1 /*
2 * Copyright (c) 1998, 2014, 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 "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/ad.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/optoreg.hpp"
39 #include "opto/output.hpp"
40 #include "opto/regalloc.hpp"
41 #include "opto/runtime.hpp"
42 #include "opto/subnode.hpp"
43 #include "opto/type.hpp"
44 #include "runtime/handles.inline.hpp"
45 #include "utilities/xmlstream.hpp"
46
47 #ifndef PRODUCT
48 #define DEBUG_ARG(x) , x
49 #else
50 #define DEBUG_ARG(x)
51 #endif
52
53 // Convert Nodes to instruction bits and pass off to the VM
54 void Compile::Output() {
55 // RootNode goes
56 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
57
58 // The number of new nodes (mostly MachNop) is proportional to
59 // the number of java calls and inner loops which are aligned.
60 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
61 C->inner_loops()*(OptoLoopAlignment-1)),
62 "out of nodes before code generation" ) ) {
63 return;
64 }
65 // Make sure I can find the Start Node
66 Block *entry = _cfg->get_block(1);
67 Block *broot = _cfg->get_root_block();
68
69 const StartNode *start = entry->head()->as_Start();
70
71 // Replace StartNode with prolog
72 MachPrologNode *prolog = new MachPrologNode();
73 entry->map_node(prolog, 0);
74 _cfg->map_node_to_block(prolog, entry);
75 _cfg->unmap_node_from_block(start); // start is no longer in any block
76
77 // Virtual methods need an unverified entry point
78
79 if( is_osr_compilation() ) {
80 if( PoisonOSREntry ) {
81 // TODO: Should use a ShouldNotReachHereNode...
82 _cfg->insert( broot, 0, new MachBreakpointNode() );
83 }
84 } else {
85 if( _method && !_method->flags().is_static() ) {
86 // Insert unvalidated entry point
87 _cfg->insert( broot, 0, new MachUEPNode() );
88 }
89
90 }
91
92
93 // Break before main entry point
94 if( (_method && _method->break_at_execute())
95 #ifndef PRODUCT
96 ||(OptoBreakpoint && is_method_compilation())
97 ||(OptoBreakpointOSR && is_osr_compilation())
98 ||(OptoBreakpointC2R && !_method)
99 #endif
100 ) {
101 // checking for _method means that OptoBreakpoint does not apply to
102 // runtime stubs or frame converters
103 _cfg->insert( entry, 1, new MachBreakpointNode() );
104 }
105
106 // Insert epilogs before every return
107 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
108 Block* block = _cfg->get_block(i);
109 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
110 Node* m = block->end();
111 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
112 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
113 block->add_inst(epilog);
114 _cfg->map_node_to_block(epilog, block);
115 }
116 }
117 }
118
119 # ifdef ENABLE_ZAP_DEAD_LOCALS
120 if (ZapDeadCompiledLocals) {
121 Insert_zap_nodes();
122 }
123 # endif
124
125 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
126 blk_starts[0] = 0;
127
128 // Initialize code buffer and process short branches.
129 CodeBuffer* cb = init_buffer(blk_starts);
130
131 if (cb == NULL || failing()) {
132 return;
133 }
134
135 ScheduleAndBundle();
136
137 #ifndef PRODUCT
138 if (trace_opto_output()) {
139 tty->print("\n---- After ScheduleAndBundle ----\n");
140 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
141 tty->print("\nBB#%03d:\n", i);
142 Block* block = _cfg->get_block(i);
143 for (uint j = 0; j < block->number_of_nodes(); j++) {
144 Node* n = block->get_node(j);
145 OptoReg::Name reg = _regalloc->get_reg_first(n);
146 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
147 n->dump();
148 }
149 }
150 }
151 #endif
152
153 if (failing()) {
154 return;
155 }
156
157 BuildOopMaps();
158
159 if (failing()) {
160 return;
161 }
162
163 fill_buffer(cb, blk_starts);
164 }
165
166 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
167 // Determine if we need to generate a stack overflow check.
168 // Do it if the method is not a stub function and
169 // has java calls or has frame size > vm_page_size/8.
170 // The debug VM checks that deoptimization doesn't trigger an
171 // unexpected stack overflow (compiled method stack banging should
172 // guarantee it doesn't happen) so we always need the stack bang in
173 // a debug VM.
174 return (UseStackBanging && stub_function() == NULL &&
175 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
176 DEBUG_ONLY(|| true)));
177 }
178
179 bool Compile::need_register_stack_bang() const {
180 // Determine if we need to generate a register stack overflow check.
181 // This is only used on architectures which have split register
182 // and memory stacks (ie. IA64).
183 // Bang if the method is not a stub function and has java calls
184 return (stub_function() == NULL && has_java_calls());
185 }
186
187 # ifdef ENABLE_ZAP_DEAD_LOCALS
188
189
190 // In order to catch compiler oop-map bugs, we have implemented
191 // a debugging mode called ZapDeadCompilerLocals.
192 // This mode causes the compiler to insert a call to a runtime routine,
193 // "zap_dead_locals", right before each place in compiled code
194 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
195 // The runtime routine checks that locations mapped as oops are really
196 // oops, that locations mapped as values do not look like oops,
197 // and that locations mapped as dead are not used later
198 // (by zapping them to an invalid address).
199
200 int Compile::_CompiledZap_count = 0;
201
202 void Compile::Insert_zap_nodes() {
203 bool skip = false;
204
205
206 // Dink with static counts because code code without the extra
207 // runtime calls is MUCH faster for debugging purposes
208
209 if ( CompileZapFirst == 0 ) ; // nothing special
210 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
211 else if ( CompileZapFirst == CompiledZap_count() )
212 warning("starting zap compilation after skipping");
213
214 if ( CompileZapLast == -1 ) ; // nothing special
215 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
216 else if ( CompileZapLast == CompiledZap_count() )
217 warning("about to compile last zap");
218
219 ++_CompiledZap_count; // counts skipped zaps, too
220
221 if ( skip ) return;
222
223
224 if ( _method == NULL )
225 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
226
227 // Insert call to zap runtime stub before every node with an oop map
228 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
229 Block *b = _cfg->get_block(i);
230 for ( uint j = 0; j < b->number_of_nodes(); ++j ) {
231 Node *n = b->get_node(j);
232
233 // Determining if we should insert a zap-a-lot node in output.
234 // We do that for all nodes that has oopmap info, except for calls
235 // to allocation. Calls to allocation passes in the old top-of-eden pointer
236 // and expect the C code to reset it. Hence, there can be no safepoints between
237 // the inlined-allocation and the call to new_Java, etc.
238 // We also cannot zap monitor calls, as they must hold the microlock
239 // during the call to Zap, which also wants to grab the microlock.
240 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
241 if ( insert ) { // it is MachSafePoint
242 if ( !n->is_MachCall() ) {
243 insert = false;
244 } else if ( n->is_MachCall() ) {
245 MachCallNode* call = n->as_MachCall();
246 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
247 call->entry_point() == OptoRuntime::new_array_Java() ||
248 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
249 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
250 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
251 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
252 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
253 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
254 ) {
255 insert = false;
256 }
257 }
258 if (insert) {
259 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
260 b->insert_node(zap, j);
261 _cfg->map_node_to_block(zap, b);
262 ++j;
263 }
264 }
265 }
266 }
267 }
268
269
270 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
271 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
272 CallStaticJavaNode* ideal_node =
273 new CallStaticJavaNode( tf,
274 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
275 "call zap dead locals stub", 0, TypePtr::BOTTOM);
276 // We need to copy the OopMap from the site we're zapping at.
277 // We have to make a copy, because the zap site might not be
278 // a call site, and zap_dead is a call site.
279 OopMap* clone = node_to_check->oop_map()->deep_copy();
280
281 // Add the cloned OopMap to the zap node
282 ideal_node->set_oop_map(clone);
283 return _matcher->match_sfpt(ideal_node);
284 }
285
286 bool Compile::is_node_getting_a_safepoint( Node* n) {
287 // This code duplicates the logic prior to the call of add_safepoint
288 // below in this file.
289 if( n->is_MachSafePoint() ) return true;
290 return false;
291 }
292
293 # endif // ENABLE_ZAP_DEAD_LOCALS
294
295 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
296 // of a loop. When aligning a loop we need to provide enough instructions
297 // in cpu's fetch buffer to feed decoders. The loop alignment could be
298 // avoided if we have enough instructions in fetch buffer at the head of a loop.
299 // By default, the size is set to 999999 by Block's constructor so that
300 // a loop will be aligned if the size is not reset here.
301 //
302 // Note: Mach instructions could contain several HW instructions
303 // so the size is estimated only.
304 //
305 void Compile::compute_loop_first_inst_sizes() {
306 // The next condition is used to gate the loop alignment optimization.
307 // Don't aligned a loop if there are enough instructions at the head of a loop
308 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
309 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
310 // equal to 11 bytes which is the largest address NOP instruction.
311 if (MaxLoopPad < OptoLoopAlignment - 1) {
312 uint last_block = _cfg->number_of_blocks() - 1;
313 for (uint i = 1; i <= last_block; i++) {
314 Block* block = _cfg->get_block(i);
315 // Check the first loop's block which requires an alignment.
316 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
317 uint sum_size = 0;
318 uint inst_cnt = NumberOfLoopInstrToAlign;
319 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
320
321 // Check subsequent fallthrough blocks if the loop's first
322 // block(s) does not have enough instructions.
323 Block *nb = block;
324 while(inst_cnt > 0 &&
325 i < last_block &&
326 !_cfg->get_block(i + 1)->has_loop_alignment() &&
327 !nb->has_successor(block)) {
328 i++;
329 nb = _cfg->get_block(i);
330 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
331 } // while( inst_cnt > 0 && i < last_block )
332
333 block->set_first_inst_size(sum_size);
334 } // f( b->head()->is_Loop() )
335 } // for( i <= last_block )
336 } // if( MaxLoopPad < OptoLoopAlignment-1 )
337 }
338
339 // The architecture description provides short branch variants for some long
340 // branch instructions. Replace eligible long branches with short branches.
341 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
342 // Compute size of each block, method size, and relocation information size
343 uint nblocks = _cfg->number_of_blocks();
344
345 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
346 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
347 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
348
349 // Collect worst case block paddings
350 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
351 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
352
353 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
354 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
355
356 bool has_short_branch_candidate = false;
357
358 // Initialize the sizes to 0
359 code_size = 0; // Size in bytes of generated code
360 stub_size = 0; // Size in bytes of all stub entries
361 // Size in bytes of all relocation entries, including those in local stubs.
362 // Start with 2-bytes of reloc info for the unvalidated entry point
363 reloc_size = 1; // Number of relocation entries
364
365 // Make three passes. The first computes pessimistic blk_starts,
366 // relative jmp_offset and reloc_size information. The second performs
367 // short branch substitution using the pessimistic sizing. The
368 // third inserts nops where needed.
369
370 // Step one, perform a pessimistic sizing pass.
371 uint last_call_adr = max_juint;
372 uint last_avoid_back_to_back_adr = max_juint;
373 uint nop_size = (new MachNopNode())->size(_regalloc);
374 for (uint i = 0; i < nblocks; i++) { // For all blocks
375 Block* block = _cfg->get_block(i);
376
377 // During short branch replacement, we store the relative (to blk_starts)
378 // offset of jump in jmp_offset, rather than the absolute offset of jump.
379 // This is so that we do not need to recompute sizes of all nodes when
380 // we compute correct blk_starts in our next sizing pass.
381 jmp_offset[i] = 0;
382 jmp_size[i] = 0;
383 jmp_nidx[i] = -1;
384 DEBUG_ONLY( jmp_target[i] = 0; )
385 DEBUG_ONLY( jmp_rule[i] = 0; )
386
387 // Sum all instruction sizes to compute block size
388 uint last_inst = block->number_of_nodes();
389 uint blk_size = 0;
390 for (uint j = 0; j < last_inst; j++) {
391 Node* nj = block->get_node(j);
392 // Handle machine instruction nodes
393 if (nj->is_Mach()) {
394 MachNode *mach = nj->as_Mach();
395 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
396 reloc_size += mach->reloc();
397 if (mach->is_MachCall()) {
398 // add size information for trampoline stub
399 // class CallStubImpl is platform-specific and defined in the *.ad files.
400 stub_size += CallStubImpl::size_call_trampoline();
401 reloc_size += CallStubImpl::reloc_call_trampoline();
402
403 MachCallNode *mcall = mach->as_MachCall();
404 // This destination address is NOT PC-relative
405
406 mcall->method_set((intptr_t)mcall->entry_point());
407
408 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
409 stub_size += CompiledStaticCall::to_interp_stub_size();
410 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
411 }
412 } else if (mach->is_MachSafePoint()) {
413 // If call/safepoint are adjacent, account for possible
414 // nop to disambiguate the two safepoints.
415 // ScheduleAndBundle() can rearrange nodes in a block,
416 // check for all offsets inside this block.
417 if (last_call_adr >= blk_starts[i]) {
418 blk_size += nop_size;
419 }
420 }
421 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
422 // Nop is inserted between "avoid back to back" instructions.
423 // ScheduleAndBundle() can rearrange nodes in a block,
424 // check for all offsets inside this block.
425 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
426 blk_size += nop_size;
427 }
428 }
429 if (mach->may_be_short_branch()) {
430 if (!nj->is_MachBranch()) {
431 #ifndef PRODUCT
432 nj->dump(3);
433 #endif
434 Unimplemented();
435 }
436 assert(jmp_nidx[i] == -1, "block should have only one branch");
437 jmp_offset[i] = blk_size;
438 jmp_size[i] = nj->size(_regalloc);
439 jmp_nidx[i] = j;
440 has_short_branch_candidate = true;
441 }
442 }
443 blk_size += nj->size(_regalloc);
444 // Remember end of call offset
445 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
446 last_call_adr = blk_starts[i]+blk_size;
447 }
448 // Remember end of avoid_back_to_back offset
449 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
450 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
451 }
452 }
453
454 // When the next block starts a loop, we may insert pad NOP
455 // instructions. Since we cannot know our future alignment,
456 // assume the worst.
457 if (i < nblocks - 1) {
458 Block* nb = _cfg->get_block(i + 1);
459 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
460 if (max_loop_pad > 0) {
461 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
462 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
463 // If either is the last instruction in this block, bump by
464 // max_loop_pad in lock-step with blk_size, so sizing
465 // calculations in subsequent blocks still can conservatively
466 // detect that it may the last instruction in this block.
467 if (last_call_adr == blk_starts[i]+blk_size) {
468 last_call_adr += max_loop_pad;
469 }
470 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
471 last_avoid_back_to_back_adr += max_loop_pad;
472 }
473 blk_size += max_loop_pad;
474 block_worst_case_pad[i + 1] = max_loop_pad;
475 }
476 }
477
478 // Save block size; update total method size
479 blk_starts[i+1] = blk_starts[i]+blk_size;
480 }
481
482 // Step two, replace eligible long jumps.
483 bool progress = true;
484 uint last_may_be_short_branch_adr = max_juint;
485 while (has_short_branch_candidate && progress) {
486 progress = false;
487 has_short_branch_candidate = false;
488 int adjust_block_start = 0;
489 for (uint i = 0; i < nblocks; i++) {
490 Block* block = _cfg->get_block(i);
491 int idx = jmp_nidx[i];
492 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
493 if (mach != NULL && mach->may_be_short_branch()) {
494 #ifdef ASSERT
495 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
496 int j;
497 // Find the branch; ignore trailing NOPs.
498 for (j = block->number_of_nodes()-1; j>=0; j--) {
499 Node* n = block->get_node(j);
500 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
501 break;
502 }
503 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
504 #endif
505 int br_size = jmp_size[i];
506 int br_offs = blk_starts[i] + jmp_offset[i];
507
508 // This requires the TRUE branch target be in succs[0]
509 uint bnum = block->non_connector_successor(0)->_pre_order;
510 int offset = blk_starts[bnum] - br_offs;
511 if (bnum > i) { // adjust following block's offset
512 offset -= adjust_block_start;
513 }
514
515 // This block can be a loop header, account for the padding
516 // in the previous block.
517 int block_padding = block_worst_case_pad[i];
518 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
519 // In the following code a nop could be inserted before
520 // the branch which will increase the backward distance.
521 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
522 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
523
524 if (needs_padding && offset <= 0)
525 offset -= nop_size;
526
527 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
528 // We've got a winner. Replace this branch.
529 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
530
531 // Update the jmp_size.
532 int new_size = replacement->size(_regalloc);
533 int diff = br_size - new_size;
534 assert(diff >= (int)nop_size, "short_branch size should be smaller");
535 // Conservatively take into account padding between
536 // avoid_back_to_back branches. Previous branch could be
537 // converted into avoid_back_to_back branch during next
538 // rounds.
539 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
540 jmp_offset[i] += nop_size;
541 diff -= nop_size;
542 }
543 adjust_block_start += diff;
544 block->map_node(replacement, idx);
545 mach->subsume_by(replacement, C);
546 mach = replacement;
547 progress = true;
548
549 jmp_size[i] = new_size;
550 DEBUG_ONLY( jmp_target[i] = bnum; );
551 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
552 } else {
553 // The jump distance is not short, try again during next iteration.
554 has_short_branch_candidate = true;
555 }
556 } // (mach->may_be_short_branch())
557 if (mach != NULL && (mach->may_be_short_branch() ||
558 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
559 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
560 }
561 blk_starts[i+1] -= adjust_block_start;
562 }
563 }
564
565 #ifdef ASSERT
566 for (uint i = 0; i < nblocks; i++) { // For all blocks
567 if (jmp_target[i] != 0) {
568 int br_size = jmp_size[i];
569 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
570 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
571 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]);
572 }
573 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
574 }
575 }
576 #endif
577
578 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
579 // after ScheduleAndBundle().
580
581 // ------------------
582 // Compute size for code buffer
583 code_size = blk_starts[nblocks];
584
585 // Relocation records
586 reloc_size += 1; // Relo entry for exception handler
587
588 // Adjust reloc_size to number of record of relocation info
589 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
590 // a relocation index.
591 // The CodeBuffer will expand the locs array if this estimate is too low.
592 reloc_size *= 10 / sizeof(relocInfo);
593 }
594
595 //------------------------------FillLocArray-----------------------------------
596 // Create a bit of debug info and append it to the array. The mapping is from
597 // Java local or expression stack to constant, register or stack-slot. For
598 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
599 // entry has been taken care of and caller should skip it).
600 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
601 // This should never have accepted Bad before
602 assert(OptoReg::is_valid(regnum), "location must be valid");
603 return (OptoReg::is_reg(regnum))
604 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
605 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
606 }
607
608
609 ObjectValue*
610 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
611 for (int i = 0; i < objs->length(); i++) {
612 assert(objs->at(i)->is_object(), "corrupt object cache");
613 ObjectValue* sv = (ObjectValue*) objs->at(i);
614 if (sv->id() == id) {
615 return sv;
616 }
617 }
618 // Otherwise..
619 return NULL;
620 }
621
622 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
623 ObjectValue* sv ) {
624 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
625 objs->append(sv);
626 }
627
628
629 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
630 GrowableArray<ScopeValue*> *array,
631 GrowableArray<ScopeValue*> *objs ) {
632 assert( local, "use _top instead of null" );
633 if (array->length() != idx) {
634 assert(array->length() == idx + 1, "Unexpected array count");
635 // Old functionality:
636 // return
637 // New functionality:
638 // Assert if the local is not top. In product mode let the new node
639 // override the old entry.
640 assert(local == top(), "LocArray collision");
641 if (local == top()) {
642 return;
643 }
644 array->pop();
645 }
646 const Type *t = local->bottom_type();
647
648 // Is it a safepoint scalar object node?
649 if (local->is_SafePointScalarObject()) {
650 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
651
652 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
653 if (sv == NULL) {
654 ciKlass* cik = t->is_oopptr()->klass();
655 assert(cik->is_instance_klass() ||
656 cik->is_array_klass(), "Not supported allocation.");
657 sv = new ObjectValue(spobj->_idx,
658 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
659 Compile::set_sv_for_object_node(objs, sv);
660
661 uint first_ind = spobj->first_index(sfpt->jvms());
662 for (uint i = 0; i < spobj->n_fields(); i++) {
663 Node* fld_node = sfpt->in(first_ind+i);
664 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
665 }
666 }
667 array->append(sv);
668 return;
669 }
670
671 // Grab the register number for the local
672 OptoReg::Name regnum = _regalloc->get_reg_first(local);
673 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
674 // Record the double as two float registers.
675 // The register mask for such a value always specifies two adjacent
676 // float registers, with the lower register number even.
677 // Normally, the allocation of high and low words to these registers
678 // is irrelevant, because nearly all operations on register pairs
679 // (e.g., StoreD) treat them as a single unit.
680 // Here, we assume in addition that the words in these two registers
681 // stored "naturally" (by operations like StoreD and double stores
682 // within the interpreter) such that the lower-numbered register
683 // is written to the lower memory address. This may seem like
684 // a machine dependency, but it is not--it is a requirement on
685 // the author of the <arch>.ad file to ensure that, for every
686 // even/odd double-register pair to which a double may be allocated,
687 // the word in the even single-register is stored to the first
688 // memory word. (Note that register numbers are completely
689 // arbitrary, and are not tied to any machine-level encodings.)
690 #ifdef _LP64
691 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
692 array->append(new ConstantIntValue(0));
693 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
694 } else if ( t->base() == Type::Long ) {
695 array->append(new ConstantIntValue(0));
696 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
697 } else if ( t->base() == Type::RawPtr ) {
698 // jsr/ret return address which must be restored into a the full
699 // width 64-bit stack slot.
700 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
701 }
702 #else //_LP64
703 #ifdef SPARC
704 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
705 // For SPARC we have to swap high and low words for
706 // long values stored in a single-register (g0-g7).
707 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
708 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
709 } else
710 #endif //SPARC
711 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
712 // Repack the double/long as two jints.
713 // The convention the interpreter uses is that the second local
714 // holds the first raw word of the native double representation.
715 // This is actually reasonable, since locals and stack arrays
716 // grow downwards in all implementations.
717 // (If, on some machine, the interpreter's Java locals or stack
718 // were to grow upwards, the embedded doubles would be word-swapped.)
719 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
720 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
721 }
722 #endif //_LP64
723 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
724 OptoReg::is_reg(regnum) ) {
725 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
726 ? Location::float_in_dbl : Location::normal ));
727 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
728 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
729 ? Location::int_in_long : Location::normal ));
730 } else if( t->base() == Type::NarrowOop ) {
731 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
732 } else {
733 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
734 }
735 return;
736 }
737
738 // No register. It must be constant data.
739 switch (t->base()) {
740 case Type::Half: // Second half of a double
741 ShouldNotReachHere(); // Caller should skip 2nd halves
742 break;
743 case Type::AnyPtr:
744 array->append(new ConstantOopWriteValue(NULL));
745 break;
746 case Type::AryPtr:
747 case Type::InstPtr: // fall through
748 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
749 break;
750 case Type::NarrowOop:
751 if (t == TypeNarrowOop::NULL_PTR) {
752 array->append(new ConstantOopWriteValue(NULL));
753 } else {
754 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
755 }
756 break;
757 case Type::Int:
758 array->append(new ConstantIntValue(t->is_int()->get_con()));
759 break;
760 case Type::RawPtr:
761 // A return address (T_ADDRESS).
762 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
763 #ifdef _LP64
764 // Must be restored to the full-width 64-bit stack slot.
765 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
766 #else
767 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
768 #endif
769 break;
770 case Type::FloatCon: {
771 float f = t->is_float_constant()->getf();
772 array->append(new ConstantIntValue(jint_cast(f)));
773 break;
774 }
775 case Type::DoubleCon: {
776 jdouble d = t->is_double_constant()->getd();
777 #ifdef _LP64
778 array->append(new ConstantIntValue(0));
779 array->append(new ConstantDoubleValue(d));
780 #else
781 // Repack the double as two jints.
782 // The convention the interpreter uses is that the second local
783 // holds the first raw word of the native double representation.
784 // This is actually reasonable, since locals and stack arrays
785 // grow downwards in all implementations.
786 // (If, on some machine, the interpreter's Java locals or stack
787 // were to grow upwards, the embedded doubles would be word-swapped.)
788 jlong_accessor acc = { jlong_cast(d) };
789 array->append(new ConstantIntValue(acc.words[1]));
790 array->append(new ConstantIntValue(acc.words[0]));
791 #endif
792 break;
793 }
794 case Type::Long: {
795 jlong d = t->is_long()->get_con();
796 #ifdef _LP64
797 array->append(new ConstantIntValue(0));
798 array->append(new ConstantLongValue(d));
799 #else
800 // Repack the long as two jints.
801 // The convention the interpreter uses is that the second local
802 // holds the first raw word of the native double representation.
803 // This is actually reasonable, since locals and stack arrays
804 // grow downwards in all implementations.
805 // (If, on some machine, the interpreter's Java locals or stack
806 // were to grow upwards, the embedded doubles would be word-swapped.)
807 jlong_accessor acc = { d };
808 array->append(new ConstantIntValue(acc.words[1]));
809 array->append(new ConstantIntValue(acc.words[0]));
810 #endif
811 break;
812 }
813 case Type::Top: // Add an illegal value here
814 array->append(new LocationValue(Location()));
815 break;
816 default:
817 ShouldNotReachHere();
818 break;
819 }
820 }
821
822 // Determine if this node starts a bundle
823 bool Compile::starts_bundle(const Node *n) const {
824 return (_node_bundling_limit > n->_idx &&
825 _node_bundling_base[n->_idx].starts_bundle());
826 }
827
828 //--------------------------Process_OopMap_Node--------------------------------
829 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
830
831 // Handle special safepoint nodes for synchronization
832 MachSafePointNode *sfn = mach->as_MachSafePoint();
833 MachCallNode *mcall;
834
835 #ifdef ENABLE_ZAP_DEAD_LOCALS
836 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
837 #endif
838
839 int safepoint_pc_offset = current_offset;
840 bool is_method_handle_invoke = false;
841 bool return_oop = false;
842
843 // Add the safepoint in the DebugInfoRecorder
844 if( !mach->is_MachCall() ) {
845 mcall = NULL;
846 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
847 } else {
848 mcall = mach->as_MachCall();
849
850 // Is the call a MethodHandle call?
851 if (mcall->is_MachCallJava()) {
852 if (mcall->as_MachCallJava()->_method_handle_invoke) {
853 assert(has_method_handle_invokes(), "must have been set during call generation");
854 is_method_handle_invoke = true;
855 }
856 }
857
858 // Check if a call returns an object.
859 if (mcall->returns_pointer()) {
860 return_oop = true;
861 }
862 safepoint_pc_offset += mcall->ret_addr_offset();
863 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
864 }
865
866 // Loop over the JVMState list to add scope information
867 // Do not skip safepoints with a NULL method, they need monitor info
868 JVMState* youngest_jvms = sfn->jvms();
869 int max_depth = youngest_jvms->depth();
870
871 // Allocate the object pool for scalar-replaced objects -- the map from
872 // small-integer keys (which can be recorded in the local and ostack
873 // arrays) to descriptions of the object state.
874 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
875
876 // Visit scopes from oldest to youngest.
877 for (int depth = 1; depth <= max_depth; depth++) {
878 JVMState* jvms = youngest_jvms->of_depth(depth);
879 int idx;
880 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
881 // Safepoints that do not have method() set only provide oop-map and monitor info
882 // to support GC; these do not support deoptimization.
883 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
884 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
885 int num_mon = jvms->nof_monitors();
886 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
887 "JVMS local count must match that of the method");
888
889 // Add Local and Expression Stack Information
890
891 // Insert locals into the locarray
892 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
893 for( idx = 0; idx < num_locs; idx++ ) {
894 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
895 }
896
897 // Insert expression stack entries into the exparray
898 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
899 for( idx = 0; idx < num_exps; idx++ ) {
900 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
901 }
902
903 // Add in mappings of the monitors
904 assert( !method ||
905 !method->is_synchronized() ||
906 method->is_native() ||
907 num_mon > 0 ||
908 !GenerateSynchronizationCode,
909 "monitors must always exist for synchronized methods");
910
911 // Build the growable array of ScopeValues for exp stack
912 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
913
914 // Loop over monitors and insert into array
915 for (idx = 0; idx < num_mon; idx++) {
916 // Grab the node that defines this monitor
917 Node* box_node = sfn->monitor_box(jvms, idx);
918 Node* obj_node = sfn->monitor_obj(jvms, idx);
919
920 // Create ScopeValue for object
921 ScopeValue *scval = NULL;
922
923 if (obj_node->is_SafePointScalarObject()) {
924 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
925 scval = Compile::sv_for_node_id(objs, spobj->_idx);
926 if (scval == NULL) {
927 const Type *t = spobj->bottom_type();
928 ciKlass* cik = t->is_oopptr()->klass();
929 assert(cik->is_instance_klass() ||
930 cik->is_array_klass(), "Not supported allocation.");
931 ObjectValue* sv = new ObjectValue(spobj->_idx,
932 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
933 Compile::set_sv_for_object_node(objs, sv);
934
935 uint first_ind = spobj->first_index(youngest_jvms);
936 for (uint i = 0; i < spobj->n_fields(); i++) {
937 Node* fld_node = sfn->in(first_ind+i);
938 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
939 }
940 scval = sv;
941 }
942 } else if (!obj_node->is_Con()) {
943 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
944 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
945 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
946 } else {
947 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
948 }
949 } else {
950 const TypePtr *tp = obj_node->get_ptr_type();
951 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
952 }
953
954 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
955 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
956 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
957 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
958 }
959
960 // We dump the object pool first, since deoptimization reads it in first.
961 debug_info()->dump_object_pool(objs);
962
963 // Build first class objects to pass to scope
964 DebugToken *locvals = debug_info()->create_scope_values(locarray);
965 DebugToken *expvals = debug_info()->create_scope_values(exparray);
966 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
967
968 // Make method available for all Safepoints
969 ciMethod* scope_method = method ? method : _method;
970 // Describe the scope here
971 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
972 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
973 // Now we can describe the scope.
974 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
975 } // End jvms loop
976
977 // Mark the end of the scope set.
978 debug_info()->end_safepoint(safepoint_pc_offset);
979 }
980
981
982
983 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
984 class NonSafepointEmitter {
985 Compile* C;
986 JVMState* _pending_jvms;
987 int _pending_offset;
988
989 void emit_non_safepoint();
990
991 public:
992 NonSafepointEmitter(Compile* compile) {
993 this->C = compile;
994 _pending_jvms = NULL;
995 _pending_offset = 0;
996 }
997
998 void observe_instruction(Node* n, int pc_offset) {
999 if (!C->debug_info()->recording_non_safepoints()) return;
1000
1001 Node_Notes* nn = C->node_notes_at(n->_idx);
1002 if (nn == NULL || nn->jvms() == NULL) return;
1003 if (_pending_jvms != NULL &&
1004 _pending_jvms->same_calls_as(nn->jvms())) {
1005 // Repeated JVMS? Stretch it up here.
1006 _pending_offset = pc_offset;
1007 } else {
1008 if (_pending_jvms != NULL &&
1009 _pending_offset < pc_offset) {
1010 emit_non_safepoint();
1011 }
1012 _pending_jvms = NULL;
1013 if (pc_offset > C->debug_info()->last_pc_offset()) {
1014 // This is the only way _pending_jvms can become non-NULL:
1015 _pending_jvms = nn->jvms();
1016 _pending_offset = pc_offset;
1017 }
1018 }
1019 }
1020
1021 // Stay out of the way of real safepoints:
1022 void observe_safepoint(JVMState* jvms, int pc_offset) {
1023 if (_pending_jvms != NULL &&
1024 !_pending_jvms->same_calls_as(jvms) &&
1025 _pending_offset < pc_offset) {
1026 emit_non_safepoint();
1027 }
1028 _pending_jvms = NULL;
1029 }
1030
1031 void flush_at_end() {
1032 if (_pending_jvms != NULL) {
1033 emit_non_safepoint();
1034 }
1035 _pending_jvms = NULL;
1036 }
1037 };
1038
1039 void NonSafepointEmitter::emit_non_safepoint() {
1040 JVMState* youngest_jvms = _pending_jvms;
1041 int pc_offset = _pending_offset;
1042
1043 // Clear it now:
1044 _pending_jvms = NULL;
1045
1046 DebugInformationRecorder* debug_info = C->debug_info();
1047 assert(debug_info->recording_non_safepoints(), "sanity");
1048
1049 debug_info->add_non_safepoint(pc_offset);
1050 int max_depth = youngest_jvms->depth();
1051
1052 // Visit scopes from oldest to youngest.
1053 for (int depth = 1; depth <= max_depth; depth++) {
1054 JVMState* jvms = youngest_jvms->of_depth(depth);
1055 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1056 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1057 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1058 }
1059
1060 // Mark the end of the scope set.
1061 debug_info->end_non_safepoint(pc_offset);
1062 }
1063
1064 //------------------------------init_buffer------------------------------------
1065 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1066
1067 // Set the initially allocated size
1068 int code_req = initial_code_capacity;
1069 int locs_req = initial_locs_capacity;
1070 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1071 int const_req = initial_const_capacity;
1072
1073 int pad_req = NativeCall::instruction_size;
1074 // The extra spacing after the code is necessary on some platforms.
1075 // Sometimes we need to patch in a jump after the last instruction,
1076 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1077
1078 // Compute the byte offset where we can store the deopt pc.
1079 if (fixed_slots() != 0) {
1080 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1081 }
1082
1083 // Compute prolog code size
1084 _method_size = 0;
1085 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1086 #if defined(IA64) && !defined(AIX)
1087 if (save_argument_registers()) {
1088 // 4815101: this is a stub with implicit and unknown precision fp args.
1089 // The usual spill mechanism can only generate stfd's in this case, which
1090 // doesn't work if the fp reg to spill contains a single-precision denorm.
1091 // Instead, we hack around the normal spill mechanism using stfspill's and
1092 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1093 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1094 //
1095 // If we ever implement 16-byte 'registers' == stack slots, we can
1096 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1097 // instead of stfd/stfs/ldfd/ldfs.
1098 _frame_slots += 8*(16/BytesPerInt);
1099 }
1100 #endif
1101 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1102
1103 if (has_mach_constant_base_node()) {
1104 uint add_size = 0;
1105 // Fill the constant table.
1106 // Note: This must happen before shorten_branches.
1107 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1108 Block* b = _cfg->get_block(i);
1109
1110 for (uint j = 0; j < b->number_of_nodes(); j++) {
1111 Node* n = b->get_node(j);
1112
1113 // If the node is a MachConstantNode evaluate the constant
1114 // value section.
1115 if (n->is_MachConstant()) {
1116 MachConstantNode* machcon = n->as_MachConstant();
1117 machcon->eval_constant(C);
1118 } else if (n->is_Mach()) {
1119 // On Power there are more nodes that issue constants.
1120 add_size += (n->as_Mach()->ins_num_consts() * 8);
1121 }
1122 }
1123 }
1124
1125 // Calculate the offsets of the constants and the size of the
1126 // constant table (including the padding to the next section).
1127 constant_table().calculate_offsets_and_size();
1128 const_req = constant_table().size() + add_size;
1129 }
1130
1131 // Initialize the space for the BufferBlob used to find and verify
1132 // instruction size in MachNode::emit_size()
1133 init_scratch_buffer_blob(const_req);
1134 if (failing()) return NULL; // Out of memory
1135
1136 // Pre-compute the length of blocks and replace
1137 // long branches with short if machine supports it.
1138 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1139
1140 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1141 // class HandlerImpl is platform-specific and defined in the *.ad files.
1142 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1143 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1144 stub_req += MAX_stubs_size; // ensure per-stub margin
1145 code_req += MAX_inst_size; // ensure per-instruction margin
1146
1147 if (StressCodeBuffers)
1148 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1149
1150 int total_req =
1151 const_req +
1152 code_req +
1153 pad_req +
1154 stub_req +
1155 exception_handler_req +
1156 deopt_handler_req; // deopt handler
1157
1158 if (has_method_handle_invokes())
1159 total_req += deopt_handler_req; // deopt MH handler
1160
1161 CodeBuffer* cb = code_buffer();
1162 cb->initialize(total_req, locs_req);
1163
1164 // Have we run out of code space?
1165 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1166 C->record_failure("CodeCache is full");
1167 CompileBroker::handle_full_code_cache();
1168 return NULL;
1169 }
1170 // Configure the code buffer.
1171 cb->initialize_consts_size(const_req);
1172 cb->initialize_stubs_size(stub_req);
1173 cb->initialize_oop_recorder(env()->oop_recorder());
1174
1175 // fill in the nop array for bundling computations
1176 MachNode *_nop_list[Bundle::_nop_count];
1177 Bundle::initialize_nops(_nop_list, this);
1178
1179 return cb;
1180 }
1181
1182 //------------------------------fill_buffer------------------------------------
1183 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1184 // blk_starts[] contains offsets calculated during short branches processing,
1185 // offsets should not be increased during following steps.
1186
1187 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1188 // of a loop. It is used to determine the padding for loop alignment.
1189 compute_loop_first_inst_sizes();
1190
1191 // Create oopmap set.
1192 _oop_map_set = new OopMapSet();
1193
1194 // !!!!! This preserves old handling of oopmaps for now
1195 debug_info()->set_oopmaps(_oop_map_set);
1196
1197 uint nblocks = _cfg->number_of_blocks();
1198 // Count and start of implicit null check instructions
1199 uint inct_cnt = 0;
1200 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1201
1202 // Count and start of calls
1203 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1204
1205 uint return_offset = 0;
1206 int nop_size = (new MachNopNode())->size(_regalloc);
1207
1208 int previous_offset = 0;
1209 int current_offset = 0;
1210 int last_call_offset = -1;
1211 int last_avoid_back_to_back_offset = -1;
1212 #ifdef ASSERT
1213 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1214 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1215 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1216 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1217 #endif
1218
1219 // Create an array of unused labels, one for each basic block, if printing is enabled
1220 #ifndef PRODUCT
1221 int *node_offsets = NULL;
1222 uint node_offset_limit = unique();
1223
1224 if (print_assembly())
1225 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1226 #endif
1227
1228 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1229
1230 // Emit the constant table.
1231 if (has_mach_constant_base_node()) {
1232 constant_table().emit(*cb);
1233 }
1234
1235 // Create an array of labels, one for each basic block
1236 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1237 for (uint i=0; i <= nblocks; i++) {
1238 blk_labels[i].init();
1239 }
1240
1241 // ------------------
1242 // Now fill in the code buffer
1243 Node *delay_slot = NULL;
1244
1245 for (uint i = 0; i < nblocks; i++) {
1246 Block* block = _cfg->get_block(i);
1247 Node* head = block->head();
1248
1249 // If this block needs to start aligned (i.e, can be reached other
1250 // than by falling-thru from the previous block), then force the
1251 // start of a new bundle.
1252 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1253 cb->flush_bundle(true);
1254 }
1255
1256 #ifdef ASSERT
1257 if (!block->is_connector()) {
1258 stringStream st;
1259 block->dump_head(_cfg, &st);
1260 MacroAssembler(cb).block_comment(st.as_string());
1261 }
1262 jmp_target[i] = 0;
1263 jmp_offset[i] = 0;
1264 jmp_size[i] = 0;
1265 jmp_rule[i] = 0;
1266 #endif
1267 int blk_offset = current_offset;
1268
1269 // Define the label at the beginning of the basic block
1270 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1271
1272 uint last_inst = block->number_of_nodes();
1273
1274 // Emit block normally, except for last instruction.
1275 // Emit means "dump code bits into code buffer".
1276 for (uint j = 0; j<last_inst; j++) {
1277
1278 // Get the node
1279 Node* n = block->get_node(j);
1280
1281 // See if delay slots are supported
1282 if (valid_bundle_info(n) &&
1283 node_bundling(n)->used_in_unconditional_delay()) {
1284 assert(delay_slot == NULL, "no use of delay slot node");
1285 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1286
1287 delay_slot = n;
1288 continue;
1289 }
1290
1291 // If this starts a new instruction group, then flush the current one
1292 // (but allow split bundles)
1293 if (Pipeline::requires_bundling() && starts_bundle(n))
1294 cb->flush_bundle(false);
1295
1296 // The following logic is duplicated in the code ifdeffed for
1297 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1298 // should be factored out. Or maybe dispersed to the nodes?
1299
1300 // Special handling for SafePoint/Call Nodes
1301 bool is_mcall = false;
1302 if (n->is_Mach()) {
1303 MachNode *mach = n->as_Mach();
1304 is_mcall = n->is_MachCall();
1305 bool is_sfn = n->is_MachSafePoint();
1306
1307 // If this requires all previous instructions be flushed, then do so
1308 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1309 cb->flush_bundle(true);
1310 current_offset = cb->insts_size();
1311 }
1312
1313 // A padding may be needed again since a previous instruction
1314 // could be moved to delay slot.
1315
1316 // align the instruction if necessary
1317 int padding = mach->compute_padding(current_offset);
1318 // Make sure safepoint node for polling is distinct from a call's
1319 // return by adding a nop if needed.
1320 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1321 padding = nop_size;
1322 }
1323 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1324 current_offset == last_avoid_back_to_back_offset) {
1325 // Avoid back to back some instructions.
1326 padding = nop_size;
1327 }
1328
1329 if(padding > 0) {
1330 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1331 int nops_cnt = padding / nop_size;
1332 MachNode *nop = new MachNopNode(nops_cnt);
1333 block->insert_node(nop, j++);
1334 last_inst++;
1335 _cfg->map_node_to_block(nop, block);
1336 nop->emit(*cb, _regalloc);
1337 cb->flush_bundle(true);
1338 current_offset = cb->insts_size();
1339 }
1340
1341 // Remember the start of the last call in a basic block
1342 if (is_mcall) {
1343 MachCallNode *mcall = mach->as_MachCall();
1344
1345 // This destination address is NOT PC-relative
1346 mcall->method_set((intptr_t)mcall->entry_point());
1347
1348 // Save the return address
1349 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1350
1351 if (mcall->is_MachCallLeaf()) {
1352 is_mcall = false;
1353 is_sfn = false;
1354 }
1355 }
1356
1357 // sfn will be valid whenever mcall is valid now because of inheritance
1358 if (is_sfn || is_mcall) {
1359
1360 // Handle special safepoint nodes for synchronization
1361 if (!is_mcall) {
1362 MachSafePointNode *sfn = mach->as_MachSafePoint();
1363 // !!!!! Stubs only need an oopmap right now, so bail out
1364 if (sfn->jvms()->method() == NULL) {
1365 // Write the oopmap directly to the code blob??!!
1366 # ifdef ENABLE_ZAP_DEAD_LOCALS
1367 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1368 # endif
1369 continue;
1370 }
1371 } // End synchronization
1372
1373 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1374 current_offset);
1375 Process_OopMap_Node(mach, current_offset);
1376 } // End if safepoint
1377
1378 // If this is a null check, then add the start of the previous instruction to the list
1379 else if( mach->is_MachNullCheck() ) {
1380 inct_starts[inct_cnt++] = previous_offset;
1381 }
1382
1383 // If this is a branch, then fill in the label with the target BB's label
1384 else if (mach->is_MachBranch()) {
1385 // This requires the TRUE branch target be in succs[0]
1386 uint block_num = block->non_connector_successor(0)->_pre_order;
1387
1388 // Try to replace long branch if delay slot is not used,
1389 // it is mostly for back branches since forward branch's
1390 // distance is not updated yet.
1391 bool delay_slot_is_used = valid_bundle_info(n) &&
1392 node_bundling(n)->use_unconditional_delay();
1393 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1394 assert(delay_slot == NULL, "not expecting delay slot node");
1395 int br_size = n->size(_regalloc);
1396 int offset = blk_starts[block_num] - current_offset;
1397 if (block_num >= i) {
1398 // Current and following block's offset are not
1399 // finalized yet, adjust distance by the difference
1400 // between calculated and final offsets of current block.
1401 offset -= (blk_starts[i] - blk_offset);
1402 }
1403 // In the following code a nop could be inserted before
1404 // the branch which will increase the backward distance.
1405 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1406 if (needs_padding && offset <= 0)
1407 offset -= nop_size;
1408
1409 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1410 // We've got a winner. Replace this branch.
1411 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1412
1413 // Update the jmp_size.
1414 int new_size = replacement->size(_regalloc);
1415 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1416 // Insert padding between avoid_back_to_back branches.
1417 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1418 MachNode *nop = new MachNopNode();
1419 block->insert_node(nop, j++);
1420 _cfg->map_node_to_block(nop, block);
1421 last_inst++;
1422 nop->emit(*cb, _regalloc);
1423 cb->flush_bundle(true);
1424 current_offset = cb->insts_size();
1425 }
1426 #ifdef ASSERT
1427 jmp_target[i] = block_num;
1428 jmp_offset[i] = current_offset - blk_offset;
1429 jmp_size[i] = new_size;
1430 jmp_rule[i] = mach->rule();
1431 #endif
1432 block->map_node(replacement, j);
1433 mach->subsume_by(replacement, C);
1434 n = replacement;
1435 mach = replacement;
1436 }
1437 }
1438 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1439 } else if (mach->ideal_Opcode() == Op_Jump) {
1440 for (uint h = 0; h < block->_num_succs; h++) {
1441 Block* succs_block = block->_succs[h];
1442 for (uint j = 1; j < succs_block->num_preds(); j++) {
1443 Node* jpn = succs_block->pred(j);
1444 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1445 uint block_num = succs_block->non_connector()->_pre_order;
1446 Label *blkLabel = &blk_labels[block_num];
1447 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1448 }
1449 }
1450 }
1451 }
1452 #ifdef ASSERT
1453 // Check that oop-store precedes the card-mark
1454 else if (mach->ideal_Opcode() == Op_StoreCM) {
1455 uint storeCM_idx = j;
1456 int count = 0;
1457 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1458 Node *oop_store = mach->in(prec); // Precedence edge
1459 if (oop_store == NULL) continue;
1460 count++;
1461 uint i4;
1462 for (i4 = 0; i4 < last_inst; ++i4) {
1463 if (block->get_node(i4) == oop_store) {
1464 break;
1465 }
1466 }
1467 // Note: This test can provide a false failure if other precedence
1468 // edges have been added to the storeCMNode.
1469 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1470 }
1471 assert(count > 0, "storeCM expects at least one precedence edge");
1472 }
1473 #endif
1474 else if (!n->is_Proj()) {
1475 // Remember the beginning of the previous instruction, in case
1476 // it's followed by a flag-kill and a null-check. Happens on
1477 // Intel all the time, with add-to-memory kind of opcodes.
1478 previous_offset = current_offset;
1479 }
1480
1481 // Not an else-if!
1482 // If this is a trap based cmp then add its offset to the list.
1483 if (mach->is_TrapBasedCheckNode()) {
1484 inct_starts[inct_cnt++] = current_offset;
1485 }
1486 }
1487
1488 // Verify that there is sufficient space remaining
1489 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1490 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1491 C->record_failure("CodeCache is full");
1492 CompileBroker::handle_full_code_cache();
1493 return;
1494 }
1495
1496 // Save the offset for the listing
1497 #ifndef PRODUCT
1498 if (node_offsets && n->_idx < node_offset_limit)
1499 node_offsets[n->_idx] = cb->insts_size();
1500 #endif
1501
1502 // "Normal" instruction case
1503 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1504 n->emit(*cb, _regalloc);
1505 current_offset = cb->insts_size();
1506
1507 #ifdef ASSERT
1508 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1509 n->dump();
1510 assert(false, "wrong size of mach node");
1511 }
1512 #endif
1513 non_safepoints.observe_instruction(n, current_offset);
1514
1515 // mcall is last "call" that can be a safepoint
1516 // record it so we can see if a poll will directly follow it
1517 // in which case we'll need a pad to make the PcDesc sites unique
1518 // see 5010568. This can be slightly inaccurate but conservative
1519 // in the case that return address is not actually at current_offset.
1520 // This is a small price to pay.
1521
1522 if (is_mcall) {
1523 last_call_offset = current_offset;
1524 }
1525
1526 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1527 // Avoid back to back some instructions.
1528 last_avoid_back_to_back_offset = current_offset;
1529 }
1530
1531 // See if this instruction has a delay slot
1532 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1533 assert(delay_slot != NULL, "expecting delay slot node");
1534
1535 // Back up 1 instruction
1536 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1537
1538 // Save the offset for the listing
1539 #ifndef PRODUCT
1540 if (node_offsets && delay_slot->_idx < node_offset_limit)
1541 node_offsets[delay_slot->_idx] = cb->insts_size();
1542 #endif
1543
1544 // Support a SafePoint in the delay slot
1545 if (delay_slot->is_MachSafePoint()) {
1546 MachNode *mach = delay_slot->as_Mach();
1547 // !!!!! Stubs only need an oopmap right now, so bail out
1548 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1549 // Write the oopmap directly to the code blob??!!
1550 # ifdef ENABLE_ZAP_DEAD_LOCALS
1551 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1552 # endif
1553 delay_slot = NULL;
1554 continue;
1555 }
1556
1557 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1558 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1559 adjusted_offset);
1560 // Generate an OopMap entry
1561 Process_OopMap_Node(mach, adjusted_offset);
1562 }
1563
1564 // Insert the delay slot instruction
1565 delay_slot->emit(*cb, _regalloc);
1566
1567 // Don't reuse it
1568 delay_slot = NULL;
1569 }
1570
1571 } // End for all instructions in block
1572
1573 // If the next block is the top of a loop, pad this block out to align
1574 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1575 if (i < nblocks-1) {
1576 Block *nb = _cfg->get_block(i + 1);
1577 int padding = nb->alignment_padding(current_offset);
1578 if( padding > 0 ) {
1579 MachNode *nop = new MachNopNode(padding / nop_size);
1580 block->insert_node(nop, block->number_of_nodes());
1581 _cfg->map_node_to_block(nop, block);
1582 nop->emit(*cb, _regalloc);
1583 current_offset = cb->insts_size();
1584 }
1585 }
1586 // Verify that the distance for generated before forward
1587 // short branches is still valid.
1588 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1589
1590 // Save new block start offset
1591 blk_starts[i] = blk_offset;
1592 } // End of for all blocks
1593 blk_starts[nblocks] = current_offset;
1594
1595 non_safepoints.flush_at_end();
1596
1597 // Offset too large?
1598 if (failing()) return;
1599
1600 // Define a pseudo-label at the end of the code
1601 MacroAssembler(cb).bind( blk_labels[nblocks] );
1602
1603 // Compute the size of the first block
1604 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1605
1606 assert(cb->insts_size() < 500000, "method is unreasonably large");
1607
1608 #ifdef ASSERT
1609 for (uint i = 0; i < nblocks; i++) { // For all blocks
1610 if (jmp_target[i] != 0) {
1611 int br_size = jmp_size[i];
1612 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1613 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1614 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]);
1615 assert(false, "Displacement too large for short jmp");
1616 }
1617 }
1618 }
1619 #endif
1620
1621 #ifndef PRODUCT
1622 // Information on the size of the method, without the extraneous code
1623 Scheduling::increment_method_size(cb->insts_size());
1624 #endif
1625
1626 // ------------------
1627 // Fill in exception table entries.
1628 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1629
1630 // Only java methods have exception handlers and deopt handlers
1631 // class HandlerImpl is platform-specific and defined in the *.ad files.
1632 if (_method) {
1633 // Emit the exception handler code.
1634 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1635 // Emit the deopt handler code.
1636 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1637
1638 // Emit the MethodHandle deopt handler code (if required).
1639 if (has_method_handle_invokes()) {
1640 // We can use the same code as for the normal deopt handler, we
1641 // just need a different entry point address.
1642 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1643 }
1644 }
1645
1646 // One last check for failed CodeBuffer::expand:
1647 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1648 C->record_failure("CodeCache is full");
1649 CompileBroker::handle_full_code_cache();
1650 return;
1651 }
1652
1653 #ifndef PRODUCT
1654 // Dump the assembly code, including basic-block numbers
1655 if (print_assembly()) {
1656 ttyLocker ttyl; // keep the following output all in one block
1657 if (!VMThread::should_terminate()) { // test this under the tty lock
1658 // This output goes directly to the tty, not the compiler log.
1659 // To enable tools to match it up with the compilation activity,
1660 // be sure to tag this tty output with the compile ID.
1661 if (xtty != NULL) {
1662 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1663 is_osr_compilation() ? " compile_kind='osr'" :
1664 "");
1665 }
1666 if (method() != NULL) {
1667 method()->print_metadata();
1668 }
1669 dump_asm(node_offsets, node_offset_limit);
1670 if (xtty != NULL) {
1671 xtty->tail("opto_assembly");
1672 }
1673 }
1674 }
1675 #endif
1676
1677 }
1678
1679 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1680 _inc_table.set_size(cnt);
1681
1682 uint inct_cnt = 0;
1683 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1684 Block* block = _cfg->get_block(i);
1685 Node *n = NULL;
1686 int j;
1687
1688 // Find the branch; ignore trailing NOPs.
1689 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1690 n = block->get_node(j);
1691 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1692 break;
1693 }
1694 }
1695
1696 // If we didn't find anything, continue
1697 if (j < 0) {
1698 continue;
1699 }
1700
1701 // Compute ExceptionHandlerTable subtable entry and add it
1702 // (skip empty blocks)
1703 if (n->is_Catch()) {
1704
1705 // Get the offset of the return from the call
1706 uint call_return = call_returns[block->_pre_order];
1707 #ifdef ASSERT
1708 assert( call_return > 0, "no call seen for this basic block" );
1709 while (block->get_node(--j)->is_MachProj()) ;
1710 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1711 #endif
1712 // last instruction is a CatchNode, find it's CatchProjNodes
1713 int nof_succs = block->_num_succs;
1714 // allocate space
1715 GrowableArray<intptr_t> handler_bcis(nof_succs);
1716 GrowableArray<intptr_t> handler_pcos(nof_succs);
1717 // iterate through all successors
1718 for (int j = 0; j < nof_succs; j++) {
1719 Block* s = block->_succs[j];
1720 bool found_p = false;
1721 for (uint k = 1; k < s->num_preds(); k++) {
1722 Node* pk = s->pred(k);
1723 if (pk->is_CatchProj() && pk->in(0) == n) {
1724 const CatchProjNode* p = pk->as_CatchProj();
1725 found_p = true;
1726 // add the corresponding handler bci & pco information
1727 if (p->_con != CatchProjNode::fall_through_index) {
1728 // p leads to an exception handler (and is not fall through)
1729 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1730 // no duplicates, please
1731 if (!handler_bcis.contains(p->handler_bci())) {
1732 uint block_num = s->non_connector()->_pre_order;
1733 handler_bcis.append(p->handler_bci());
1734 handler_pcos.append(blk_labels[block_num].loc_pos());
1735 }
1736 }
1737 }
1738 }
1739 assert(found_p, "no matching predecessor found");
1740 // Note: Due to empty block removal, one block may have
1741 // several CatchProj inputs, from the same Catch.
1742 }
1743
1744 // Set the offset of the return from the call
1745 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1746 continue;
1747 }
1748
1749 // Handle implicit null exception table updates
1750 if (n->is_MachNullCheck()) {
1751 uint block_num = block->non_connector_successor(0)->_pre_order;
1752 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1753 continue;
1754 }
1755 // Handle implicit exception table updates: trap instructions.
1756 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1757 uint block_num = block->non_connector_successor(0)->_pre_order;
1758 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1759 continue;
1760 }
1761 } // End of for all blocks fill in exception table entries
1762 }
1763
1764 // Static Variables
1765 #ifndef PRODUCT
1766 uint Scheduling::_total_nop_size = 0;
1767 uint Scheduling::_total_method_size = 0;
1768 uint Scheduling::_total_branches = 0;
1769 uint Scheduling::_total_unconditional_delays = 0;
1770 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1771 #endif
1772
1773 // Initializer for class Scheduling
1774
1775 Scheduling::Scheduling(Arena *arena, Compile &compile)
1776 : _arena(arena),
1777 _cfg(compile.cfg()),
1778 _regalloc(compile.regalloc()),
1779 _reg_node(arena),
1780 _bundle_instr_count(0),
1781 _bundle_cycle_number(0),
1782 _scheduled(arena),
1783 _available(arena),
1784 _next_node(NULL),
1785 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1786 _pinch_free_list(arena)
1787 #ifndef PRODUCT
1788 , _branches(0)
1789 , _unconditional_delays(0)
1790 #endif
1791 {
1792 // Create a MachNopNode
1793 _nop = new MachNopNode();
1794
1795 // Now that the nops are in the array, save the count
1796 // (but allow entries for the nops)
1797 _node_bundling_limit = compile.unique();
1798 uint node_max = _regalloc->node_regs_max_index();
1799
1800 compile.set_node_bundling_limit(_node_bundling_limit);
1801
1802 // This one is persistent within the Compile class
1803 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1804
1805 // Allocate space for fixed-size arrays
1806 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1807 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1808 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1809
1810 // Clear the arrays
1811 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1812 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1813 memset(_uses, 0, node_max * sizeof(short));
1814 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1815
1816 // Clear the bundling information
1817 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1818
1819 // Get the last node
1820 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1821
1822 _next_node = block->get_node(block->number_of_nodes() - 1);
1823 }
1824
1825 #ifndef PRODUCT
1826 // Scheduling destructor
1827 Scheduling::~Scheduling() {
1828 _total_branches += _branches;
1829 _total_unconditional_delays += _unconditional_delays;
1830 }
1831 #endif
1832
1833 // Step ahead "i" cycles
1834 void Scheduling::step(uint i) {
1835
1836 Bundle *bundle = node_bundling(_next_node);
1837 bundle->set_starts_bundle();
1838
1839 // Update the bundle record, but leave the flags information alone
1840 if (_bundle_instr_count > 0) {
1841 bundle->set_instr_count(_bundle_instr_count);
1842 bundle->set_resources_used(_bundle_use.resourcesUsed());
1843 }
1844
1845 // Update the state information
1846 _bundle_instr_count = 0;
1847 _bundle_cycle_number += i;
1848 _bundle_use.step(i);
1849 }
1850
1851 void Scheduling::step_and_clear() {
1852 Bundle *bundle = node_bundling(_next_node);
1853 bundle->set_starts_bundle();
1854
1855 // Update the bundle record
1856 if (_bundle_instr_count > 0) {
1857 bundle->set_instr_count(_bundle_instr_count);
1858 bundle->set_resources_used(_bundle_use.resourcesUsed());
1859
1860 _bundle_cycle_number += 1;
1861 }
1862
1863 // Clear the bundling information
1864 _bundle_instr_count = 0;
1865 _bundle_use.reset();
1866
1867 memcpy(_bundle_use_elements,
1868 Pipeline_Use::elaborated_elements,
1869 sizeof(Pipeline_Use::elaborated_elements));
1870 }
1871
1872 // Perform instruction scheduling and bundling over the sequence of
1873 // instructions in backwards order.
1874 void Compile::ScheduleAndBundle() {
1875
1876 // Don't optimize this if it isn't a method
1877 if (!_method)
1878 return;
1879
1880 // Don't optimize this if scheduling is disabled
1881 if (!do_scheduling())
1882 return;
1883
1884 // Scheduling code works only with pairs (8 bytes) maximum.
1885 if (max_vector_size() > 8)
1886 return;
1887
1888 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1889
1890 // Create a data structure for all the scheduling information
1891 Scheduling scheduling(Thread::current()->resource_area(), *this);
1892
1893 // Walk backwards over each basic block, computing the needed alignment
1894 // Walk over all the basic blocks
1895 scheduling.DoScheduling();
1896 }
1897
1898 // Compute the latency of all the instructions. This is fairly simple,
1899 // because we already have a legal ordering. Walk over the instructions
1900 // from first to last, and compute the latency of the instruction based
1901 // on the latency of the preceding instruction(s).
1902 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1903 #ifndef PRODUCT
1904 if (_cfg->C->trace_opto_output())
1905 tty->print("# -> ComputeLocalLatenciesForward\n");
1906 #endif
1907
1908 // Walk over all the schedulable instructions
1909 for( uint j=_bb_start; j < _bb_end; j++ ) {
1910
1911 // This is a kludge, forcing all latency calculations to start at 1.
1912 // Used to allow latency 0 to force an instruction to the beginning
1913 // of the bb
1914 uint latency = 1;
1915 Node *use = bb->get_node(j);
1916 uint nlen = use->len();
1917
1918 // Walk over all the inputs
1919 for ( uint k=0; k < nlen; k++ ) {
1920 Node *def = use->in(k);
1921 if (!def)
1922 continue;
1923
1924 uint l = _node_latency[def->_idx] + use->latency(k);
1925 if (latency < l)
1926 latency = l;
1927 }
1928
1929 _node_latency[use->_idx] = latency;
1930
1931 #ifndef PRODUCT
1932 if (_cfg->C->trace_opto_output()) {
1933 tty->print("# latency %4d: ", latency);
1934 use->dump();
1935 }
1936 #endif
1937 }
1938
1939 #ifndef PRODUCT
1940 if (_cfg->C->trace_opto_output())
1941 tty->print("# <- ComputeLocalLatenciesForward\n");
1942 #endif
1943
1944 } // end ComputeLocalLatenciesForward
1945
1946 // See if this node fits into the present instruction bundle
1947 bool Scheduling::NodeFitsInBundle(Node *n) {
1948 uint n_idx = n->_idx;
1949
1950 // If this is the unconditional delay instruction, then it fits
1951 if (n == _unconditional_delay_slot) {
1952 #ifndef PRODUCT
1953 if (_cfg->C->trace_opto_output())
1954 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1955 #endif
1956 return (true);
1957 }
1958
1959 // If the node cannot be scheduled this cycle, skip it
1960 if (_current_latency[n_idx] > _bundle_cycle_number) {
1961 #ifndef PRODUCT
1962 if (_cfg->C->trace_opto_output())
1963 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1964 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1965 #endif
1966 return (false);
1967 }
1968
1969 const Pipeline *node_pipeline = n->pipeline();
1970
1971 uint instruction_count = node_pipeline->instructionCount();
1972 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1973 instruction_count = 0;
1974 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1975 instruction_count++;
1976
1977 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1978 #ifndef PRODUCT
1979 if (_cfg->C->trace_opto_output())
1980 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1981 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1982 #endif
1983 return (false);
1984 }
1985
1986 // Don't allow non-machine nodes to be handled this way
1987 if (!n->is_Mach() && instruction_count == 0)
1988 return (false);
1989
1990 // See if there is any overlap
1991 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1992
1993 if (delay > 0) {
1994 #ifndef PRODUCT
1995 if (_cfg->C->trace_opto_output())
1996 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1997 #endif
1998 return false;
1999 }
2000
2001 #ifndef PRODUCT
2002 if (_cfg->C->trace_opto_output())
2003 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2004 #endif
2005
2006 return true;
2007 }
2008
2009 Node * Scheduling::ChooseNodeToBundle() {
2010 uint siz = _available.size();
2011
2012 if (siz == 0) {
2013
2014 #ifndef PRODUCT
2015 if (_cfg->C->trace_opto_output())
2016 tty->print("# ChooseNodeToBundle: NULL\n");
2017 #endif
2018 return (NULL);
2019 }
2020
2021 // Fast path, if only 1 instruction in the bundle
2022 if (siz == 1) {
2023 #ifndef PRODUCT
2024 if (_cfg->C->trace_opto_output()) {
2025 tty->print("# ChooseNodeToBundle (only 1): ");
2026 _available[0]->dump();
2027 }
2028 #endif
2029 return (_available[0]);
2030 }
2031
2032 // Don't bother, if the bundle is already full
2033 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2034 for ( uint i = 0; i < siz; i++ ) {
2035 Node *n = _available[i];
2036
2037 // Skip projections, we'll handle them another way
2038 if (n->is_Proj())
2039 continue;
2040
2041 // This presupposed that instructions are inserted into the
2042 // available list in a legality order; i.e. instructions that
2043 // must be inserted first are at the head of the list
2044 if (NodeFitsInBundle(n)) {
2045 #ifndef PRODUCT
2046 if (_cfg->C->trace_opto_output()) {
2047 tty->print("# ChooseNodeToBundle: ");
2048 n->dump();
2049 }
2050 #endif
2051 return (n);
2052 }
2053 }
2054 }
2055
2056 // Nothing fits in this bundle, choose the highest priority
2057 #ifndef PRODUCT
2058 if (_cfg->C->trace_opto_output()) {
2059 tty->print("# ChooseNodeToBundle: ");
2060 _available[0]->dump();
2061 }
2062 #endif
2063
2064 return _available[0];
2065 }
2066
2067 void Scheduling::AddNodeToAvailableList(Node *n) {
2068 assert( !n->is_Proj(), "projections never directly made available" );
2069 #ifndef PRODUCT
2070 if (_cfg->C->trace_opto_output()) {
2071 tty->print("# AddNodeToAvailableList: ");
2072 n->dump();
2073 }
2074 #endif
2075
2076 int latency = _current_latency[n->_idx];
2077
2078 // Insert in latency order (insertion sort)
2079 uint i;
2080 for ( i=0; i < _available.size(); i++ )
2081 if (_current_latency[_available[i]->_idx] > latency)
2082 break;
2083
2084 // Special Check for compares following branches
2085 if( n->is_Mach() && _scheduled.size() > 0 ) {
2086 int op = n->as_Mach()->ideal_Opcode();
2087 Node *last = _scheduled[0];
2088 if( last->is_MachIf() && last->in(1) == n &&
2089 ( op == Op_CmpI ||
2090 op == Op_CmpU ||
2091 op == Op_CmpP ||
2092 op == Op_CmpF ||
2093 op == Op_CmpD ||
2094 op == Op_CmpL ) ) {
2095
2096 // Recalculate position, moving to front of same latency
2097 for ( i=0 ; i < _available.size(); i++ )
2098 if (_current_latency[_available[i]->_idx] >= latency)
2099 break;
2100 }
2101 }
2102
2103 // Insert the node in the available list
2104 _available.insert(i, n);
2105
2106 #ifndef PRODUCT
2107 if (_cfg->C->trace_opto_output())
2108 dump_available();
2109 #endif
2110 }
2111
2112 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2113 for ( uint i=0; i < n->len(); i++ ) {
2114 Node *def = n->in(i);
2115 if (!def) continue;
2116 if( def->is_Proj() ) // If this is a machine projection, then
2117 def = def->in(0); // propagate usage thru to the base instruction
2118
2119 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2120 continue;
2121 }
2122
2123 // Compute the latency
2124 uint l = _bundle_cycle_number + n->latency(i);
2125 if (_current_latency[def->_idx] < l)
2126 _current_latency[def->_idx] = l;
2127
2128 // If this does not have uses then schedule it
2129 if ((--_uses[def->_idx]) == 0)
2130 AddNodeToAvailableList(def);
2131 }
2132 }
2133
2134 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2135 #ifndef PRODUCT
2136 if (_cfg->C->trace_opto_output()) {
2137 tty->print("# AddNodeToBundle: ");
2138 n->dump();
2139 }
2140 #endif
2141
2142 // Remove this from the available list
2143 uint i;
2144 for (i = 0; i < _available.size(); i++)
2145 if (_available[i] == n)
2146 break;
2147 assert(i < _available.size(), "entry in _available list not found");
2148 _available.remove(i);
2149
2150 // See if this fits in the current bundle
2151 const Pipeline *node_pipeline = n->pipeline();
2152 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2153
2154 // Check for instructions to be placed in the delay slot. We
2155 // do this before we actually schedule the current instruction,
2156 // because the delay slot follows the current instruction.
2157 if (Pipeline::_branch_has_delay_slot &&
2158 node_pipeline->hasBranchDelay() &&
2159 !_unconditional_delay_slot) {
2160
2161 uint siz = _available.size();
2162
2163 // Conditional branches can support an instruction that
2164 // is unconditionally executed and not dependent by the
2165 // branch, OR a conditionally executed instruction if
2166 // the branch is taken. In practice, this means that
2167 // the first instruction at the branch target is
2168 // copied to the delay slot, and the branch goes to
2169 // the instruction after that at the branch target
2170 if ( n->is_MachBranch() ) {
2171
2172 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2173 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2174
2175 #ifndef PRODUCT
2176 _branches++;
2177 #endif
2178
2179 // At least 1 instruction is on the available list
2180 // that is not dependent on the branch
2181 for (uint i = 0; i < siz; i++) {
2182 Node *d = _available[i];
2183 const Pipeline *avail_pipeline = d->pipeline();
2184
2185 // Don't allow safepoints in the branch shadow, that will
2186 // cause a number of difficulties
2187 if ( avail_pipeline->instructionCount() == 1 &&
2188 !avail_pipeline->hasMultipleBundles() &&
2189 !avail_pipeline->hasBranchDelay() &&
2190 Pipeline::instr_has_unit_size() &&
2191 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2192 NodeFitsInBundle(d) &&
2193 !node_bundling(d)->used_in_delay()) {
2194
2195 if (d->is_Mach() && !d->is_MachSafePoint()) {
2196 // A node that fits in the delay slot was found, so we need to
2197 // set the appropriate bits in the bundle pipeline information so
2198 // that it correctly indicates resource usage. Later, when we
2199 // attempt to add this instruction to the bundle, we will skip
2200 // setting the resource usage.
2201 _unconditional_delay_slot = d;
2202 node_bundling(n)->set_use_unconditional_delay();
2203 node_bundling(d)->set_used_in_unconditional_delay();
2204 _bundle_use.add_usage(avail_pipeline->resourceUse());
2205 _current_latency[d->_idx] = _bundle_cycle_number;
2206 _next_node = d;
2207 ++_bundle_instr_count;
2208 #ifndef PRODUCT
2209 _unconditional_delays++;
2210 #endif
2211 break;
2212 }
2213 }
2214 }
2215 }
2216
2217 // No delay slot, add a nop to the usage
2218 if (!_unconditional_delay_slot) {
2219 // See if adding an instruction in the delay slot will overflow
2220 // the bundle.
2221 if (!NodeFitsInBundle(_nop)) {
2222 #ifndef PRODUCT
2223 if (_cfg->C->trace_opto_output())
2224 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2225 #endif
2226 step(1);
2227 }
2228
2229 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2230 _next_node = _nop;
2231 ++_bundle_instr_count;
2232 }
2233
2234 // See if the instruction in the delay slot requires a
2235 // step of the bundles
2236 if (!NodeFitsInBundle(n)) {
2237 #ifndef PRODUCT
2238 if (_cfg->C->trace_opto_output())
2239 tty->print("# *** STEP(branch won't fit) ***\n");
2240 #endif
2241 // Update the state information
2242 _bundle_instr_count = 0;
2243 _bundle_cycle_number += 1;
2244 _bundle_use.step(1);
2245 }
2246 }
2247
2248 // Get the number of instructions
2249 uint instruction_count = node_pipeline->instructionCount();
2250 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2251 instruction_count = 0;
2252
2253 // Compute the latency information
2254 uint delay = 0;
2255
2256 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2257 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2258 if (relative_latency < 0)
2259 relative_latency = 0;
2260
2261 delay = _bundle_use.full_latency(relative_latency, node_usage);
2262
2263 // Does not fit in this bundle, start a new one
2264 if (delay > 0) {
2265 step(delay);
2266
2267 #ifndef PRODUCT
2268 if (_cfg->C->trace_opto_output())
2269 tty->print("# *** STEP(%d) ***\n", delay);
2270 #endif
2271 }
2272 }
2273
2274 // If this was placed in the delay slot, ignore it
2275 if (n != _unconditional_delay_slot) {
2276
2277 if (delay == 0) {
2278 if (node_pipeline->hasMultipleBundles()) {
2279 #ifndef PRODUCT
2280 if (_cfg->C->trace_opto_output())
2281 tty->print("# *** STEP(multiple instructions) ***\n");
2282 #endif
2283 step(1);
2284 }
2285
2286 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2287 #ifndef PRODUCT
2288 if (_cfg->C->trace_opto_output())
2289 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2290 instruction_count + _bundle_instr_count,
2291 Pipeline::_max_instrs_per_cycle);
2292 #endif
2293 step(1);
2294 }
2295 }
2296
2297 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2298 _bundle_instr_count++;
2299
2300 // Set the node's latency
2301 _current_latency[n->_idx] = _bundle_cycle_number;
2302
2303 // Now merge the functional unit information
2304 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2305 _bundle_use.add_usage(node_usage);
2306
2307 // Increment the number of instructions in this bundle
2308 _bundle_instr_count += instruction_count;
2309
2310 // Remember this node for later
2311 if (n->is_Mach())
2312 _next_node = n;
2313 }
2314
2315 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2316 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2317 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2318 // into the block. All other scheduled nodes get put in the schedule here.
2319 int op = n->Opcode();
2320 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2321 (op != Op_Node && // Not an unused antidepedence node and
2322 // not an unallocated boxlock
2323 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2324
2325 // Push any trailing projections
2326 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2327 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2328 Node *foi = n->fast_out(i);
2329 if( foi->is_Proj() )
2330 _scheduled.push(foi);
2331 }
2332 }
2333
2334 // Put the instruction in the schedule list
2335 _scheduled.push(n);
2336 }
2337
2338 #ifndef PRODUCT
2339 if (_cfg->C->trace_opto_output())
2340 dump_available();
2341 #endif
2342
2343 // Walk all the definitions, decrementing use counts, and
2344 // if a definition has a 0 use count, place it in the available list.
2345 DecrementUseCounts(n,bb);
2346 }
2347
2348 // This method sets the use count within a basic block. We will ignore all
2349 // uses outside the current basic block. As we are doing a backwards walk,
2350 // any node we reach that has a use count of 0 may be scheduled. This also
2351 // avoids the problem of cyclic references from phi nodes, as long as phi
2352 // nodes are at the front of the basic block. This method also initializes
2353 // the available list to the set of instructions that have no uses within this
2354 // basic block.
2355 void Scheduling::ComputeUseCount(const Block *bb) {
2356 #ifndef PRODUCT
2357 if (_cfg->C->trace_opto_output())
2358 tty->print("# -> ComputeUseCount\n");
2359 #endif
2360
2361 // Clear the list of available and scheduled instructions, just in case
2362 _available.clear();
2363 _scheduled.clear();
2364
2365 // No delay slot specified
2366 _unconditional_delay_slot = NULL;
2367
2368 #ifdef ASSERT
2369 for( uint i=0; i < bb->number_of_nodes(); i++ )
2370 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2371 #endif
2372
2373 // Force the _uses count to never go to zero for unscheduable pieces
2374 // of the block
2375 for( uint k = 0; k < _bb_start; k++ )
2376 _uses[bb->get_node(k)->_idx] = 1;
2377 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2378 _uses[bb->get_node(l)->_idx] = 1;
2379
2380 // Iterate backwards over the instructions in the block. Don't count the
2381 // branch projections at end or the block header instructions.
2382 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2383 Node *n = bb->get_node(j);
2384 if( n->is_Proj() ) continue; // Projections handled another way
2385
2386 // Account for all uses
2387 for ( uint k = 0; k < n->len(); k++ ) {
2388 Node *inp = n->in(k);
2389 if (!inp) continue;
2390 assert(inp != n, "no cycles allowed" );
2391 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2392 if (inp->is_Proj()) { // Skip through Proj's
2393 inp = inp->in(0);
2394 }
2395 ++_uses[inp->_idx]; // Count 1 block-local use
2396 }
2397 }
2398
2399 // If this instruction has a 0 use count, then it is available
2400 if (!_uses[n->_idx]) {
2401 _current_latency[n->_idx] = _bundle_cycle_number;
2402 AddNodeToAvailableList(n);
2403 }
2404
2405 #ifndef PRODUCT
2406 if (_cfg->C->trace_opto_output()) {
2407 tty->print("# uses: %3d: ", _uses[n->_idx]);
2408 n->dump();
2409 }
2410 #endif
2411 }
2412
2413 #ifndef PRODUCT
2414 if (_cfg->C->trace_opto_output())
2415 tty->print("# <- ComputeUseCount\n");
2416 #endif
2417 }
2418
2419 // This routine performs scheduling on each basic block in reverse order,
2420 // using instruction latencies and taking into account function unit
2421 // availability.
2422 void Scheduling::DoScheduling() {
2423 #ifndef PRODUCT
2424 if (_cfg->C->trace_opto_output())
2425 tty->print("# -> DoScheduling\n");
2426 #endif
2427
2428 Block *succ_bb = NULL;
2429 Block *bb;
2430
2431 // Walk over all the basic blocks in reverse order
2432 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2433 bb = _cfg->get_block(i);
2434
2435 #ifndef PRODUCT
2436 if (_cfg->C->trace_opto_output()) {
2437 tty->print("# Schedule BB#%03d (initial)\n", i);
2438 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2439 bb->get_node(j)->dump();
2440 }
2441 }
2442 #endif
2443
2444 // On the head node, skip processing
2445 if (bb == _cfg->get_root_block()) {
2446 continue;
2447 }
2448
2449 // Skip empty, connector blocks
2450 if (bb->is_connector())
2451 continue;
2452
2453 // If the following block is not the sole successor of
2454 // this one, then reset the pipeline information
2455 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2456 #ifndef PRODUCT
2457 if (_cfg->C->trace_opto_output()) {
2458 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2459 _next_node->_idx, _bundle_instr_count);
2460 }
2461 #endif
2462 step_and_clear();
2463 }
2464
2465 // Leave untouched the starting instruction, any Phis, a CreateEx node
2466 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2467 _bb_end = bb->number_of_nodes()-1;
2468 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2469 Node *n = bb->get_node(_bb_start);
2470 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2471 // Also, MachIdealNodes do not get scheduled
2472 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2473 MachNode *mach = n->as_Mach();
2474 int iop = mach->ideal_Opcode();
2475 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2476 if( iop == Op_Con ) continue; // Do not schedule Top
2477 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2478 mach->pipeline() == MachNode::pipeline_class() &&
2479 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2480 continue;
2481 break; // Funny loop structure to be sure...
2482 }
2483 // Compute last "interesting" instruction in block - last instruction we
2484 // might schedule. _bb_end points just after last schedulable inst. We
2485 // normally schedule conditional branches (despite them being forced last
2486 // in the block), because they have delay slots we can fill. Calls all
2487 // have their delay slots filled in the template expansions, so we don't
2488 // bother scheduling them.
2489 Node *last = bb->get_node(_bb_end);
2490 // Ignore trailing NOPs.
2491 while (_bb_end > 0 && last->is_Mach() &&
2492 last->as_Mach()->ideal_Opcode() == Op_Con) {
2493 last = bb->get_node(--_bb_end);
2494 }
2495 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2496 if( last->is_Catch() ||
2497 // Exclude unreachable path case when Halt node is in a separate block.
2498 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2499 // There must be a prior call. Skip it.
2500 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2501 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2502 }
2503 } else if( last->is_MachNullCheck() ) {
2504 // Backup so the last null-checked memory instruction is
2505 // outside the schedulable range. Skip over the nullcheck,
2506 // projection, and the memory nodes.
2507 Node *mem = last->in(1);
2508 do {
2509 _bb_end--;
2510 } while (mem != bb->get_node(_bb_end));
2511 } else {
2512 // Set _bb_end to point after last schedulable inst.
2513 _bb_end++;
2514 }
2515
2516 assert( _bb_start <= _bb_end, "inverted block ends" );
2517
2518 // Compute the register antidependencies for the basic block
2519 ComputeRegisterAntidependencies(bb);
2520 if (_cfg->C->failing()) return; // too many D-U pinch points
2521
2522 // Compute intra-bb latencies for the nodes
2523 ComputeLocalLatenciesForward(bb);
2524
2525 // Compute the usage within the block, and set the list of all nodes
2526 // in the block that have no uses within the block.
2527 ComputeUseCount(bb);
2528
2529 // Schedule the remaining instructions in the block
2530 while ( _available.size() > 0 ) {
2531 Node *n = ChooseNodeToBundle();
2532 guarantee(n != NULL, "no nodes available");
2533 AddNodeToBundle(n,bb);
2534 }
2535
2536 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2537 #ifdef ASSERT
2538 for( uint l = _bb_start; l < _bb_end; l++ ) {
2539 Node *n = bb->get_node(l);
2540 uint m;
2541 for( m = 0; m < _bb_end-_bb_start; m++ )
2542 if( _scheduled[m] == n )
2543 break;
2544 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2545 }
2546 #endif
2547
2548 // Now copy the instructions (in reverse order) back to the block
2549 for ( uint k = _bb_start; k < _bb_end; k++ )
2550 bb->map_node(_scheduled[_bb_end-k-1], k);
2551
2552 #ifndef PRODUCT
2553 if (_cfg->C->trace_opto_output()) {
2554 tty->print("# Schedule BB#%03d (final)\n", i);
2555 uint current = 0;
2556 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2557 Node *n = bb->get_node(j);
2558 if( valid_bundle_info(n) ) {
2559 Bundle *bundle = node_bundling(n);
2560 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2561 tty->print("*** Bundle: ");
2562 bundle->dump();
2563 }
2564 n->dump();
2565 }
2566 }
2567 }
2568 #endif
2569 #ifdef ASSERT
2570 verify_good_schedule(bb,"after block local scheduling");
2571 #endif
2572 }
2573
2574 #ifndef PRODUCT
2575 if (_cfg->C->trace_opto_output())
2576 tty->print("# <- DoScheduling\n");
2577 #endif
2578
2579 // Record final node-bundling array location
2580 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2581
2582 } // end DoScheduling
2583
2584 // Verify that no live-range used in the block is killed in the block by a
2585 // wrong DEF. This doesn't verify live-ranges that span blocks.
2586
2587 // Check for edge existence. Used to avoid adding redundant precedence edges.
2588 static bool edge_from_to( Node *from, Node *to ) {
2589 for( uint i=0; i<from->len(); i++ )
2590 if( from->in(i) == to )
2591 return true;
2592 return false;
2593 }
2594
2595 #ifdef ASSERT
2596 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2597 // Check for bad kills
2598 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2599 Node *prior_use = _reg_node[def];
2600 if( prior_use && !edge_from_to(prior_use,n) ) {
2601 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2602 n->dump();
2603 tty->print_cr("...");
2604 prior_use->dump();
2605 assert(edge_from_to(prior_use,n),msg);
2606 }
2607 _reg_node.map(def,NULL); // Kill live USEs
2608 }
2609 }
2610
2611 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2612
2613 // Zap to something reasonable for the verify code
2614 _reg_node.clear();
2615
2616 // Walk over the block backwards. Check to make sure each DEF doesn't
2617 // kill a live value (other than the one it's supposed to). Add each
2618 // USE to the live set.
2619 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2620 Node *n = b->get_node(i);
2621 int n_op = n->Opcode();
2622 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2623 // Fat-proj kills a slew of registers
2624 RegMask rm = n->out_RegMask();// Make local copy
2625 while( rm.is_NotEmpty() ) {
2626 OptoReg::Name kill = rm.find_first_elem();
2627 rm.Remove(kill);
2628 verify_do_def( n, kill, msg );
2629 }
2630 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2631 // Get DEF'd registers the normal way
2632 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2633 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2634 }
2635
2636 // Now make all USEs live
2637 for( uint i=1; i<n->req(); i++ ) {
2638 Node *def = n->in(i);
2639 assert(def != 0, "input edge required");
2640 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2641 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2642 if( OptoReg::is_valid(reg_lo) ) {
2643 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2644 _reg_node.map(reg_lo,n);
2645 }
2646 if( OptoReg::is_valid(reg_hi) ) {
2647 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2648 _reg_node.map(reg_hi,n);
2649 }
2650 }
2651
2652 }
2653
2654 // Zap to something reasonable for the Antidependence code
2655 _reg_node.clear();
2656 }
2657 #endif
2658
2659 // Conditionally add precedence edges. Avoid putting edges on Projs.
2660 static void add_prec_edge_from_to( Node *from, Node *to ) {
2661 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2662 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2663 from = from->in(0);
2664 }
2665 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2666 !edge_from_to( from, to ) ) // Avoid duplicate edge
2667 from->add_prec(to);
2668 }
2669
2670 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2671 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2672 return;
2673
2674 Node *pinch = _reg_node[def_reg]; // Get pinch point
2675 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2676 is_def ) { // Check for a true def (not a kill)
2677 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2678 return;
2679 }
2680
2681 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2682 debug_only( def = (Node*)0xdeadbeef; )
2683
2684 // After some number of kills there _may_ be a later def
2685 Node *later_def = NULL;
2686
2687 // Finding a kill requires a real pinch-point.
2688 // Check for not already having a pinch-point.
2689 // Pinch points are Op_Node's.
2690 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2691 later_def = pinch; // Must be def/kill as optimistic pinch-point
2692 if ( _pinch_free_list.size() > 0) {
2693 pinch = _pinch_free_list.pop();
2694 } else {
2695 pinch = new Node(1); // Pinch point to-be
2696 }
2697 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2698 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2699 return;
2700 }
2701 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2702 _reg_node.map(def_reg,pinch); // Record pinch-point
2703 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2704 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2705 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2706 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2707 later_def = NULL; // and no later def
2708 }
2709 pinch->set_req(0,later_def); // Hook later def so we can find it
2710 } else { // Else have valid pinch point
2711 if( pinch->in(0) ) // If there is a later-def
2712 later_def = pinch->in(0); // Get it
2713 }
2714
2715 // Add output-dependence edge from later def to kill
2716 if( later_def ) // If there is some original def
2717 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2718
2719 // See if current kill is also a use, and so is forced to be the pinch-point.
2720 if( pinch->Opcode() == Op_Node ) {
2721 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2722 for( uint i=1; i<uses->req(); i++ ) {
2723 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2724 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2725 // Yes, found a use/kill pinch-point
2726 pinch->set_req(0,NULL); //
2727 pinch->replace_by(kill); // Move anti-dep edges up
2728 pinch = kill;
2729 _reg_node.map(def_reg,pinch);
2730 return;
2731 }
2732 }
2733 }
2734
2735 // Add edge from kill to pinch-point
2736 add_prec_edge_from_to(kill,pinch);
2737 }
2738
2739 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2740 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2741 return;
2742 Node *pinch = _reg_node[use_reg]; // Get pinch point
2743 // Check for no later def_reg/kill in block
2744 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2745 // Use has to be block-local as well
2746 _cfg->get_block_for_node(use) == b) {
2747 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2748 pinch->req() == 1 ) { // pinch not yet in block?
2749 pinch->del_req(0); // yank pointer to later-def, also set flag
2750 // Insert the pinch-point in the block just after the last use
2751 b->insert_node(pinch, b->find_node(use) + 1);
2752 _bb_end++; // Increase size scheduled region in block
2753 }
2754
2755 add_prec_edge_from_to(pinch,use);
2756 }
2757 }
2758
2759 // We insert antidependences between the reads and following write of
2760 // allocated registers to prevent illegal code motion. Hopefully, the
2761 // number of added references should be fairly small, especially as we
2762 // are only adding references within the current basic block.
2763 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2764
2765 #ifdef ASSERT
2766 verify_good_schedule(b,"before block local scheduling");
2767 #endif
2768
2769 // A valid schedule, for each register independently, is an endless cycle
2770 // of: a def, then some uses (connected to the def by true dependencies),
2771 // then some kills (defs with no uses), finally the cycle repeats with a new
2772 // def. The uses are allowed to float relative to each other, as are the
2773 // kills. No use is allowed to slide past a kill (or def). This requires
2774 // antidependencies between all uses of a single def and all kills that
2775 // follow, up to the next def. More edges are redundant, because later defs
2776 // & kills are already serialized with true or antidependencies. To keep
2777 // the edge count down, we add a 'pinch point' node if there's more than
2778 // one use or more than one kill/def.
2779
2780 // We add dependencies in one bottom-up pass.
2781
2782 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2783
2784 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2785 // register. If not, we record the DEF/KILL in _reg_node, the
2786 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2787 // "pinch point", a new Node that's in the graph but not in the block.
2788 // We put edges from the prior and current DEF/KILLs to the pinch point.
2789 // We put the pinch point in _reg_node. If there's already a pinch point
2790 // we merely add an edge from the current DEF/KILL to the pinch point.
2791
2792 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2793 // put an edge from the pinch point to the USE.
2794
2795 // To be expedient, the _reg_node array is pre-allocated for the whole
2796 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2797 // or a valid def/kill/pinch-point, or a leftover node from some prior
2798 // block. Leftover node from some prior block is treated like a NULL (no
2799 // prior def, so no anti-dependence needed). Valid def is distinguished by
2800 // it being in the current block.
2801 bool fat_proj_seen = false;
2802 uint last_safept = _bb_end-1;
2803 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2804 Node* last_safept_node = end_node;
2805 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2806 Node *n = b->get_node(i);
2807 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2808 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2809 // Fat-proj kills a slew of registers
2810 // This can add edges to 'n' and obscure whether or not it was a def,
2811 // hence the is_def flag.
2812 fat_proj_seen = true;
2813 RegMask rm = n->out_RegMask();// Make local copy
2814 while( rm.is_NotEmpty() ) {
2815 OptoReg::Name kill = rm.find_first_elem();
2816 rm.Remove(kill);
2817 anti_do_def( b, n, kill, is_def );
2818 }
2819 } else {
2820 // Get DEF'd registers the normal way
2821 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2822 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2823 }
2824
2825 // Kill projections on a branch should appear to occur on the
2826 // branch, not afterwards, so grab the masks from the projections
2827 // and process them.
2828 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2829 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2830 Node* use = n->fast_out(i);
2831 if (use->is_Proj()) {
2832 RegMask rm = use->out_RegMask();// Make local copy
2833 while( rm.is_NotEmpty() ) {
2834 OptoReg::Name kill = rm.find_first_elem();
2835 rm.Remove(kill);
2836 anti_do_def( b, n, kill, false );
2837 }
2838 }
2839 }
2840 }
2841
2842 // Check each register used by this instruction for a following DEF/KILL
2843 // that must occur afterward and requires an anti-dependence edge.
2844 for( uint j=0; j<n->req(); j++ ) {
2845 Node *def = n->in(j);
2846 if( def ) {
2847 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2848 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2849 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2850 }
2851 }
2852 // Do not allow defs of new derived values to float above GC
2853 // points unless the base is definitely available at the GC point.
2854
2855 Node *m = b->get_node(i);
2856
2857 // Add precedence edge from following safepoint to use of derived pointer
2858 if( last_safept_node != end_node &&
2859 m != last_safept_node) {
2860 for (uint k = 1; k < m->req(); k++) {
2861 const Type *t = m->in(k)->bottom_type();
2862 if( t->isa_oop_ptr() &&
2863 t->is_ptr()->offset() != 0 ) {
2864 last_safept_node->add_prec( m );
2865 break;
2866 }
2867 }
2868 }
2869
2870 if( n->jvms() ) { // Precedence edge from derived to safept
2871 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2872 if( b->get_node(last_safept) != last_safept_node ) {
2873 last_safept = b->find_node(last_safept_node);
2874 }
2875 for( uint j=last_safept; j > i; j-- ) {
2876 Node *mach = b->get_node(j);
2877 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2878 mach->add_prec( n );
2879 }
2880 last_safept = i;
2881 last_safept_node = m;
2882 }
2883 }
2884
2885 if (fat_proj_seen) {
2886 // Garbage collect pinch nodes that were not consumed.
2887 // They are usually created by a fat kill MachProj for a call.
2888 garbage_collect_pinch_nodes();
2889 }
2890 }
2891
2892 // Garbage collect pinch nodes for reuse by other blocks.
2893 //
2894 // The block scheduler's insertion of anti-dependence
2895 // edges creates many pinch nodes when the block contains
2896 // 2 or more Calls. A pinch node is used to prevent a
2897 // combinatorial explosion of edges. If a set of kills for a
2898 // register is anti-dependent on a set of uses (or defs), rather
2899 // than adding an edge in the graph between each pair of kill
2900 // and use (or def), a pinch is inserted between them:
2901 //
2902 // use1 use2 use3
2903 // \ | /
2904 // \ | /
2905 // pinch
2906 // / | \
2907 // / | \
2908 // kill1 kill2 kill3
2909 //
2910 // One pinch node is created per register killed when
2911 // the second call is encountered during a backwards pass
2912 // over the block. Most of these pinch nodes are never
2913 // wired into the graph because the register is never
2914 // used or def'ed in the block.
2915 //
2916 void Scheduling::garbage_collect_pinch_nodes() {
2917 #ifndef PRODUCT
2918 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2919 #endif
2920 int trace_cnt = 0;
2921 for (uint k = 0; k < _reg_node.Size(); k++) {
2922 Node* pinch = _reg_node[k];
2923 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2924 // no predecence input edges
2925 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2926 cleanup_pinch(pinch);
2927 _pinch_free_list.push(pinch);
2928 _reg_node.map(k, NULL);
2929 #ifndef PRODUCT
2930 if (_cfg->C->trace_opto_output()) {
2931 trace_cnt++;
2932 if (trace_cnt > 40) {
2933 tty->print("\n");
2934 trace_cnt = 0;
2935 }
2936 tty->print(" %d", pinch->_idx);
2937 }
2938 #endif
2939 }
2940 }
2941 #ifndef PRODUCT
2942 if (_cfg->C->trace_opto_output()) tty->print("\n");
2943 #endif
2944 }
2945
2946 // Clean up a pinch node for reuse.
2947 void Scheduling::cleanup_pinch( Node *pinch ) {
2948 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2949
2950 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2951 Node* use = pinch->last_out(i);
2952 uint uses_found = 0;
2953 for (uint j = use->req(); j < use->len(); j++) {
2954 if (use->in(j) == pinch) {
2955 use->rm_prec(j);
2956 uses_found++;
2957 }
2958 }
2959 assert(uses_found > 0, "must be a precedence edge");
2960 i -= uses_found; // we deleted 1 or more copies of this edge
2961 }
2962 // May have a later_def entry
2963 pinch->set_req(0, NULL);
2964 }
2965
2966 #ifndef PRODUCT
2967
2968 void Scheduling::dump_available() const {
2969 tty->print("#Availist ");
2970 for (uint i = 0; i < _available.size(); i++)
2971 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2972 tty->cr();
2973 }
2974
2975 // Print Scheduling Statistics
2976 void Scheduling::print_statistics() {
2977 // Print the size added by nops for bundling
2978 tty->print("Nops added %d bytes to total of %d bytes",
2979 _total_nop_size, _total_method_size);
2980 if (_total_method_size > 0)
2981 tty->print(", for %.2f%%",
2982 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2983 tty->print("\n");
2984
2985 // Print the number of branch shadows filled
2986 if (Pipeline::_branch_has_delay_slot) {
2987 tty->print("Of %d branches, %d had unconditional delay slots filled",
2988 _total_branches, _total_unconditional_delays);
2989 if (_total_branches > 0)
2990 tty->print(", for %.2f%%",
2991 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2992 tty->print("\n");
2993 }
2994
2995 uint total_instructions = 0, total_bundles = 0;
2996
2997 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2998 uint bundle_count = _total_instructions_per_bundle[i];
2999 total_instructions += bundle_count * i;
3000 total_bundles += bundle_count;
3001 }
3002
3003 if (total_bundles > 0)
3004 tty->print("Average ILP (excluding nops) is %.2f\n",
3005 ((double)total_instructions) / ((double)total_bundles));
3006 }
3007 #endif