we need to provide enough instructions
392 // in cpu's fetch buffer to feed decoders. The loop alignment could be
393 // avoided if we have enough instructions in fetch buffer at the head of a loop.
394 // By default, the size is set to 999999 by Block's constructor so that
395 // a loop will be aligned if the size is not reset here.
396 //
397 // Note: Mach instructions could contain several HW instructions
398 // so the size is estimated only.
399 //
400 void PhaseOutput::compute_loop_first_inst_sizes() {
401 // The next condition is used to gate the loop alignment optimization.
402 // Don't aligned a loop if there are enough instructions at the head of a loop
403 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
404 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
405 // equal to 11 bytes which is the largest address NOP instruction.
406 if (MaxLoopPad < OptoLoopAlignment - 1) {
407 uint last_block = C->cfg()->number_of_blocks() - 1;
408 for (uint i = 1; i <= last_block; i++) {
409 Block* block = C->cfg()->get_block(i);
410 // Check the first loop's block which requires an alignment.
411 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
412 uint sum_size = 0;
413 uint inst_cnt = NumberOfLoopInstrToAlign;
414 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
415
416 // Check subsequent fallthrough blocks if the loop's first
417 // block(s) does not have enough instructions.
418 Block *nb = block;
419 while(inst_cnt > 0 &&
420 i < last_block &&
421 !C->cfg()->get_block(i + 1)->has_loop_alignment() &&
422 !nb->has_successor(block)) {
423 i++;
424 nb = C->cfg()->get_block(i);
425 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, C->regalloc());
426 } // while( inst_cnt > 0 && i < last_block )
427
428 block->set_first_inst_size(sum_size);
429 } // f( b->head()->is_Loop() )
430 } // for( i <= last_block )
431 } // if( MaxLoopPad < OptoLoopAlignment-1 )
432 }
433
434 // The architecture description provides short branch variants for some long
435 // branch instructions. Replace eligible long branches with short branches.
436 void PhaseOutput::shorten_branches(uint* blk_starts) {
437 // Compute size of each block, method size, and relocation information size
438 uint nblocks = C->cfg()->number_of_blocks();
439
440 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
441 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
442 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
443
444 // Collect worst case block paddings
445 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
446 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
447
448 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
449 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
450
451 bool has_short_branch_candidate = false;
452
453 // Initialize the sizes to 0
454 int code_size = 0; // Size in bytes of generated code
455 int stub_size = 0; // Size in bytes of all stub entries
456 // Size in bytes of all relocation entries, including those in local stubs.
457 // Start with 2-bytes of reloc info for the unvalidated entry point
458 int reloc_size = 1; // Number of relocation entries
459
460 // Make three passes. The first computes pessimistic blk_starts,
461 // relative jmp_offset and reloc_size information. The second performs
462 // short branch substitution using the pessimistic sizing. The
463 // third inserts nops where needed.
464
465 // Step one, perform a pessimistic sizing pass.
466 uint last_call_adr = max_juint;
467 uint last_avoid_back_to_back_adr = max_juint;
468 uint nop_size = (new MachNopNode())->size(C->regalloc());
469 for (uint i = 0; i < nblocks; i++) { // For all blocks
470 Block* block = C->cfg()->get_block(i);
471 _block = block;
472
473 // During short branch replacement, we store the relative (to blk_starts)
474 // offset of jump in jmp_offset, rather than the absolute offset of jump.
475 // This is so that we do not need to recompute sizes of all nodes when
476 // we compute correct blk_starts in our next sizing pass.
477 jmp_offset[i] = 0;
478 jmp_size[i] = 0;
479 jmp_nidx[i] = -1;
480 DEBUG_ONLY( jmp_target[i] = 0; )
481 DEBUG_ONLY( jmp_rule[i] = 0; )
482
483 // Sum all instruction sizes to compute block size
484 uint last_inst = block->number_of_nodes();
485 uint blk_size = 0;
486 for (uint j = 0; j < last_inst; j++) {
487 _index = j;
488 Node* nj = block->get_node(_index);
489 // Handle machine instruction nodes
490 if (nj->is_Mach()) {
491 MachNode* mach = nj->as_Mach();
492 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
493 reloc_size += mach->reloc();
494 if (mach->is_MachCall()) {
495 // add size information for trampoline stub
496 // class CallStubImpl is platform-specific and defined in the *.ad files.
497 stub_size += CallStubImpl::size_call_trampoline();
498 reloc_size += CallStubImpl::reloc_call_trampoline();
499
500 MachCallNode *mcall = mach->as_MachCall();
501 // This destination address is NOT PC-relative
502
503 mcall->method_set((intptr_t)mcall->entry_point());
504
505 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
506 stub_size += CompiledStaticCall::to_interp_stub_size();
507 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
508 #if INCLUDE_AOT
509 stub_size += CompiledStaticCall::to_aot_stub_size();
510 reloc_size += CompiledStaticCall::reloc_to_aot_stub();
511 #endif
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) ? NULL: block->get_node(idx)->as_Mach();
594 if (mach != NULL && 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 != NULL && (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 NULL;
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()) == NULL, "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 == NULL) {
759 ciKlass* cik = t->is_oopptr()->klass();
760 assert(cik->is_instance_klass() ||
761 cik->is_array_klass(), "Not supported allocation.");
762 if (spobj->stack_allocated()) {
763 Node *box_lock = spobj->in(1);
764 assert(box_lock != NULL, "Need to have a box lock");
765 sv = new StackObjectValue(spobj->_idx,
766 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()),
767 Location::new_stk_loc(Location::oop, C->regalloc()->reg2offset(BoxLockNode::reg(box_lock))),
768 new ConstantIntValue(spobj->n_fields()));
769 set_sv_for_object_node(objs, sv);
770 } else {
771 sv = new ObjectValue(spobj->_idx,
772 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
773 set_sv_for_object_node(objs, sv);
774
775 uint first_ind = spobj->first_index(sfpt->jvms());
776 for (uint i = 0; i < spobj->n_fields(); i++) {
777 Node* fld_node = sfpt->in(first_ind+i);
778 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
779 }
780 }
781 }
782 array->append(sv);
783 return;
784 }
785
786 // Grab the register number for the local
787 OptoReg::Name regnum = C->regalloc()->get_reg_first(local);
788 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
789 // Record the double as two float registers.
790 // The register mask for such a value always specifies two adjacent
791 // float registers, with the lower register number even.
792 // Normally, the allocation of high and low words to these registers
793 // is irrelevant, because nearly all operations on register pairs
794 // (e.g., StoreD) treat them as a single unit.
795 // Here, we assume in addition that the words in these two registers
796 // stored "naturally" (by operations like StoreD and double stores
797 // within the interpreter) such that the lower-numbered register
798 // is written to the lower memory address. This may seem like
799 // a machine dependency, but it is not--it is a requirement on
800 // the author of the <arch>.ad file to ensure that, for every
801 // even/odd double-register pair to which a double may be allocated,
802 // the word in the even single-register is stored to the first
803 // memory word. (Note that register numbers are completely
804 // arbitrary, and are not tied to any machine-level encodings.)
805 #ifdef _LP64
806 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
807 array->append(new ConstantIntValue((jint)0));
808 array->append(new_loc_value( C->regalloc(), regnum, Location::dbl ));
809 } else if ( t->base() == Type::Long ) {
810 array->append(new ConstantIntValue((jint)0));
811 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
812 } else if ( t->base() == Type::RawPtr ) {
813 // jsr/ret return address which must be restored into a the full
814 // width 64-bit stack slot.
815 array->append(new_loc_value( C->regalloc(), regnum, Location::lng ));
816 }
817 #else //_LP64
818 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
819 // Repack the double/long as two jints.
820 // The convention the interpreter uses is that the second local
821 // holds the first raw word of the native double representation.
822 // This is actually reasonable, since locals and stack arrays
823 // grow downwards in all implementations.
824 // (If, on some machine, the interpreter's Java locals or stack
825 // were to grow upwards, the embedded doubles would be word-swapped.)
826 array->append(new_loc_value( C->regalloc(), OptoReg::add(regnum,1), Location::normal ));
827 array->append(new_loc_value( C->regalloc(), regnum , Location::normal ));
828 }
829 #endif //_LP64
830 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
831 OptoReg::is_reg(regnum) ) {
832 array->append(new_loc_value( C->regalloc(), regnum, Matcher::float_in_double()
833 ? Location::float_in_dbl : Location::normal ));
834 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
835 array->append(new_loc_value( C->regalloc(), regnum, Matcher::int_in_long
836 ? Location::int_in_long : Location::normal ));
837 } else if( t->base() == Type::NarrowOop ) {
838 array->append(new_loc_value( C->regalloc(), regnum, Location::narrowoop ));
839 } else {
840 array->append(new_loc_value( C->regalloc(), regnum, C->regalloc()->is_oop(local) ? Location::oop : Location::normal ));
841 }
842 return;
843 }
844
845 // No register. It must be constant data.
846 switch (t->base()) {
847 case Type::Half: // Second half of a double
848 ShouldNotReachHere(); // Caller should skip 2nd halves
849 break;
850 case Type::AnyPtr:
851 array->append(new ConstantOopWriteValue(NULL));
852 break;
853 case Type::AryPtr:
854 case Type::InstPtr: // fall through
855 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
856 break;
857 case Type::NarrowOop:
858 if (t == TypeNarrowOop::NULL_PTR) {
859 array->append(new ConstantOopWriteValue(NULL));
860 } else {
861 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
862 }
863 break;
864 case Type::Int:
865 array->append(new ConstantIntValue(t->is_int()->get_con()));
866 break;
867 case Type::RawPtr:
868 // A return address (T_ADDRESS).
869 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
870 #ifdef _LP64
871 // Must be restored to the full-width 64-bit stack slot.
872 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
873 #else
874 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
875 #endif
876 break;
877 case Type::FloatCon: {
878 float f = t->is_float_constant()->getf();
879 array->append(new ConstantIntValue(jint_cast(f)));
880 break;
881 }
882 case Type::DoubleCon: {
883 jdouble d = t->is_double_constant()->getd();
884 #ifdef _LP64
885 array->append(new ConstantIntValue((jint)0));
886 array->append(new ConstantDoubleValue(d));
887 #else
888 // Repack the double as two jints.
889 // The convention the interpreter uses is that the second local
890 // holds the first raw word of the native double representation.
891 // This is actually reasonable, since locals and stack arrays
892 // grow downwards in all implementations.
893 // (If, on some machine, the interpreter's Java locals or stack
894 // were to grow upwards, the embedded doubles would be word-swapped.)
895 jlong_accessor acc;
896 acc.long_value = jlong_cast(d);
897 array->append(new ConstantIntValue(acc.words[1]));
898 array->append(new ConstantIntValue(acc.words[0]));
899 #endif
900 break;
901 }
902 case Type::Long: {
903 jlong d = t->is_long()->get_con();
904 #ifdef _LP64
905 array->append(new ConstantIntValue((jint)0));
906 array->append(new ConstantLongValue(d));
907 #else
908 // Repack the long as two jints.
909 // The convention the interpreter uses is that the second local
910 // holds the first raw word of the native double representation.
911 // This is actually reasonable, since locals and stack arrays
912 // grow downwards in all implementations.
913 // (If, on some machine, the interpreter's Java locals or stack
914 // were to grow upwards, the embedded doubles would be word-swapped.)
915 jlong_accessor acc;
916 acc.long_value = d;
917 array->append(new ConstantIntValue(acc.words[1]));
918 array->append(new ConstantIntValue(acc.words[0]));
919 #endif
920 break;
921 }
922 case Type::Top: // Add an illegal value here
923 array->append(new LocationValue(Location()));
924 break;
925 default:
926 ShouldNotReachHere();
927 break;
928 }
929 }
930
931 // Determine if this node starts a bundle
932 bool PhaseOutput::starts_bundle(const Node *n) const {
933 return (_node_bundling_limit > n->_idx &&
934 _node_bundling_base[n->_idx].starts_bundle());
935 }
936
937 //--------------------------Process_OopMap_Node--------------------------------
938 void PhaseOutput::Process_OopMap_Node(MachNode *mach, int current_offset) {
939 // Handle special safepoint nodes for synchronization
940 MachSafePointNode *sfn = mach->as_MachSafePoint();
941 MachCallNode *mcall;
942
943 int safepoint_pc_offset = current_offset;
944 bool is_method_handle_invoke = false;
945 bool return_oop = false;
946
947 // Add the safepoint in the DebugInfoRecorder
948 if( !mach->is_MachCall() ) {
949 mcall = NULL;
950 C->debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
951 } else {
952 mcall = mach->as_MachCall();
953
954 // Is the call a MethodHandle call?
955 if (mcall->is_MachCallJava()) {
956 if (mcall->as_MachCallJava()->_method_handle_invoke) {
957 assert(C->has_method_handle_invokes(), "must have been set during call generation");
958 is_method_handle_invoke = true;
959 }
960 }
961
962 // Check if a call returns an object.
963 if (mcall->returns_pointer()) {
964 return_oop = true;
965 }
966 safepoint_pc_offset += mcall->ret_addr_offset();
967 C->debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
968 }
969
970 // Loop over the JVMState list to add scope information
971 // Do not skip safepoints with a NULL method, they need monitor info
972 JVMState* youngest_jvms = sfn->jvms();
973 int max_depth = youngest_jvms->depth();
974
975 // Allocate the object pool for scalar-replaced objects -- the map from
976 // small-integer keys (which can be recorded in the local and ostack
977 // arrays) to descriptions of the object state.
978 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
979
980 // Visit scopes from oldest to youngest.
981 for (int depth = 1; depth <= max_depth; depth++) {
982 JVMState* jvms = youngest_jvms->of_depth(depth);
983 int idx;
984 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
985 // Safepoints that do not have method() set only provide oop-map and monitor info
986 // to support GC; these do not support deoptimization.
987 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
988 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
989 int num_mon = jvms->nof_monitors();
990 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
991 "JVMS local count must match that of the method");
992
993 // Add Local and Expression Stack Information
994
995 // Insert locals into the locarray
996 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
997 for( idx = 0; idx < num_locs; idx++ ) {
998 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
999 }
1000
1001 // Insert expression stack entries into the exparray
1002 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
1003 for( idx = 0; idx < num_exps; idx++ ) {
1004 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
1005 }
1006
1007 // Add in mappings of the monitors
1008 assert( !method ||
1009 !method->is_synchronized() ||
1010 method->is_native() ||
1011 num_mon > 0 ||
1012 !GenerateSynchronizationCode,
1013 "monitors must always exist for synchronized methods");
1014
1015 // Build the growable array of ScopeValues for exp stack
1016 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
1017
1018 // Loop over monitors and insert into array
1019 for (idx = 0; idx < num_mon; idx++) {
1020 // Grab the node that defines this monitor
1021 Node* box_node = sfn->monitor_box(jvms, idx);
1022 Node* obj_node = sfn->monitor_obj(jvms, idx);
1023 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
1024
1025 // Create ScopeValue for object
1026 ScopeValue *scval = NULL;
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 == NULL) {
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 ObjectValue* sv = NULL;
1037 if (spobj->stack_allocated()) {
1038 Node *box_lock = spobj->in(1);
1039 assert(box_lock != NULL, "Need to have a box lock");
1040 assert(eliminated, "monitor has to be eliminated for stack allocation");
1041 sv = new StackObjectValue(spobj->_idx,
1042 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()),
1043 Location::new_stk_loc(Location::oop, C->regalloc()->reg2offset(BoxLockNode::reg(box_lock))),
1044 new ConstantIntValue(spobj->n_fields()));
1045 set_sv_for_object_node(objs, sv);
1046 } else {
1047 sv = new ObjectValue(spobj->_idx,
1048 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
1049 set_sv_for_object_node(objs, sv);
1050
1051 uint first_ind = spobj->first_index(youngest_jvms);
1052 for (uint i = 0; i < spobj->n_fields(); i++) {
1053 Node* fld_node = sfn->in(first_ind+i);
1054 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
1055 }
1056 }
1057 scval = sv;
1058 }
1059 } else if (!obj_node->is_Con()) {
1060 OptoReg::Name obj_reg = C->regalloc()->get_reg_first(obj_node);
1061 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
1062 scval = new_loc_value( C->regalloc(), obj_reg, Location::narrowoop );
1063 } else {
1064 scval = new_loc_value( C->regalloc(), obj_reg, Location::oop );
1065 }
1066 } else {
1067 const TypePtr *tp = obj_node->get_ptr_type();
1068 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
1069 }
1070
1071 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
1072 Location basic_lock = Location::new_stk_loc(Location::normal,C->regalloc()->reg2offset(box_reg));
1073 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
1074 }
1075
1076 for (idx = 0; idx < jvms->scl_size(); idx++ ) {
1077 Node* obj_node = sfn->scalar(jvms, idx);
1078
1079 if (obj_node->is_SafePointScalarObject()) {
1080 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
1081 if (sv_for_node_id(objs, spobj->_idx) == NULL) {
1082 const Type *t = spobj->bottom_type();
1083 ciKlass* cik = t->is_oopptr()->klass();
1084 assert(cik->is_instance_klass() ||
1085 cik->is_array_klass(), "Not supported allocation.");
1086 assert(spobj->stack_allocated(), "has to be stack allocated");
1087 Node *box_lock = spobj->in(1);
1088 assert(box_lock != NULL, "Need to have a box lock");
1089 StackObjectValue* sv = new StackObjectValue(spobj->_idx,
1090 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()),
1091 Location::new_stk_loc(Location::oop, C->regalloc()->reg2offset(BoxLockNode::reg(box_lock))),
1092 new ConstantIntValue(spobj->n_fields()));
1093 set_sv_for_object_node(objs, sv);
1094 }
1095 }
1096 }
1097 // We dump the object pool first, since deoptimization reads it in first.
1098 C->debug_info()->dump_object_pool(objs);
1099
1100 // Build first class objects to pass to scope
1101 DebugToken *locvals = C->debug_info()->create_scope_values(locarray);
1102 DebugToken *expvals = C->debug_info()->create_scope_values(exparray);
1103 DebugToken *monvals = C->debug_info()->create_monitor_values(monarray);
1104
1105 // Make method available for all Safepoints
1106 ciMethod* scope_method = method ? method : C->method();
1107 // Describe the scope here
1108 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1109 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1110 // Now we can describe the scope.
1111 methodHandle null_mh;
1112 bool rethrow_exception = false;
1113 C->debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, locvals, expvals, monvals);
1114 } // End jvms loop
1115
1116 // Mark the end of the scope set.
1117 C->debug_info()->end_safepoint(safepoint_pc_offset);
1118 }
1119
1120
1121
1122 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1123 class NonSafepointEmitter {
1124 Compile* C;
1125 JVMState* _pending_jvms;
1126 int _pending_offset;
1127
1128 void emit_non_safepoint();
1129
1130 public:
1131 NonSafepointEmitter(Compile* compile) {
1132 this->C = compile;
1133 _pending_jvms = NULL;
1134 _pending_offset = 0;
1135 }
1136
1137 void observe_instruction(Node* n, int pc_offset) {
1138 if (!C->debug_info()->recording_non_safepoints()) return;
1139
1140 Node_Notes* nn = C->node_notes_at(n->_idx);
1141 if (nn == NULL || nn->jvms() == NULL) return;
1142 if (_pending_jvms != NULL &&
1143 _pending_jvms->same_calls_as(nn->jvms())) {
1144 // Repeated JVMS? Stretch it up here.
1145 _pending_offset = pc_offset;
1146 } else {
1147 if (_pending_jvms != NULL &&
1148 _pending_offset < pc_offset) {
1149 emit_non_safepoint();
1150 }
1151 _pending_jvms = NULL;
1152 if (pc_offset > C->debug_info()->last_pc_offset()) {
1153 // This is the only way _pending_jvms can become non-NULL:
1154 _pending_jvms = nn->jvms();
1155 _pending_offset = pc_offset;
1156 }
1157 }
1158 }
1159
1160 // Stay out of the way of real safepoints:
1161 void observe_safepoint(JVMState* jvms, int pc_offset) {
1162 if (_pending_jvms != NULL &&
1163 !_pending_jvms->same_calls_as(jvms) &&
1164 _pending_offset < pc_offset) {
1165 emit_non_safepoint();
1166 }
1167 _pending_jvms = NULL;
1168 }
1169
1170 void flush_at_end() {
1171 if (_pending_jvms != NULL) {
1172 emit_non_safepoint();
1173 }
1174 _pending_jvms = NULL;
1175 }
1176 };
1177
1178 void NonSafepointEmitter::emit_non_safepoint() {
1179 JVMState* youngest_jvms = _pending_jvms;
1180 int pc_offset = _pending_offset;
1181
1182 // Clear it now:
1183 _pending_jvms = NULL;
1184
1185 DebugInformationRecorder* debug_info = C->debug_info();
1186 assert(debug_info->recording_non_safepoints(), "sanity");
1187
1188 debug_info->add_non_safepoint(pc_offset);
1189 int max_depth = youngest_jvms->depth();
1190
1191 // Visit scopes from oldest to youngest.
1192 for (int depth = 1; depth <= max_depth; depth++) {
1193 JVMState* jvms = youngest_jvms->of_depth(depth);
1194 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1195 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1196 methodHandle null_mh;
1197 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
1198 }
1199
1200 // Mark the end of the scope set.
1201 debug_info->end_non_safepoint(pc_offset);
1202 }
1203
1204 //------------------------------init_buffer------------------------------------
1205 void PhaseOutput::estimate_buffer_size(int& const_req) {
1206
1207 // Set the initially allocated size
1208 const_req = initial_const_capacity;
1209
1210 // The extra spacing after the code is necessary on some platforms.
1211 // Sometimes we need to patch in a jump after the last instruction,
1212 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1213
1214 // Compute the byte offset where we can store the deopt pc.
1215 if (C->fixed_slots() != 0) {
1216 _orig_pc_slot_offset_in_bytes = C->regalloc()->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1217 }
1218
1219 // Compute prolog code size
1220 _method_size = 0;
1221 _frame_slots = OptoReg::reg2stack(C->matcher()->_old_SP) + C->regalloc()->_framesize;
1222 #if defined(IA64) && !defined(AIX)
1223 if (save_argument_registers()) {
1224 // 4815101: this is a stub with implicit and unknown precision fp args.
1225 // The usual spill mechanism can only generate stfd's in this case, which
1226 // doesn't work if the fp reg to spill contains a single-precision denorm.
1227 // Instead, we hack around the normal spill mechanism using stfspill's and
1228 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1229 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1230 //
1231 // If we ever implement 16-byte 'registers' == stack slots, we can
1232 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1233 // instead of stfd/stfs/ldfd/ldfs.
1234 _frame_slots += 8*(16/BytesPerInt);
1235 }
1236 #endif
1237 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1238
1239 if (C->has_mach_constant_base_node()) {
1240 uint add_size = 0;
1241 // Fill the constant table.
1242 // Note: This must happen before shorten_branches.
1243 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1244 Block* b = C->cfg()->get_block(i);
1245
1246 for (uint j = 0; j < b->number_of_nodes(); j++) {
1247 Node* n = b->get_node(j);
1248
1249 // If the node is a MachConstantNode evaluate the constant
1250 // value section.
1251 if (n->is_MachConstant()) {
1252 MachConstantNode* machcon = n->as_MachConstant();
1253 machcon->eval_constant(C);
1254 } else if (n->is_Mach()) {
1255 // On Power there are more nodes that issue constants.
1256 add_size += (n->as_Mach()->ins_num_consts() * 8);
1257 }
1258 }
1259 }
1260
1261 // Calculate the offsets of the constants and the size of the
1262 // constant table (including the padding to the next section).
1263 constant_table().calculate_offsets_and_size();
1264 const_req = constant_table().size() + add_size;
1265 }
1266
1267 // Initialize the space for the BufferBlob used to find and verify
1268 // instruction size in MachNode::emit_size()
1269 init_scratch_buffer_blob(const_req);
1270 }
1271
1272 CodeBuffer* PhaseOutput::init_buffer() {
1273 int stub_req = _buf_sizes._stub;
1274 int code_req = _buf_sizes._code;
1275 int const_req = _buf_sizes._const;
1276
1277 int pad_req = NativeCall::instruction_size;
1278
1279 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1280 stub_req += bs->estimate_stub_size();
1281
1282 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1283 // class HandlerImpl is platform-specific and defined in the *.ad files.
1284 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1285 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1286 stub_req += MAX_stubs_size; // ensure per-stub margin
1287 code_req += MAX_inst_size; // ensure per-instruction margin
1288
1289 if (StressCodeBuffers)
1290 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1291
1292 int total_req =
1293 const_req +
1294 code_req +
1295 pad_req +
1296 stub_req +
1297 exception_handler_req +
1298 deopt_handler_req; // deopt handler
1299
1300 if (C->has_method_handle_invokes())
1301 total_req += deopt_handler_req; // deopt MH handler
1302
1303 CodeBuffer* cb = code_buffer();
1304 cb->initialize(total_req, _buf_sizes._reloc);
1305
1306 // Have we run out of code space?
1307 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1308 C->record_failure("CodeCache is full");
1309 return NULL;
1310 }
1311 // Configure the code buffer.
1312 cb->initialize_consts_size(const_req);
1313 cb->initialize_stubs_size(stub_req);
1314 cb->initialize_oop_recorder(C->env()->oop_recorder());
1315
1316 // fill in the nop array for bundling computations
1317 MachNode *_nop_list[Bundle::_nop_count];
1318 Bundle::initialize_nops(_nop_list);
1319
1320 // if we are using stack allocation enable the runtime part
1321 // stack allocation can be enabled selectively via compiler directive
1322 // so we need to enable the runtime part
1323 if (!UseStackAllocationRuntime && C->do_stack_allocation()) {
1324 FLAG_SET_ERGO(UseStackAllocationRuntime, true);
1325 }
1326
1327 return cb;
1328 }
1329
1330 //------------------------------fill_buffer------------------------------------
1331 void PhaseOutput::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1332 // blk_starts[] contains offsets calculated during short branches processing,
1333 // offsets should not be increased during following steps.
1334
1335 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1336 // of a loop. It is used to determine the padding for loop alignment.
1337 compute_loop_first_inst_sizes();
1338
1339 // Create oopmap set.
1340 _oop_map_set = new OopMapSet();
1341
1342 // !!!!! This preserves old handling of oopmaps for now
1343 C->debug_info()->set_oopmaps(_oop_map_set);
1344
1345 uint nblocks = C->cfg()->number_of_blocks();
1346 // Count and start of implicit null check instructions
1347 uint inct_cnt = 0;
1348 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1349
1350 // Count and start of calls
1351 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1352
1353 uint return_offset = 0;
1354 int nop_size = (new MachNopNode())->size(C->regalloc());
1355
1356 int previous_offset = 0;
1357 int current_offset = 0;
1358 int last_call_offset = -1;
1359 int last_avoid_back_to_back_offset = -1;
1360 #ifdef ASSERT
1361 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1362 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1363 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1364 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1365 #endif
1366
1367 // Create an array of unused labels, one for each basic block, if printing is enabled
1368 #if defined(SUPPORT_OPTO_ASSEMBLY)
1369 int *node_offsets = NULL;
1370 uint node_offset_limit = C->unique();
1371
1372 if (C->print_assembly()) {
1373 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1374 }
1375 if (node_offsets != NULL) {
1376 // We need to initialize. Unused array elements may contain garbage and mess up PrintOptoAssembly.
1377 memset(node_offsets, 0, node_offset_limit*sizeof(int));
1378 }
1379 #endif
1380
1381 NonSafepointEmitter non_safepoints(C); // emit non-safepoints lazily
1382
1383 // Emit the constant table.
1384 if (C->has_mach_constant_base_node()) {
1385 constant_table().emit(*cb);
1386 }
1387
1388 // Create an array of labels, one for each basic block
1389 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1390 for (uint i=0; i <= nblocks; i++) {
1391 blk_labels[i].init();
1392 }
1393
1394 // ------------------
1395 // Now fill in the code buffer
1396 Node *delay_slot = NULL;
1397
1398 for (uint i = 0; i < nblocks; i++) {
1399 Block* block = C->cfg()->get_block(i);
1400 _block = block;
1401 Node* head = block->head();
1402
1403 // If this block needs to start aligned (i.e, can be reached other
1404 // than by falling-thru from the previous block), then force the
1405 // start of a new bundle.
1406 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1407 cb->flush_bundle(true);
1408 }
1409
1410 #ifdef ASSERT
1411 if (!block->is_connector()) {
1412 stringStream st;
1413 block->dump_head(C->cfg(), &st);
1414 MacroAssembler(cb).block_comment(st.as_string());
1415 }
1416 jmp_target[i] = 0;
1417 jmp_offset[i] = 0;
1418 jmp_size[i] = 0;
1419 jmp_rule[i] = 0;
1420 #endif
1421 int blk_offset = current_offset;
1422
1423 // Define the label at the beginning of the basic block
1424 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1425
1426 uint last_inst = block->number_of_nodes();
1427
1428 // Emit block normally, except for last instruction.
1429 // Emit means "dump code bits into code buffer".
1430 for (uint j = 0; j<last_inst; j++) {
1431 _index = j;
1432
1433 // Get the node
1434 Node* n = block->get_node(j);
1435
1436 // See if delay slots are supported
1437 if (valid_bundle_info(n) && node_bundling(n)->used_in_unconditional_delay()) {
1438 assert(delay_slot == NULL, "no use of delay slot node");
1439 assert(n->size(C->regalloc()) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1440
1441 delay_slot = n;
1442 continue;
1443 }
1444
1445 // If this starts a new instruction group, then flush the current one
1446 // (but allow split bundles)
1447 if (Pipeline::requires_bundling() && starts_bundle(n))
1448 cb->flush_bundle(false);
1449
1450 // Special handling for SafePoint/Call Nodes
1451 bool is_mcall = false;
1452 if (n->is_Mach()) {
1453 MachNode *mach = n->as_Mach();
1454 is_mcall = n->is_MachCall();
1455 bool is_sfn = n->is_MachSafePoint();
1456
1457 // If this requires all previous instructions be flushed, then do so
1458 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1459 cb->flush_bundle(true);
1460 current_offset = cb->insts_size();
1461 }
1462
1463 // A padding may be needed again since a previous instruction
1464 // could be moved to delay slot.
1465
1466 // align the instruction if necessary
1467 int padding = mach->compute_padding(current_offset);
1468 // Make sure safepoint node for polling is distinct from a call's
1469 // return by adding a nop if needed.
1470 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1471 padding = nop_size;
1472 }
1473 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1474 current_offset == last_avoid_back_to_back_offset) {
1475 // Avoid back to back some instructions.
1476 padding = nop_size;
1477 }
1478
1479 if (padding > 0) {
1480 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1481 int nops_cnt = padding / nop_size;
1482 MachNode *nop = new MachNopNode(nops_cnt);
1483 block->insert_node(nop, j++);
1484 last_inst++;
1485 C->cfg()->map_node_to_block(nop, block);
1486 // Ensure enough space.
1487 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1488 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1489 C->record_failure("CodeCache is full");
1490 return;
1491 }
1492 nop->emit(*cb, C->regalloc());
1493 cb->flush_bundle(true);
1494 current_offset = cb->insts_size();
1495 }
1496
1497 // Remember the start of the last call in a basic block
1498 if (is_mcall) {
1499 MachCallNode *mcall = mach->as_MachCall();
1500
1501 // This destination address is NOT PC-relative
1502 mcall->method_set((intptr_t)mcall->entry_point());
1503
1504 // Save the return address
1505 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1506
1507 if (mcall->is_MachCallLeaf()) {
1508 is_mcall = false;
1509 is_sfn = false;
1510 }
1511 }
1512
1513 // sfn will be valid whenever mcall is valid now because of inheritance
1514 if (is_sfn || is_mcall) {
1515
1516 // Handle special safepoint nodes for synchronization
1517 if (!is_mcall) {
1518 MachSafePointNode *sfn = mach->as_MachSafePoint();
1519 // !!!!! Stubs only need an oopmap right now, so bail out
1520 if (sfn->jvms()->method() == NULL) {
1521 // Write the oopmap directly to the code blob??!!
1522 continue;
1523 }
1524 } // End synchronization
1525
1526 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1527 current_offset);
1528 Process_OopMap_Node(mach, current_offset);
1529 } // End if safepoint
1530
1531 // If this is a null check, then add the start of the previous instruction to the list
1532 else if( mach->is_MachNullCheck() ) {
1533 inct_starts[inct_cnt++] = previous_offset;
1534 }
1535
1536 // If this is a branch, then fill in the label with the target BB's label
1537 else if (mach->is_MachBranch()) {
1538 // This requires the TRUE branch target be in succs[0]
1539 uint block_num = block->non_connector_successor(0)->_pre_order;
1540
1541 // Try to replace long branch if delay slot is not used,
1542 // it is mostly for back branches since forward branch's
1543 // distance is not updated yet.
1544 bool delay_slot_is_used = valid_bundle_info(n) &&
1545 C->output()->node_bundling(n)->use_unconditional_delay();
1546 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1547 assert(delay_slot == NULL, "not expecting delay slot node");
1548 int br_size = n->size(C->regalloc());
1549 int offset = blk_starts[block_num] - current_offset;
1550 if (block_num >= i) {
1551 // Current and following block's offset are not
1552 // finalized yet, adjust distance by the difference
1553 // between calculated and final offsets of current block.
1554 offset -= (blk_starts[i] - blk_offset);
1555 }
1556 // In the following code a nop could be inserted before
1557 // the branch which will increase the backward distance.
1558 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1559 if (needs_padding && offset <= 0)
1560 offset -= nop_size;
1561
1562 if (C->matcher()->is_short_branch_offset(mach->rule(), br_size, offset)) {
1563 // We've got a winner. Replace this branch.
1564 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1565
1566 // Update the jmp_size.
1567 int new_size = replacement->size(C->regalloc());
1568 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1569 // Insert padding between avoid_back_to_back branches.
1570 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1571 MachNode *nop = new MachNopNode();
1572 block->insert_node(nop, j++);
1573 C->cfg()->map_node_to_block(nop, block);
1574 last_inst++;
1575 nop->emit(*cb, C->regalloc());
1576 cb->flush_bundle(true);
1577 current_offset = cb->insts_size();
1578 }
1579 #ifdef ASSERT
1580 jmp_target[i] = block_num;
1581 jmp_offset[i] = current_offset - blk_offset;
1582 jmp_size[i] = new_size;
1583 jmp_rule[i] = mach->rule();
1584 #endif
1585 block->map_node(replacement, j);
1586 mach->subsume_by(replacement, C);
1587 n = replacement;
1588 mach = replacement;
1589 }
1590 }
1591 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1592 } else if (mach->ideal_Opcode() == Op_Jump) {
1593 for (uint h = 0; h < block->_num_succs; h++) {
1594 Block* succs_block = block->_succs[h];
1595 for (uint j = 1; j < succs_block->num_preds(); j++) {
1596 Node* jpn = succs_block->pred(j);
1597 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1598 uint block_num = succs_block->non_connector()->_pre_order;
1599 Label *blkLabel = &blk_labels[block_num];
1600 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1601 }
1602 }
1603 }
1604 }
1605 #ifdef ASSERT
1606 // Check that oop-store precedes the card-mark
1607 else if (mach->ideal_Opcode() == Op_StoreCM) {
1608 uint storeCM_idx = j;
1609 int count = 0;
1610 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1611 Node *oop_store = mach->in(prec); // Precedence edge
1612 if (oop_store == NULL) continue;
1613 count++;
1614 uint i4;
1615 for (i4 = 0; i4 < last_inst; ++i4) {
1616 if (block->get_node(i4) == oop_store) {
1617 break;
1618 }
1619 }
1620 // Note: This test can provide a false failure if other precedence
1621 // edges have been added to the storeCMNode.
1622 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1623 }
1624 assert(count > 0, "storeCM expects at least one precedence edge");
1625 }
1626 #endif
1627 else if (!n->is_Proj()) {
1628 // Remember the beginning of the previous instruction, in case
1629 // it's followed by a flag-kill and a null-check. Happens on
1630 // Intel all the time, with add-to-memory kind of opcodes.
1631 previous_offset = current_offset;
1632 }
1633
1634 // Not an else-if!
1635 // If this is a trap based cmp then add its offset to the list.
1636 if (mach->is_TrapBasedCheckNode()) {
1637 inct_starts[inct_cnt++] = current_offset;
1638 }
1639 }
1640
1641 // Verify that there is sufficient space remaining
1642 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1643 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1644 C->record_failure("CodeCache is full");
1645 return;
1646 }
1647
1648 // Save the offset for the listing
1649 #if defined(SUPPORT_OPTO_ASSEMBLY)
1650 if ((node_offsets != NULL) && (n->_idx < node_offset_limit)) {
1651 node_offsets[n->_idx] = cb->insts_size();
1652 }
1653 #endif
1654
1655 // "Normal" instruction case
1656 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1657 n->emit(*cb, C->regalloc());
1658 current_offset = cb->insts_size();
1659
1660 // Above we only verified that there is enough space in the instruction section.
1661 // However, the instruction may emit stubs that cause code buffer expansion.
1662 // Bail out here if expansion failed due to a lack of code cache space.
1663 if (C->failing()) {
1664 return;
1665 }
1666
1667 #ifdef ASSERT
1668 uint n_size = n->size(C->regalloc());
1669 if (n_size < (current_offset-instr_offset)) {
1670 MachNode* mach = n->as_Mach();
1671 n->dump();
1672 mach->dump_format(C->regalloc(), tty);
1673 tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
1674 Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty);
1675 tty->print_cr(" ------------------- ");
1676 BufferBlob* blob = this->scratch_buffer_blob();
1677 address blob_begin = blob->content_begin();
1678 Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
1679 assert(false, "wrong size of mach node");
1680 }
1681 #endif
1682 non_safepoints.observe_instruction(n, current_offset);
1683
1684 // mcall is last "call" that can be a safepoint
1685 // record it so we can see if a poll will directly follow it
1686 // in which case we'll need a pad to make the PcDesc sites unique
1687 // see 5010568. This can be slightly inaccurate but conservative
1688 // in the case that return address is not actually at current_offset.
1689 // This is a small price to pay.
1690
1691 if (is_mcall) {
1692 last_call_offset = current_offset;
1693 }
1694
1695 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1696 // Avoid back to back some instructions.
1697 last_avoid_back_to_back_offset = current_offset;
1698 }
1699
1700 // See if this instruction has a delay slot
1701 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1702 guarantee(delay_slot != NULL, "expecting delay slot node");
1703
1704 // Back up 1 instruction
1705 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1706
1707 // Save the offset for the listing
1708 #if defined(SUPPORT_OPTO_ASSEMBLY)
1709 if ((node_offsets != NULL) && (delay_slot->_idx < node_offset_limit)) {
1710 node_offsets[delay_slot->_idx] = cb->insts_size();
1711 }
1712 #endif
1713
1714 // Support a SafePoint in the delay slot
1715 if (delay_slot->is_MachSafePoint()) {
1716 MachNode *mach = delay_slot->as_Mach();
1717 // !!!!! Stubs only need an oopmap right now, so bail out
1718 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1719 // Write the oopmap directly to the code blob??!!
1720 delay_slot = NULL;
1721 continue;
1722 }
1723
1724 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1725 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1726 adjusted_offset);
1727 // Generate an OopMap entry
1728 Process_OopMap_Node(mach, adjusted_offset);
1729 }
1730
1731 // Insert the delay slot instruction
1732 delay_slot->emit(*cb, C->regalloc());
1733
1734 // Don't reuse it
1735 delay_slot = NULL;
1736 }
1737
1738 } // End for all instructions in block
1739
1740 // If the next block is the top of a loop, pad this block out to align
1741 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1742 if (i < nblocks-1) {
1743 Block *nb = C->cfg()->get_block(i + 1);
1744 int padding = nb->alignment_padding(current_offset);
1745 if( padding > 0 ) {
1746 MachNode *nop = new MachNopNode(padding / nop_size);
1747 block->insert_node(nop, block->number_of_nodes());
1748 C->cfg()->map_node_to_block(nop, block);
1749 nop->emit(*cb, C->regalloc());
1750 current_offset = cb->insts_size();
1751 }
1752 }
1753 // Verify that the distance for generated before forward
1754 // short branches is still valid.
1755 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1756
1757 // Save new block start offset
1758 blk_starts[i] = blk_offset;
1759 } // End of for all blocks
1760 blk_starts[nblocks] = current_offset;
1761
1762 non_safepoints.flush_at_end();
1763
1764 // Offset too large?
1765 if (C->failing()) return;
1766
1767 // Define a pseudo-label at the end of the code
1768 MacroAssembler(cb).bind( blk_labels[nblocks] );
1769
1770 // Compute the size of the first block
1771 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1772
1773 #ifdef ASSERT
1774 for (uint i = 0; i < nblocks; i++) { // For all blocks
1775 if (jmp_target[i] != 0) {
1776 int br_size = jmp_size[i];
1777 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1778 if (!C->matcher()->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1779 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]);
1780 assert(false, "Displacement too large for short jmp");
1781 }
1782 }
1783 }
1784 #endif
1785
1786 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1787 bs->emit_stubs(*cb);
1788 if (C->failing()) return;
1789
1790 #ifndef PRODUCT
1791 // Information on the size of the method, without the extraneous code
1792 Scheduling::increment_method_size(cb->insts_size());
1793 #endif
1794
1795 // ------------------
1796 // Fill in exception table entries.
1797 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1798
1799 // Only java methods have exception handlers and deopt handlers
1800 // class HandlerImpl is platform-specific and defined in the *.ad files.
1801 if (C->method()) {
1802 // Emit the exception handler code.
1803 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1804 if (C->failing()) {
1805 return; // CodeBuffer::expand failed
1806 }
1807 // Emit the deopt handler code.
1808 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1809
1810 // Emit the MethodHandle deopt handler code (if required).
1811 if (C->has_method_handle_invokes() && !C->failing()) {
1812 // We can use the same code as for the normal deopt handler, we
1813 // just need a different entry point address.
1814 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1815 }
1816 }
1817
1818 // One last check for failed CodeBuffer::expand:
1819 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1820 C->record_failure("CodeCache is full");
1821 return;
1822 }
1823
1824 #if defined(SUPPORT_ABSTRACT_ASSEMBLY) || defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_OPTO_ASSEMBLY)
1825 if (C->print_assembly()) {
1826 tty->cr();
1827 tty->print_cr("============================= C2-compiled nmethod ==============================");
1828 }
1829 #endif
1830
1831 #if defined(SUPPORT_OPTO_ASSEMBLY)
1832 // Dump the assembly code, including basic-block numbers
1833 if (C->print_assembly()) {
1834 ttyLocker ttyl; // keep the following output all in one block
1835 if (!VMThread::should_terminate()) { // test this under the tty lock
1836 // This output goes directly to the tty, not the compiler log.
1837 // To enable tools to match it up with the compilation activity,
1838 // be sure to tag this tty output with the compile ID.
1839 if (xtty != NULL) {
1840 xtty->head("opto_assembly compile_id='%d'%s", C->compile_id(),
1841 C->is_osr_compilation() ? " compile_kind='osr'" :
1842 "");
1843 }
1844 if (C->method() != NULL) {
1845 tty->print_cr("----------------------- MetaData before Compile_id = %d ------------------------", C->compile_id());
1846 C->method()->print_metadata();
1847 } else if (C->stub_name() != NULL) {
1848 tty->print_cr("----------------------------- RuntimeStub %s -------------------------------", C->stub_name());
1849 }
1850 tty->cr();
1851 tty->print_cr("------------------------ OptoAssembly for Compile_id = %d -----------------------", C->compile_id());
1852 dump_asm(node_offsets, node_offset_limit);
1853 tty->print_cr("--------------------------------------------------------------------------------");
1854 if (xtty != NULL) {
1855 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock.
1856 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done
1857 // thread safe
1858 ttyLocker ttyl2;
1859 xtty->tail("opto_assembly");
1860 }
1861 }
1862 }
1863 #endif
1864 }
1865
1866 void PhaseOutput::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1867 _inc_table.set_size(cnt);
1868
1869 uint inct_cnt = 0;
1870 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
1871 Block* block = C->cfg()->get_block(i);
1872 Node *n = NULL;
1873 int j;
1874
1875 // Find the branch; ignore trailing NOPs.
1876 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1877 n = block->get_node(j);
1878 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1879 break;
1880 }
1881 }
1882
1883 // If we didn't find anything, continue
1884 if (j < 0) {
1885 continue;
1886 }
1887
1888 // Compute ExceptionHandlerTable subtable entry and add it
1889 // (skip empty blocks)
1890 if (n->is_Catch()) {
1891
1892 // Get the offset of the return from the call
1893 uint call_return = call_returns[block->_pre_order];
1894 #ifdef ASSERT
1895 assert( call_return > 0, "no call seen for this basic block" );
1896 while (block->get_node(--j)->is_MachProj()) ;
1897 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1898 #endif
1899 // last instruction is a CatchNode, find it's CatchProjNodes
1900 int nof_succs = block->_num_succs;
1901 // allocate space
1902 GrowableArray<intptr_t> handler_bcis(nof_succs);
1903 GrowableArray<intptr_t> handler_pcos(nof_succs);
1904 // iterate through all successors
1905 for (int j = 0; j < nof_succs; j++) {
1906 Block* s = block->_succs[j];
1907 bool found_p = false;
1908 for (uint k = 1; k < s->num_preds(); k++) {
1909 Node* pk = s->pred(k);
1910 if (pk->is_CatchProj() && pk->in(0) == n) {
1911 const CatchProjNode* p = pk->as_CatchProj();
1912 found_p = true;
1913 // add the corresponding handler bci & pco information
1914 if (p->_con != CatchProjNode::fall_through_index) {
1915 // p leads to an exception handler (and is not fall through)
1916 assert(s == C->cfg()->get_block(s->_pre_order), "bad numbering");
1917 // no duplicates, please
1918 if (!handler_bcis.contains(p->handler_bci())) {
1919 uint block_num = s->non_connector()->_pre_order;
1920 handler_bcis.append(p->handler_bci());
1921 handler_pcos.append(blk_labels[block_num].loc_pos());
1922 }
1923 }
1924 }
1925 }
1926 assert(found_p, "no matching predecessor found");
1927 // Note: Due to empty block removal, one block may have
1928 // several CatchProj inputs, from the same Catch.
1929 }
1930
1931 // Set the offset of the return from the call
1932 assert(handler_bcis.find(-1) != -1, "must have default handler");
1933 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1934 continue;
1935 }
1936
1937 // Handle implicit null exception table updates
1938 if (n->is_MachNullCheck()) {
1939 uint block_num = block->non_connector_successor(0)->_pre_order;
1940 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1941 continue;
1942 }
1943 // Handle implicit exception table updates: trap instructions.
1944 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1945 uint block_num = block->non_connector_successor(0)->_pre_order;
1946 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1947 continue;
1948 }
1949 } // End of for all blocks fill in exception table entries
1950 }
1951
1952 // Static Variables
1953 #ifndef PRODUCT
1954 uint Scheduling::_total_nop_size = 0;
1955 uint Scheduling::_total_method_size = 0;
1956 uint Scheduling::_total_branches = 0;
1957 uint Scheduling::_total_unconditional_delays = 0;
1958 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1959 #endif
1960
1961 // Initializer for class Scheduling
1962
1963 Scheduling::Scheduling(Arena *arena, Compile &compile)
1964 : _arena(arena),
1965 _cfg(compile.cfg()),
1966 _regalloc(compile.regalloc()),
1967 _scheduled(arena),
1968 _available(arena),
1969 _reg_node(arena),
1970 _pinch_free_list(arena),
1971 _next_node(NULL),
1972 _bundle_instr_count(0),
1973 _bundle_cycle_number(0),
1974 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
1975 #ifndef PRODUCT
1976 , _branches(0)
1977 , _unconditional_delays(0)
1978 #endif
1979 {
1980 // Create a MachNopNode
1981 _nop = new MachNopNode();
1982
1983 // Now that the nops are in the array, save the count
1984 // (but allow entries for the nops)
1985 _node_bundling_limit = compile.unique();
1986 uint node_max = _regalloc->node_regs_max_index();
1987
1988 compile.output()->set_node_bundling_limit(_node_bundling_limit);
1989
1990 // This one is persistent within the Compile class
1991 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1992
1993 // Allocate space for fixed-size arrays
1994 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1995 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1996 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1997
1998 // Clear the arrays
1999 for (uint i = 0; i < node_max; i++) {
2000 ::new (&_node_bundling_base[i]) Bundle();
2001 }
2002 memset(_node_latency, 0, node_max * sizeof(unsigned short));
2003 memset(_uses, 0, node_max * sizeof(short));
2004 memset(_current_latency, 0, node_max * sizeof(unsigned short));
2005
2006 // Clear the bundling information
2007 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
2008
2009 // Get the last node
2010 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
2011
2012 _next_node = block->get_node(block->number_of_nodes() - 1);
2013 }
2014
2015 #ifndef PRODUCT
2016 // Scheduling destructor
2017 Scheduling::~Scheduling() {
2018 _total_branches += _branches;
2019 _total_unconditional_delays += _unconditional_delays;
2020 }
2021 #endif
2022
2023 // Step ahead "i" cycles
2024 void Scheduling::step(uint i) {
2025
2026 Bundle *bundle = node_bundling(_next_node);
2027 bundle->set_starts_bundle();
2028
2029 // Update the bundle record, but leave the flags information alone
2030 if (_bundle_instr_count > 0) {
2031 bundle->set_instr_count(_bundle_instr_count);
2032 bundle->set_resources_used(_bundle_use.resourcesUsed());
2033 }
2034
2035 // Update the state information
2036 _bundle_instr_count = 0;
2037 _bundle_cycle_number += i;
2038 _bundle_use.step(i);
2039 }
2040
2041 void Scheduling::step_and_clear() {
2042 Bundle *bundle = node_bundling(_next_node);
2043 bundle->set_starts_bundle();
2044
2045 // Update the bundle record
2046 if (_bundle_instr_count > 0) {
2047 bundle->set_instr_count(_bundle_instr_count);
2048 bundle->set_resources_used(_bundle_use.resourcesUsed());
2049
2050 _bundle_cycle_number += 1;
2051 }
2052
2053 // Clear the bundling information
2054 _bundle_instr_count = 0;
2055 _bundle_use.reset();
2056
2057 memcpy(_bundle_use_elements,
2058 Pipeline_Use::elaborated_elements,
2059 sizeof(Pipeline_Use::elaborated_elements));
2060 }
2061
2062 // Perform instruction scheduling and bundling over the sequence of
2063 // instructions in backwards order.
2064 void PhaseOutput::ScheduleAndBundle() {
2065
2066 // Don't optimize this if it isn't a method
2067 if (!C->method())
2068 return;
2069
2070 // Don't optimize this if scheduling is disabled
2071 if (!C->do_scheduling())
2072 return;
2073
2074 // Scheduling code works only with pairs (16 bytes) maximum.
2075 if (C->max_vector_size() > 16)
2076 return;
2077
2078 Compile::TracePhase tp("isched", &timers[_t_instrSched]);
2079
2080 // Create a data structure for all the scheduling information
2081 Scheduling scheduling(Thread::current()->resource_area(), *C);
2082
2083 // Walk backwards over each basic block, computing the needed alignment
2084 // Walk over all the basic blocks
2085 scheduling.DoScheduling();
2086
2087 #ifndef PRODUCT
2088 if (C->trace_opto_output()) {
2089 tty->print("\n---- After ScheduleAndBundle ----\n");
2090 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
2091 tty->print("\nBB#%03d:\n", i);
2092 Block* block = C->cfg()->get_block(i);
2093 for (uint j = 0; j < block->number_of_nodes(); j++) {
2094 Node* n = block->get_node(j);
2095 OptoReg::Name reg = C->regalloc()->get_reg_first(n);
2096 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
2097 n->dump();
2098 }
2099 }
2100 }
2101 #endif
2102 }
2103
2104 // Compute the latency of all the instructions. This is fairly simple,
2105 // because we already have a legal ordering. Walk over the instructions
2106 // from first to last, and compute the latency of the instruction based
2107 // on the latency of the preceding instruction(s).
2108 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
2109 #ifndef PRODUCT
2110 if (_cfg->C->trace_opto_output())
2111 tty->print("# -> ComputeLocalLatenciesForward\n");
2112 #endif
2113
2114 // Walk over all the schedulable instructions
2115 for( uint j=_bb_start; j < _bb_end; j++ ) {
2116
2117 // This is a kludge, forcing all latency calculations to start at 1.
2118 // Used to allow latency 0 to force an instruction to the beginning
2119 // of the bb
2120 uint latency = 1;
2121 Node *use = bb->get_node(j);
2122 uint nlen = use->len();
2123
2124 // Walk over all the inputs
2125 for ( uint k=0; k < nlen; k++ ) {
2126 Node *def = use->in(k);
2127 if (!def)
2128 continue;
2129
2130 uint l = _node_latency[def->_idx] + use->latency(k);
2131 if (latency < l)
2132 latency = l;
2133 }
2134
2135 _node_latency[use->_idx] = latency;
2136
2137 #ifndef PRODUCT
2138 if (_cfg->C->trace_opto_output()) {
2139 tty->print("# latency %4d: ", latency);
2140 use->dump();
2141 }
2142 #endif
2143 }
2144
2145 #ifndef PRODUCT
2146 if (_cfg->C->trace_opto_output())
2147 tty->print("# <- ComputeLocalLatenciesForward\n");
2148 #endif
2149
2150 } // end ComputeLocalLatenciesForward
2151
2152 // See if this node fits into the present instruction bundle
2153 bool Scheduling::NodeFitsInBundle(Node *n) {
2154 uint n_idx = n->_idx;
2155
2156 // If this is the unconditional delay instruction, then it fits
2157 if (n == _unconditional_delay_slot) {
2158 #ifndef PRODUCT
2159 if (_cfg->C->trace_opto_output())
2160 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2161 #endif
2162 return (true);
2163 }
2164
2165 // If the node cannot be scheduled this cycle, skip it
2166 if (_current_latency[n_idx] > _bundle_cycle_number) {
2167 #ifndef PRODUCT
2168 if (_cfg->C->trace_opto_output())
2169 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2170 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2171 #endif
2172 return (false);
2173 }
2174
2175 const Pipeline *node_pipeline = n->pipeline();
2176
2177 uint instruction_count = node_pipeline->instructionCount();
2178 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2179 instruction_count = 0;
2180 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2181 instruction_count++;
2182
2183 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2184 #ifndef PRODUCT
2185 if (_cfg->C->trace_opto_output())
2186 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2187 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2188 #endif
2189 return (false);
2190 }
2191
2192 // Don't allow non-machine nodes to be handled this way
2193 if (!n->is_Mach() && instruction_count == 0)
2194 return (false);
2195
2196 // See if there is any overlap
2197 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2198
2199 if (delay > 0) {
2200 #ifndef PRODUCT
2201 if (_cfg->C->trace_opto_output())
2202 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2203 #endif
2204 return false;
2205 }
2206
2207 #ifndef PRODUCT
2208 if (_cfg->C->trace_opto_output())
2209 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2210 #endif
2211
2212 return true;
2213 }
2214
2215 Node * Scheduling::ChooseNodeToBundle() {
2216 uint siz = _available.size();
2217
2218 if (siz == 0) {
2219
2220 #ifndef PRODUCT
2221 if (_cfg->C->trace_opto_output())
2222 tty->print("# ChooseNodeToBundle: NULL\n");
2223 #endif
2224 return (NULL);
2225 }
2226
2227 // Fast path, if only 1 instruction in the bundle
2228 if (siz == 1) {
2229 #ifndef PRODUCT
2230 if (_cfg->C->trace_opto_output()) {
2231 tty->print("# ChooseNodeToBundle (only 1): ");
2232 _available[0]->dump();
2233 }
2234 #endif
2235 return (_available[0]);
2236 }
2237
2238 // Don't bother, if the bundle is already full
2239 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2240 for ( uint i = 0; i < siz; i++ ) {
2241 Node *n = _available[i];
2242
2243 // Skip projections, we'll handle them another way
2244 if (n->is_Proj())
2245 continue;
2246
2247 // This presupposed that instructions are inserted into the
2248 // available list in a legality order; i.e. instructions that
2249 // must be inserted first are at the head of the list
2250 if (NodeFitsInBundle(n)) {
2251 #ifndef PRODUCT
2252 if (_cfg->C->trace_opto_output()) {
2253 tty->print("# ChooseNodeToBundle: ");
2254 n->dump();
2255 }
2256 #endif
2257 return (n);
2258 }
2259 }
2260 }
2261
2262 // Nothing fits in this bundle, choose the highest priority
2263 #ifndef PRODUCT
2264 if (_cfg->C->trace_opto_output()) {
2265 tty->print("# ChooseNodeToBundle: ");
2266 _available[0]->dump();
2267 }
2268 #endif
2269
2270 return _available[0];
2271 }
2272
2273 void Scheduling::AddNodeToAvailableList(Node *n) {
2274 assert( !n->is_Proj(), "projections never directly made available" );
2275 #ifndef PRODUCT
2276 if (_cfg->C->trace_opto_output()) {
2277 tty->print("# AddNodeToAvailableList: ");
2278 n->dump();
2279 }
2280 #endif
2281
2282 int latency = _current_latency[n->_idx];
2283
2284 // Insert in latency order (insertion sort)
2285 uint i;
2286 for ( i=0; i < _available.size(); i++ )
2287 if (_current_latency[_available[i]->_idx] > latency)
2288 break;
2289
2290 // Special Check for compares following branches
2291 if( n->is_Mach() && _scheduled.size() > 0 ) {
2292 int op = n->as_Mach()->ideal_Opcode();
2293 Node *last = _scheduled[0];
2294 if( last->is_MachIf() && last->in(1) == n &&
2295 ( op == Op_CmpI ||
2296 op == Op_CmpU ||
2297 op == Op_CmpUL ||
2298 op == Op_CmpP ||
2299 op == Op_CmpF ||
2300 op == Op_CmpD ||
2301 op == Op_CmpL ) ) {
2302
2303 // Recalculate position, moving to front of same latency
2304 for ( i=0 ; i < _available.size(); i++ )
2305 if (_current_latency[_available[i]->_idx] >= latency)
2306 break;
2307 }
2308 }
2309
2310 // Insert the node in the available list
2311 _available.insert(i, n);
2312
2313 #ifndef PRODUCT
2314 if (_cfg->C->trace_opto_output())
2315 dump_available();
2316 #endif
2317 }
2318
2319 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2320 for ( uint i=0; i < n->len(); i++ ) {
2321 Node *def = n->in(i);
2322 if (!def) continue;
2323 if( def->is_Proj() ) // If this is a machine projection, then
2324 def = def->in(0); // propagate usage thru to the base instruction
2325
2326 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2327 continue;
2328 }
2329
2330 // Compute the latency
2331 uint l = _bundle_cycle_number + n->latency(i);
2332 if (_current_latency[def->_idx] < l)
2333 _current_latency[def->_idx] = l;
2334
2335 // If this does not have uses then schedule it
2336 if ((--_uses[def->_idx]) == 0)
2337 AddNodeToAvailableList(def);
2338 }
2339 }
2340
2341 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2342 #ifndef PRODUCT
2343 if (_cfg->C->trace_opto_output()) {
2344 tty->print("# AddNodeToBundle: ");
2345 n->dump();
2346 }
2347 #endif
2348
2349 // Remove this from the available list
2350 uint i;
2351 for (i = 0; i < _available.size(); i++)
2352 if (_available[i] == n)
2353 break;
2354 assert(i < _available.size(), "entry in _available list not found");
2355 _available.remove(i);
2356
2357 // See if this fits in the current bundle
2358 const Pipeline *node_pipeline = n->pipeline();
2359 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2360
2361 // Check for instructions to be placed in the delay slot. We
2362 // do this before we actually schedule the current instruction,
2363 // because the delay slot follows the current instruction.
2364 if (Pipeline::_branch_has_delay_slot &&
2365 node_pipeline->hasBranchDelay() &&
2366 !_unconditional_delay_slot) {
2367
2368 uint siz = _available.size();
2369
2370 // Conditional branches can support an instruction that
2371 // is unconditionally executed and not dependent by the
2372 // branch, OR a conditionally executed instruction if
2373 // the branch is taken. In practice, this means that
2374 // the first instruction at the branch target is
2375 // copied to the delay slot, and the branch goes to
2376 // the instruction after that at the branch target
2377 if ( n->is_MachBranch() ) {
2378
2379 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2380 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2381
2382 #ifndef PRODUCT
2383 _branches++;
2384 #endif
2385
2386 // At least 1 instruction is on the available list
2387 // that is not dependent on the branch
2388 for (uint i = 0; i < siz; i++) {
2389 Node *d = _available[i];
2390 const Pipeline *avail_pipeline = d->pipeline();
2391
2392 // Don't allow safepoints in the branch shadow, that will
2393 // cause a number of difficulties
2394 if ( avail_pipeline->instructionCount() == 1 &&
2395 !avail_pipeline->hasMultipleBundles() &&
2396 !avail_pipeline->hasBranchDelay() &&
2397 Pipeline::instr_has_unit_size() &&
2398 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2399 NodeFitsInBundle(d) &&
2400 !node_bundling(d)->used_in_delay()) {
2401
2402 if (d->is_Mach() && !d->is_MachSafePoint()) {
2403 // A node that fits in the delay slot was found, so we need to
2404 // set the appropriate bits in the bundle pipeline information so
2405 // that it correctly indicates resource usage. Later, when we
2406 // attempt to add this instruction to the bundle, we will skip
2407 // setting the resource usage.
2408 _unconditional_delay_slot = d;
2409 node_bundling(n)->set_use_unconditional_delay();
2410 node_bundling(d)->set_used_in_unconditional_delay();
2411 _bundle_use.add_usage(avail_pipeline->resourceUse());
2412 _current_latency[d->_idx] = _bundle_cycle_number;
2413 _next_node = d;
2414 ++_bundle_instr_count;
2415 #ifndef PRODUCT
2416 _unconditional_delays++;
2417 #endif
2418 break;
2419 }
2420 }
2421 }
2422 }
2423
2424 // No delay slot, add a nop to the usage
2425 if (!_unconditional_delay_slot) {
2426 // See if adding an instruction in the delay slot will overflow
2427 // the bundle.
2428 if (!NodeFitsInBundle(_nop)) {
2429 #ifndef PRODUCT
2430 if (_cfg->C->trace_opto_output())
2431 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2432 #endif
2433 step(1);
2434 }
2435
2436 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2437 _next_node = _nop;
2438 ++_bundle_instr_count;
2439 }
2440
2441 // See if the instruction in the delay slot requires a
2442 // step of the bundles
2443 if (!NodeFitsInBundle(n)) {
2444 #ifndef PRODUCT
2445 if (_cfg->C->trace_opto_output())
2446 tty->print("# *** STEP(branch won't fit) ***\n");
2447 #endif
2448 // Update the state information
2449 _bundle_instr_count = 0;
2450 _bundle_cycle_number += 1;
2451 _bundle_use.step(1);
2452 }
2453 }
2454
2455 // Get the number of instructions
2456 uint instruction_count = node_pipeline->instructionCount();
2457 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2458 instruction_count = 0;
2459
2460 // Compute the latency information
2461 uint delay = 0;
2462
2463 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2464 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2465 if (relative_latency < 0)
2466 relative_latency = 0;
2467
2468 delay = _bundle_use.full_latency(relative_latency, node_usage);
2469
2470 // Does not fit in this bundle, start a new one
2471 if (delay > 0) {
2472 step(delay);
2473
2474 #ifndef PRODUCT
2475 if (_cfg->C->trace_opto_output())
2476 tty->print("# *** STEP(%d) ***\n", delay);
2477 #endif
2478 }
2479 }
2480
2481 // If this was placed in the delay slot, ignore it
2482 if (n != _unconditional_delay_slot) {
2483
2484 if (delay == 0) {
2485 if (node_pipeline->hasMultipleBundles()) {
2486 #ifndef PRODUCT
2487 if (_cfg->C->trace_opto_output())
2488 tty->print("# *** STEP(multiple instructions) ***\n");
2489 #endif
2490 step(1);
2491 }
2492
2493 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2494 #ifndef PRODUCT
2495 if (_cfg->C->trace_opto_output())
2496 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2497 instruction_count + _bundle_instr_count,
2498 Pipeline::_max_instrs_per_cycle);
2499 #endif
2500 step(1);
2501 }
2502 }
2503
2504 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2505 _bundle_instr_count++;
2506
2507 // Set the node's latency
2508 _current_latency[n->_idx] = _bundle_cycle_number;
2509
2510 // Now merge the functional unit information
2511 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2512 _bundle_use.add_usage(node_usage);
2513
2514 // Increment the number of instructions in this bundle
2515 _bundle_instr_count += instruction_count;
2516
2517 // Remember this node for later
2518 if (n->is_Mach())
2519 _next_node = n;
2520 }
2521
2522 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2523 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2524 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2525 // into the block. All other scheduled nodes get put in the schedule here.
2526 int op = n->Opcode();
2527 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2528 (op != Op_Node && // Not an unused antidepedence node and
2529 // not an unallocated boxlock
2530 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2531
2532 // Push any trailing projections
2533 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2534 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2535 Node *foi = n->fast_out(i);
2536 if( foi->is_Proj() )
2537 _scheduled.push(foi);
2538 }
2539 }
2540
2541 // Put the instruction in the schedule list
2542 _scheduled.push(n);
2543 }
2544
2545 #ifndef PRODUCT
2546 if (_cfg->C->trace_opto_output())
2547 dump_available();
2548 #endif
2549
2550 // Walk all the definitions, decrementing use counts, and
2551 // if a definition has a 0 use count, place it in the available list.
2552 DecrementUseCounts(n,bb);
2553 }
2554
2555 // This method sets the use count within a basic block. We will ignore all
2556 // uses outside the current basic block. As we are doing a backwards walk,
2557 // any node we reach that has a use count of 0 may be scheduled. This also
2558 // avoids the problem of cyclic references from phi nodes, as long as phi
2559 // nodes are at the front of the basic block. This method also initializes
2560 // the available list to the set of instructions that have no uses within this
2561 // basic block.
2562 void Scheduling::ComputeUseCount(const Block *bb) {
2563 #ifndef PRODUCT
2564 if (_cfg->C->trace_opto_output())
2565 tty->print("# -> ComputeUseCount\n");
2566 #endif
2567
2568 // Clear the list of available and scheduled instructions, just in case
2569 _available.clear();
2570 _scheduled.clear();
2571
2572 // No delay slot specified
2573 _unconditional_delay_slot = NULL;
2574
2575 #ifdef ASSERT
2576 for( uint i=0; i < bb->number_of_nodes(); i++ )
2577 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2578 #endif
2579
2580 // Force the _uses count to never go to zero for unscheduable pieces
2581 // of the block
2582 for( uint k = 0; k < _bb_start; k++ )
2583 _uses[bb->get_node(k)->_idx] = 1;
2584 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2585 _uses[bb->get_node(l)->_idx] = 1;
2586
2587 // Iterate backwards over the instructions in the block. Don't count the
2588 // branch projections at end or the block header instructions.
2589 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2590 Node *n = bb->get_node(j);
2591 if( n->is_Proj() ) continue; // Projections handled another way
2592
2593 // Account for all uses
2594 for ( uint k = 0; k < n->len(); k++ ) {
2595 Node *inp = n->in(k);
2596 if (!inp) continue;
2597 assert(inp != n, "no cycles allowed" );
2598 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2599 if (inp->is_Proj()) { // Skip through Proj's
2600 inp = inp->in(0);
2601 }
2602 ++_uses[inp->_idx]; // Count 1 block-local use
2603 }
2604 }
2605
2606 // If this instruction has a 0 use count, then it is available
2607 if (!_uses[n->_idx]) {
2608 _current_latency[n->_idx] = _bundle_cycle_number;
2609 AddNodeToAvailableList(n);
2610 }
2611
2612 #ifndef PRODUCT
2613 if (_cfg->C->trace_opto_output()) {
2614 tty->print("# uses: %3d: ", _uses[n->_idx]);
2615 n->dump();
2616 }
2617 #endif
2618 }
2619
2620 #ifndef PRODUCT
2621 if (_cfg->C->trace_opto_output())
2622 tty->print("# <- ComputeUseCount\n");
2623 #endif
2624 }
2625
2626 // This routine performs scheduling on each basic block in reverse order,
2627 // using instruction latencies and taking into account function unit
2628 // availability.
2629 void Scheduling::DoScheduling() {
2630 #ifndef PRODUCT
2631 if (_cfg->C->trace_opto_output())
2632 tty->print("# -> DoScheduling\n");
2633 #endif
2634
2635 Block *succ_bb = NULL;
2636 Block *bb;
2637 Compile* C = Compile::current();
2638
2639 // Walk over all the basic blocks in reverse order
2640 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2641 bb = _cfg->get_block(i);
2642
2643 #ifndef PRODUCT
2644 if (_cfg->C->trace_opto_output()) {
2645 tty->print("# Schedule BB#%03d (initial)\n", i);
2646 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2647 bb->get_node(j)->dump();
2648 }
2649 }
2650 #endif
2651
2652 // On the head node, skip processing
2653 if (bb == _cfg->get_root_block()) {
2654 continue;
2655 }
2656
2657 // Skip empty, connector blocks
2658 if (bb->is_connector())
2659 continue;
2660
2661 // If the following block is not the sole successor of
2662 // this one, then reset the pipeline information
2663 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2664 #ifndef PRODUCT
2665 if (_cfg->C->trace_opto_output()) {
2666 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2667 _next_node->_idx, _bundle_instr_count);
2668 }
2669 #endif
2670 step_and_clear();
2671 }
2672
2673 // Leave untouched the starting instruction, any Phis, a CreateEx node
2674 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2675 _bb_end = bb->number_of_nodes()-1;
2676 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2677 Node *n = bb->get_node(_bb_start);
2678 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2679 // Also, MachIdealNodes do not get scheduled
2680 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2681 MachNode *mach = n->as_Mach();
2682 int iop = mach->ideal_Opcode();
2683 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2684 if( iop == Op_Con ) continue; // Do not schedule Top
2685 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2686 mach->pipeline() == MachNode::pipeline_class() &&
2687 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2688 continue;
2689 break; // Funny loop structure to be sure...
2690 }
2691 // Compute last "interesting" instruction in block - last instruction we
2692 // might schedule. _bb_end points just after last schedulable inst. We
2693 // normally schedule conditional branches (despite them being forced last
2694 // in the block), because they have delay slots we can fill. Calls all
2695 // have their delay slots filled in the template expansions, so we don't
2696 // bother scheduling them.
2697 Node *last = bb->get_node(_bb_end);
2698 // Ignore trailing NOPs.
2699 while (_bb_end > 0 && last->is_Mach() &&
2700 last->as_Mach()->ideal_Opcode() == Op_Con) {
2701 last = bb->get_node(--_bb_end);
2702 }
2703 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2704 if( last->is_Catch() ||
2705 (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2706 // There might be a prior call. Skip it.
2707 while (_bb_start < _bb_end && bb->get_node(--_bb_end)->is_MachProj());
2708 } else if( last->is_MachNullCheck() ) {
2709 // Backup so the last null-checked memory instruction is
2710 // outside the schedulable range. Skip over the nullcheck,
2711 // projection, and the memory nodes.
2712 Node *mem = last->in(1);
2713 do {
2714 _bb_end--;
2715 } while (mem != bb->get_node(_bb_end));
2716 } else {
2717 // Set _bb_end to point after last schedulable inst.
2718 _bb_end++;
2719 }
2720
2721 assert( _bb_start <= _bb_end, "inverted block ends" );
2722
2723 // Compute the register antidependencies for the basic block
2724 ComputeRegisterAntidependencies(bb);
2725 if (C->failing()) return; // too many D-U pinch points
2726
2727 // Compute intra-bb latencies for the nodes
2728 ComputeLocalLatenciesForward(bb);
2729
2730 // Compute the usage within the block, and set the list of all nodes
2731 // in the block that have no uses within the block.
2732 ComputeUseCount(bb);
2733
2734 // Schedule the remaining instructions in the block
2735 while ( _available.size() > 0 ) {
2736 Node *n = ChooseNodeToBundle();
2737 guarantee(n != NULL, "no nodes available");
2738 AddNodeToBundle(n,bb);
2739 }
2740
2741 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2742 #ifdef ASSERT
2743 for( uint l = _bb_start; l < _bb_end; l++ ) {
2744 Node *n = bb->get_node(l);
2745 uint m;
2746 for( m = 0; m < _bb_end-_bb_start; m++ )
2747 if( _scheduled[m] == n )
2748 break;
2749 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2750 }
2751 #endif
2752
2753 // Now copy the instructions (in reverse order) back to the block
2754 for ( uint k = _bb_start; k < _bb_end; k++ )
2755 bb->map_node(_scheduled[_bb_end-k-1], k);
2756
2757 #ifndef PRODUCT
2758 if (_cfg->C->trace_opto_output()) {
2759 tty->print("# Schedule BB#%03d (final)\n", i);
2760 uint current = 0;
2761 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2762 Node *n = bb->get_node(j);
2763 if( valid_bundle_info(n) ) {
2764 Bundle *bundle = node_bundling(n);
2765 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2766 tty->print("*** Bundle: ");
2767 bundle->dump();
2768 }
2769 n->dump();
2770 }
2771 }
2772 }
2773 #endif
2774 #ifdef ASSERT
2775 verify_good_schedule(bb,"after block local scheduling");
2776 #endif
2777 }
2778
2779 #ifndef PRODUCT
2780 if (_cfg->C->trace_opto_output())
2781 tty->print("# <- DoScheduling\n");
2782 #endif
2783
2784 // Record final node-bundling array location
2785 _regalloc->C->output()->set_node_bundling_base(_node_bundling_base);
2786
2787 } // end DoScheduling
2788
2789 // Verify that no live-range used in the block is killed in the block by a
2790 // wrong DEF. This doesn't verify live-ranges that span blocks.
2791
2792 // Check for edge existence. Used to avoid adding redundant precedence edges.
2793 static bool edge_from_to( Node *from, Node *to ) {
2794 for( uint i=0; i<from->len(); i++ )
2795 if( from->in(i) == to )
2796 return true;
2797 return false;
2798 }
2799
2800 #ifdef ASSERT
2801 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2802 // Check for bad kills
2803 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2804 Node *prior_use = _reg_node[def];
2805 if( prior_use && !edge_from_to(prior_use,n) ) {
2806 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2807 n->dump();
2808 tty->print_cr("...");
2809 prior_use->dump();
2810 assert(edge_from_to(prior_use,n), "%s", msg);
2811 }
2812 _reg_node.map(def,NULL); // Kill live USEs
2813 }
2814 }
2815
2816 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2817
2818 // Zap to something reasonable for the verify code
2819 _reg_node.clear();
2820
2821 // Walk over the block backwards. Check to make sure each DEF doesn't
2822 // kill a live value (other than the one it's supposed to). Add each
2823 // USE to the live set.
2824 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2825 Node *n = b->get_node(i);
2826 int n_op = n->Opcode();
2827 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2828 // Fat-proj kills a slew of registers
2829 RegMask rm = n->out_RegMask();// Make local copy
2830 while( rm.is_NotEmpty() ) {
2831 OptoReg::Name kill = rm.find_first_elem();
2832 rm.Remove(kill);
2833 verify_do_def( n, kill, msg );
2834 }
2835 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2836 // Get DEF'd registers the normal way
2837 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2838 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2839 }
2840
2841 // Now make all USEs live
2842 for( uint i=1; i<n->req(); i++ ) {
2843 Node *def = n->in(i);
2844 assert(def != 0, "input edge required");
2845 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2846 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2847 if( OptoReg::is_valid(reg_lo) ) {
2848 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2849 _reg_node.map(reg_lo,n);
2850 }
2851 if( OptoReg::is_valid(reg_hi) ) {
2852 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2853 _reg_node.map(reg_hi,n);
2854 }
2855 }
2856
2857 }
2858
2859 // Zap to something reasonable for the Antidependence code
2860 _reg_node.clear();
2861 }
2862 #endif
2863
2864 // Conditionally add precedence edges. Avoid putting edges on Projs.
2865 static void add_prec_edge_from_to( Node *from, Node *to ) {
2866 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2867 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2868 from = from->in(0);
2869 }
2870 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2871 !edge_from_to( from, to ) ) // Avoid duplicate edge
2872 from->add_prec(to);
2873 }
2874
2875 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2876 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2877 return;
2878
2879 Node *pinch = _reg_node[def_reg]; // Get pinch point
2880 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2881 is_def ) { // Check for a true def (not a kill)
2882 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2883 return;
2884 }
2885
2886 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2887 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2888
2889 // After some number of kills there _may_ be a later def
2890 Node *later_def = NULL;
2891
2892 Compile* C = Compile::current();
2893
2894 // Finding a kill requires a real pinch-point.
2895 // Check for not already having a pinch-point.
2896 // Pinch points are Op_Node's.
2897 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2898 later_def = pinch; // Must be def/kill as optimistic pinch-point
2899 if ( _pinch_free_list.size() > 0) {
2900 pinch = _pinch_free_list.pop();
2901 } else {
2902 pinch = new Node(1); // Pinch point to-be
2903 }
2904 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2905 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2906 return;
2907 }
2908 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2909 _reg_node.map(def_reg,pinch); // Record pinch-point
2910 //regalloc()->set_bad(pinch->_idx); // Already initialized this way.
2911 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2912 pinch->init_req(0, C->top()); // set not NULL for the next call
2913 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2914 later_def = NULL; // and no later def
2915 }
2916 pinch->set_req(0,later_def); // Hook later def so we can find it
2917 } else { // Else have valid pinch point
2918 if( pinch->in(0) ) // If there is a later-def
2919 later_def = pinch->in(0); // Get it
2920 }
2921
2922 // Add output-dependence edge from later def to kill
2923 if( later_def ) // If there is some original def
2924 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2925
2926 // See if current kill is also a use, and so is forced to be the pinch-point.
2927 if( pinch->Opcode() == Op_Node ) {
2928 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2929 for( uint i=1; i<uses->req(); i++ ) {
2930 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2931 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2932 // Yes, found a use/kill pinch-point
2933 pinch->set_req(0,NULL); //
2934 pinch->replace_by(kill); // Move anti-dep edges up
2935 pinch = kill;
2936 _reg_node.map(def_reg,pinch);
2937 return;
2938 }
2939 }
2940 }
2941
2942 // Add edge from kill to pinch-point
2943 add_prec_edge_from_to(kill,pinch);
2944 }
2945
2946 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2947 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2948 return;
2949 Node *pinch = _reg_node[use_reg]; // Get pinch point
2950 // Check for no later def_reg/kill in block
2951 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2952 // Use has to be block-local as well
2953 _cfg->get_block_for_node(use) == b) {
2954 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2955 pinch->req() == 1 ) { // pinch not yet in block?
2956 pinch->del_req(0); // yank pointer to later-def, also set flag
2957 // Insert the pinch-point in the block just after the last use
2958 b->insert_node(pinch, b->find_node(use) + 1);
2959 _bb_end++; // Increase size scheduled region in block
2960 }
2961
2962 add_prec_edge_from_to(pinch,use);
2963 }
2964 }
2965
2966 // We insert antidependences between the reads and following write of
2967 // allocated registers to prevent illegal code motion. Hopefully, the
2968 // number of added references should be fairly small, especially as we
2969 // are only adding references within the current basic block.
2970 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2971
2972 #ifdef ASSERT
2973 verify_good_schedule(b,"before block local scheduling");
2974 #endif
2975
2976 // A valid schedule, for each register independently, is an endless cycle
2977 // of: a def, then some uses (connected to the def by true dependencies),
2978 // then some kills (defs with no uses), finally the cycle repeats with a new
2979 // def. The uses are allowed to float relative to each other, as are the
2980 // kills. No use is allowed to slide past a kill (or def). This requires
2981 // antidependencies between all uses of a single def and all kills that
2982 // follow, up to the next def. More edges are redundant, because later defs
2983 // & kills are already serialized with true or antidependencies. To keep
2984 // the edge count down, we add a 'pinch point' node if there's more than
2985 // one use or more than one kill/def.
2986
2987 // We add dependencies in one bottom-up pass.
2988
2989 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2990
2991 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2992 // register. If not, we record the DEF/KILL in _reg_node, the
2993 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2994 // "pinch point", a new Node that's in the graph but not in the block.
2995 // We put edges from the prior and current DEF/KILLs to the pinch point.
2996 // We put the pinch point in _reg_node. If there's already a pinch point
2997 // we merely add an edge from the current DEF/KILL to the pinch point.
2998
2999 // After doing the DEF/KILLs, we handle USEs. For each used register, we
3000 // put an edge from the pinch point to the USE.
3001
3002 // To be expedient, the _reg_node array is pre-allocated for the whole
3003 // compilation. _reg_node is lazily initialized; it either contains a NULL,
3004 // or a valid def/kill/pinch-point, or a leftover node from some prior
3005 // block. Leftover node from some prior block is treated like a NULL (no
3006 // prior def, so no anti-dependence needed). Valid def is distinguished by
3007 // it being in the current block.
3008 bool fat_proj_seen = false;
3009 uint last_safept = _bb_end-1;
3010 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
3011 Node* last_safept_node = end_node;
3012 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
3013 Node *n = b->get_node(i);
3014 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
3015 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
3016 // Fat-proj kills a slew of registers
3017 // This can add edges to 'n' and obscure whether or not it was a def,
3018 // hence the is_def flag.
3019 fat_proj_seen = true;
3020 RegMask rm = n->out_RegMask();// Make local copy
3021 while( rm.is_NotEmpty() ) {
3022 OptoReg::Name kill = rm.find_first_elem();
3023 rm.Remove(kill);
3024 anti_do_def( b, n, kill, is_def );
3025 }
3026 } else {
3027 // Get DEF'd registers the normal way
3028 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
3029 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
3030 }
3031
3032 // Kill projections on a branch should appear to occur on the
3033 // branch, not afterwards, so grab the masks from the projections
3034 // and process them.
3035 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
3036 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3037 Node* use = n->fast_out(i);
3038 if (use->is_Proj()) {
3039 RegMask rm = use->out_RegMask();// Make local copy
3040 while( rm.is_NotEmpty() ) {
3041 OptoReg::Name kill = rm.find_first_elem();
3042 rm.Remove(kill);
3043 anti_do_def( b, n, kill, false );
3044 }
3045 }
3046 }
3047 }
3048
3049 // Check each register used by this instruction for a following DEF/KILL
3050 // that must occur afterward and requires an anti-dependence edge.
3051 for( uint j=0; j<n->req(); j++ ) {
3052 Node *def = n->in(j);
3053 if( def ) {
3054 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
3055 anti_do_use( b, n, _regalloc->get_reg_first(def) );
3056 anti_do_use( b, n, _regalloc->get_reg_second(def) );
3057 }
3058 }
3059 // Do not allow defs of new derived values to float above GC
3060 // points unless the base is definitely available at the GC point.
3061
3062 Node *m = b->get_node(i);
3063
3064 // Add precedence edge from following safepoint to use of derived pointer
3065 if( last_safept_node != end_node &&
3066 m != last_safept_node) {
3067 for (uint k = 1; k < m->req(); k++) {
3068 const Type *t = m->in(k)->bottom_type();
3069 if( t->isa_oop_ptr() &&
3070 t->is_ptr()->offset() != 0 ) {
3071 last_safept_node->add_prec( m );
3072 break;
3073 }
3074 }
3075 }
3076
3077 if( n->jvms() ) { // Precedence edge from derived to safept
3078 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
3079 if( b->get_node(last_safept) != last_safept_node ) {
3080 last_safept = b->find_node(last_safept_node);
3081 }
3082 for( uint j=last_safept; j > i; j-- ) {
3083 Node *mach = b->get_node(j);
3084 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
3085 mach->add_prec( n );
3086 }
3087 last_safept = i;
3088 last_safept_node = m;
3089 }
3090 }
3091
3092 if (fat_proj_seen) {
3093 // Garbage collect pinch nodes that were not consumed.
3094 // They are usually created by a fat kill MachProj for a call.
3095 garbage_collect_pinch_nodes();
3096 }
3097 }
3098
3099 // Garbage collect pinch nodes for reuse by other blocks.
3100 //
3101 // The block scheduler's insertion of anti-dependence
3102 // edges creates many pinch nodes when the block contains
3103 // 2 or more Calls. A pinch node is used to prevent a
3104 // combinatorial explosion of edges. If a set of kills for a
3105 // register is anti-dependent on a set of uses (or defs), rather
3106 // than adding an edge in the graph between each pair of kill
3107 // and use (or def), a pinch is inserted between them:
3108 //
3109 // use1 use2 use3
3110 // \ | /
3111 // \ | /
3112 // pinch
3113 // / | \
3114 // / | \
3115 // kill1 kill2 kill3
3116 //
3117 // One pinch node is created per register killed when
3118 // the second call is encountered during a backwards pass
3119 // over the block. Most of these pinch nodes are never
3120 // wired into the graph because the register is never
3121 // used or def'ed in the block.
3122 //
3123 void Scheduling::garbage_collect_pinch_nodes() {
3124 #ifndef PRODUCT
3125 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
3126 #endif
3127 int trace_cnt = 0;
3128 for (uint k = 0; k < _reg_node.Size(); k++) {
3129 Node* pinch = _reg_node[k];
3130 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
3131 // no predecence input edges
3132 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
3133 cleanup_pinch(pinch);
3134 _pinch_free_list.push(pinch);
3135 _reg_node.map(k, NULL);
3136 #ifndef PRODUCT
3137 if (_cfg->C->trace_opto_output()) {
3138 trace_cnt++;
3139 if (trace_cnt > 40) {
3140 tty->print("\n");
3141 trace_cnt = 0;
3142 }
3143 tty->print(" %d", pinch->_idx);
3144 }
3145 #endif
3146 }
3147 }
3148 #ifndef PRODUCT
3149 if (_cfg->C->trace_opto_output()) tty->print("\n");
3150 #endif
3151 }
3152
3153 // Clean up a pinch node for reuse.
3154 void Scheduling::cleanup_pinch( Node *pinch ) {
3155 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3156
3157 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3158 Node* use = pinch->last_out(i);
3159 uint uses_found = 0;
3160 for (uint j = use->req(); j < use->len(); j++) {
3161 if (use->in(j) == pinch) {
3162 use->rm_prec(j);
3163 uses_found++;
3164 }
3165 }
3166 assert(uses_found > 0, "must be a precedence edge");
3167 i -= uses_found; // we deleted 1 or more copies of this edge
3168 }
3169 // May have a later_def entry
3170 pinch->set_req(0, NULL);
3171 }
3172
3173 #ifndef PRODUCT
3174
3175 void Scheduling::dump_available() const {
3176 tty->print("#Availist ");
3177 for (uint i = 0; i < _available.size(); i++)
3178 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3179 tty->cr();
3180 }
3181
3182 // Print Scheduling Statistics
3183 void Scheduling::print_statistics() {
3184 // Print the size added by nops for bundling
3185 tty->print("Nops added %d bytes to total of %d bytes",
3186 _total_nop_size, _total_method_size);
3187 if (_total_method_size > 0)
3188 tty->print(", for %.2f%%",
3189 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3190 tty->print("\n");
3191
3192 // Print the number of branch shadows filled
3193 if (Pipeline::_branch_has_delay_slot) {
3194 tty->print("Of %d branches, %d had unconditional delay slots filled",
3195 _total_branches, _total_unconditional_delays);
3196 if (_total_branches > 0)
3197 tty->print(", for %.2f%%",
3198 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3199 tty->print("\n");
3200 }
3201
3202 uint total_instructions = 0, total_bundles = 0;
3203
3204 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3205 uint bundle_count = _total_instructions_per_bundle[i];
3206 total_instructions += bundle_count * i;
3207 total_bundles += bundle_count;
3208 }
3209
3210 if (total_bundles > 0)
3211 tty->print("Average ILP (excluding nops) is %.2f\n",
3212 ((double)total_instructions) / ((double)total_bundles));
3213 }
3214 #endif
3215
3216 //-----------------------init_scratch_buffer_blob------------------------------
3217 // Construct a temporary BufferBlob and cache it for this compile.
3218 void PhaseOutput::init_scratch_buffer_blob(int const_size) {
3219 // If there is already a scratch buffer blob allocated and the
3220 // constant section is big enough, use it. Otherwise free the
3221 // current and allocate a new one.
3222 BufferBlob* blob = scratch_buffer_blob();
3223 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
3224 // Use the current blob.
3225 } else {
3226 if (blob != NULL) {
3227 BufferBlob::free(blob);
3228 }
3229
3230 ResourceMark rm;
3231 _scratch_const_size = const_size;
3232 int size = C2Compiler::initial_code_buffer_size(const_size);
3233 blob = BufferBlob::create("Compile::scratch_buffer", size);
3234 // Record the buffer blob for next time.
3235 set_scratch_buffer_blob(blob);
3236 // Have we run out of code space?
3237 if (scratch_buffer_blob() == NULL) {
3238 // Let CompilerBroker disable further compilations.
3239 C->record_failure("Not enough space for scratch buffer in CodeCache");
3240 return;
3241 }
3242 }
3243
3244 // Initialize the relocation buffers
3245 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
3246 set_scratch_locs_memory(locs_buf);
3247 }
3248
3249
3250 //-----------------------scratch_emit_size-------------------------------------
3251 // Helper function that computes size by emitting code
3252 uint PhaseOutput::scratch_emit_size(const Node* n) {
3253 // Start scratch_emit_size section.
3254 set_in_scratch_emit_size(true);
3255
3256 // Emit into a trash buffer and count bytes emitted.
3257 // This is a pretty expensive way to compute a size,
3258 // but it works well enough if seldom used.
3259 // All common fixed-size instructions are given a size
3260 // method by the AD file.
3261 // Note that the scratch buffer blob and locs memory are
3262 // allocated at the beginning of the compile task, and
3263 // may be shared by several calls to scratch_emit_size.
3264 // The allocation of the scratch buffer blob is particularly
3265 // expensive, since it has to grab the code cache lock.
3266 BufferBlob* blob = this->scratch_buffer_blob();
3267 assert(blob != NULL, "Initialize BufferBlob at start");
3268 assert(blob->size() > MAX_inst_size, "sanity");
3269 relocInfo* locs_buf = scratch_locs_memory();
3270 address blob_begin = blob->content_begin();
3271 address blob_end = (address)locs_buf;
3272 assert(blob->contains(blob_end), "sanity");
3273 CodeBuffer buf(blob_begin, blob_end - blob_begin);
3274 buf.initialize_consts_size(_scratch_const_size);
3275 buf.initialize_stubs_size(MAX_stubs_size);
3276 assert(locs_buf != NULL, "sanity");
3277 int lsize = MAX_locs_size / 3;
3278 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
3279 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
3280 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
3281 // Mark as scratch buffer.
3282 buf.consts()->set_scratch_emit();
3283 buf.insts()->set_scratch_emit();
3284 buf.stubs()->set_scratch_emit();
3285
3286 // Do the emission.
3287
3288 Label fakeL; // Fake label for branch instructions.
3289 Label* saveL = NULL;
3290 uint save_bnum = 0;
3291 bool is_branch = n->is_MachBranch();
3292 if (is_branch) {
3293 MacroAssembler masm(&buf);
3294 masm.bind(fakeL);
3295 n->as_MachBranch()->save_label(&saveL, &save_bnum);
3296 n->as_MachBranch()->label_set(&fakeL, 0);
3297 }
3298 n->emit(buf, C->regalloc());
3299
3300 // Emitting into the scratch buffer should not fail
3301 assert (!C->failing(), "Must not have pending failure. Reason is: %s", C->failure_reason());
3302
3303 if (is_branch) // Restore label.
3304 n->as_MachBranch()->label_set(saveL, save_bnum);
3305
3306 // End scratch_emit_size section.
3307 set_in_scratch_emit_size(false);
3308
3309 return buf.insts_size();
3310 }
3311
3312 void PhaseOutput::install() {
3313 if (C->stub_function() != NULL) {
3314 install_stub(C->stub_name(),
3315 C->save_argument_registers());
3316 } else {
3317 install_code(C->method(),
3318 C->entry_bci(),
3319 CompileBroker::compiler2(),
3320 C->has_unsafe_access(),
3321 SharedRuntime::is_wide_vector(C->max_vector_size()),
3322 C->rtm_state());
3323 }
3324 }
3325
3326 void PhaseOutput::install_code(ciMethod* target,
3327 int entry_bci,
3328 AbstractCompiler* compiler,
3329 bool has_unsafe_access,
3330 bool has_wide_vectors,
3331 RTMState rtm_state) {
3332 // Check if we want to skip execution of all compiled code.
3333 {
3334 #ifndef PRODUCT
3335 if (OptoNoExecute) {
3336 C->record_method_not_compilable("+OptoNoExecute"); // Flag as failed
3337 return;
3338 }
3339 #endif
3340 Compile::TracePhase tp("install_code", &timers[_t_registerMethod]);
3341
3342 if (C->is_osr_compilation()) {
3343 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
3344 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
3345 } else {
3346 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
3347 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
3348 }
3349
3350 C->env()->register_method(target,
3351 entry_bci,
3352 &_code_offsets,
3353 _orig_pc_slot_offset_in_bytes,
3354 code_buffer(),
3355 frame_size_in_words(),
3356 oop_map_set(),
3357 &_handler_table,
3358 inc_table(),
3359 compiler,
3360 has_unsafe_access,
3361 SharedRuntime::is_wide_vector(C->max_vector_size()),
3362 C->rtm_state());
3363
3364 if (C->log() != NULL) { // Print code cache state into compiler log
3365 C->log()->code_cache_state();
3366 }
3367 }
3368 }
3369 void PhaseOutput::install_stub(const char* stub_name,
3370 bool caller_must_gc_arguments) {
3371 // Entry point will be accessed using stub_entry_point();
3372 if (code_buffer() == NULL) {
3373 Matcher::soft_match_failure();
3374 } else {
3375 if (PrintAssembly && (WizardMode || Verbose))
3376 tty->print_cr("### Stub::%s", stub_name);
3377
3378 if (!C->failing()) {
3379 assert(C->fixed_slots() == 0, "no fixed slots used for runtime stubs");
3380
3381 // Make the NMethod
3382 // For now we mark the frame as never safe for profile stackwalking
3383 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
3384 code_buffer(),
3385 CodeOffsets::frame_never_safe,
3386 // _code_offsets.value(CodeOffsets::Frame_Complete),
3387 frame_size_in_words(),
3388 oop_map_set(),
3389 caller_must_gc_arguments);
3390 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
3391
3392 C->set_stub_entry_point(rs->entry_point());
3393 }
3394 }
3395 }
3396
3397 // Support for bundling info
3398 Bundle* PhaseOutput::node_bundling(const Node *n) {
3399 assert(valid_bundle_info(n), "oob");
3400 return &_node_bundling_base[n->_idx];
3401 }
3402
3403 bool PhaseOutput::valid_bundle_info(const Node *n) {
3404 return (_node_bundling_limit > n->_idx);
3405 }
3406
3407 //------------------------------frame_size_in_words-----------------------------
3408 // frame_slots in units of words
3409 int PhaseOutput::frame_size_in_words() const {
3410 // shift is 0 in LP32 and 1 in LP64
3411 const int shift = (LogBytesPerWord - LogBytesPerInt);
3412 int words = _frame_slots >> shift;
3413 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
3414 return words;
3415 }
3416
3417 // To bang the stack of this compiled method we use the stack size
3418 // that the interpreter would need in case of a deoptimization. This
3419 // removes the need to bang the stack in the deoptimization blob which
3420 // in turn simplifies stack overflow handling.
3421 int PhaseOutput::bang_size_in_bytes() const {
3422 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), C->interpreter_frame_size());
3423 }
3424
3425 //------------------------------dump_asm---------------------------------------
3426 // Dump formatted assembly
3427 #if defined(SUPPORT_OPTO_ASSEMBLY)
3428 void PhaseOutput::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
3429
3430 int pc_digits = 3; // #chars required for pc
3431 int sb_chars = 3; // #chars for "start bundle" indicator
3432 int tab_size = 8;
3433 if (pcs != NULL) {
3434 int max_pc = 0;
3435 for (uint i = 0; i < pc_limit; i++) {
3436 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3437 }
3438 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3439 }
3440 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3441
3442 bool cut_short = false;
3443 st->print_cr("#");
3444 st->print("# "); C->tf()->dump_on(st); st->cr();
3445 st->print_cr("#");
3446
3447 // For all blocks
3448 int pc = 0x0; // Program counter
3449 char starts_bundle = ' ';
3450 C->regalloc()->dump_frame();
3451
3452 Node *n = NULL;
3453 for (uint i = 0; i < C->cfg()->number_of_blocks(); i++) {
3454 if (VMThread::should_terminate()) {
3455 cut_short = true;
3456 break;
3457 }
3458 Block* block = C->cfg()->get_block(i);
3459 if (block->is_connector() && !Verbose) {
3460 continue;
3461 }
3462 n = block->head();
3463 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3464 pc = pcs[n->_idx];
3465 st->print("%*.*x", pc_digits, pc_digits, pc);
3466 }
3467 st->fill_to(prefix_len);
3468 block->dump_head(C->cfg(), st);
3469 if (block->is_connector()) {
3470 st->fill_to(prefix_len);
3471 st->print_cr("# Empty connector block");
3472 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3473 st->fill_to(prefix_len);
3474 st->print_cr("# Block is sole successor of call");
3475 }
3476
3477 // For all instructions
3478 Node *delay = NULL;
3479 for (uint j = 0; j < block->number_of_nodes(); j++) {
3480 if (VMThread::should_terminate()) {
3481 cut_short = true;
3482 break;
3483 }
3484 n = block->get_node(j);
3485 if (valid_bundle_info(n)) {
3486 Bundle* bundle = node_bundling(n);
3487 if (bundle->used_in_unconditional_delay()) {
3488 delay = n;
3489 continue;
3490 }
3491 if (bundle->starts_bundle()) {
3492 starts_bundle = '+';
3493 }
3494 }
3495
3496 if (WizardMode) {
3497 n->dump();
3498 }
3499
3500 if( !n->is_Region() && // Dont print in the Assembly
3501 !n->is_Phi() && // a few noisely useless nodes
3502 !n->is_Proj() &&
3503 !n->is_MachTemp() &&
3504 !n->is_SafePointScalarObject() &&
3505 !n->is_Catch() && // Would be nice to print exception table targets
3506 !n->is_MergeMem() && // Not very interesting
3507 !n->is_top() && // Debug info table constants
3508 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3509 ) {
3510 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3511 pc = pcs[n->_idx];
3512 st->print("%*.*x", pc_digits, pc_digits, pc);
3513 } else {
3514 st->fill_to(pc_digits);
3515 }
3516 st->print(" %c ", starts_bundle);
3517 starts_bundle = ' ';
3518 st->fill_to(prefix_len);
3519 n->format(C->regalloc(), st);
3520 st->cr();
3521 }
3522
3523 // If we have an instruction with a delay slot, and have seen a delay,
3524 // then back up and print it
3525 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3526 // Coverity finding - Explicit null dereferenced.
3527 guarantee(delay != NULL, "no unconditional delay instruction");
3528 if (WizardMode) delay->dump();
3529
3530 if (node_bundling(delay)->starts_bundle())
3531 starts_bundle = '+';
3532 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3533 pc = pcs[n->_idx];
3534 st->print("%*.*x", pc_digits, pc_digits, pc);
3535 } else {
3536 st->fill_to(pc_digits);
3537 }
3538 st->print(" %c ", starts_bundle);
3539 starts_bundle = ' ';
3540 st->fill_to(prefix_len);
3541 delay->format(C->regalloc(), st);
3542 st->cr();
3543 delay = NULL;
3544 }
3545
3546 // Dump the exception table as well
3547 if( n->is_Catch() && (Verbose || WizardMode) ) {
3548 // Print the exception table for this offset
3549 _handler_table.print_subtable_for(pc);
3550 }
3551 st->bol(); // Make sure we start on a new line
3552 }
3553 st->cr(); // one empty line between blocks
3554 assert(cut_short || delay == NULL, "no unconditional delay branch");
3555 } // End of per-block dump
3556
3557 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3558 }
3559 #endif
3560
3561 #ifndef PRODUCT
3562 void PhaseOutput::print_statistics() {
3563 Scheduling::print_statistics();
3564 }
3565 #endif