Thu, 24 May 2018 19:24:53 +0800
#7046 Disable the compilation when branch offset is beyond short branch
Contributed-by: fujie, aoqi
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 */
25 /*
26 * This file has been modified by Loongson Technology in 2015. These
27 * modifications are Copyright (c) 2015 Loongson Technology, and are made
28 * available on the same license terms set forth above.
29 */
31 #include "precompiled.hpp"
32 #include "asm/assembler.inline.hpp"
33 #include "code/compiledIC.hpp"
34 #include "code/debugInfo.hpp"
35 #include "code/debugInfoRec.hpp"
36 #include "compiler/compileBroker.hpp"
37 #include "compiler/oopMap.hpp"
38 #include "memory/allocation.inline.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/output.hpp"
44 #include "opto/regalloc.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subnode.hpp"
47 #include "opto/type.hpp"
48 #include "runtime/handles.inline.hpp"
49 #include "utilities/xmlstream.hpp"
51 #ifndef PRODUCT
52 #define DEBUG_ARG(x) , x
53 #else
54 #define DEBUG_ARG(x)
55 #endif
57 // Convert Nodes to instruction bits and pass off to the VM
58 void Compile::Output() {
59 // RootNode goes
60 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
62 // The number of new nodes (mostly MachNop) is proportional to
63 // the number of java calls and inner loops which are aligned.
64 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
65 C->inner_loops()*(OptoLoopAlignment-1)),
66 "out of nodes before code generation" ) ) {
67 return;
68 }
69 // Make sure I can find the Start Node
70 Block *entry = _cfg->get_block(1);
71 Block *broot = _cfg->get_root_block();
73 const StartNode *start = entry->head()->as_Start();
75 // Replace StartNode with prolog
76 MachPrologNode *prolog = new (this) MachPrologNode();
77 entry->map_node(prolog, 0);
78 _cfg->map_node_to_block(prolog, entry);
79 _cfg->unmap_node_from_block(start); // start is no longer in any block
81 // Virtual methods need an unverified entry point
83 if( is_osr_compilation() ) {
84 if( PoisonOSREntry ) {
85 // TODO: Should use a ShouldNotReachHereNode...
86 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
87 }
88 } else {
89 if( _method && !_method->flags().is_static() ) {
90 // Insert unvalidated entry point
91 _cfg->insert( broot, 0, new (this) MachUEPNode() );
92 }
94 }
97 // Break before main entry point
98 if( (_method && _method->break_at_execute())
99 #ifndef PRODUCT
100 ||(OptoBreakpoint && is_method_compilation())
101 ||(OptoBreakpointOSR && is_osr_compilation())
102 ||(OptoBreakpointC2R && !_method)
103 #endif
104 ) {
105 // checking for _method means that OptoBreakpoint does not apply to
106 // runtime stubs or frame converters
107 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
108 }
110 // Insert epilogs before every return
111 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
112 Block* block = _cfg->get_block(i);
113 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
114 Node* m = block->end();
115 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
116 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
117 block->add_inst(epilog);
118 _cfg->map_node_to_block(epilog, block);
119 }
120 }
121 }
123 # ifdef ENABLE_ZAP_DEAD_LOCALS
124 if (ZapDeadCompiledLocals) {
125 Insert_zap_nodes();
126 }
127 # endif
129 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
130 blk_starts[0] = 0;
132 // Initialize code buffer and process short branches.
133 CodeBuffer* cb = init_buffer(blk_starts);
135 if (cb == NULL || failing()) {
136 return;
137 }
139 ScheduleAndBundle();
141 #ifndef PRODUCT
142 if (trace_opto_output()) {
143 tty->print("\n---- After ScheduleAndBundle ----\n");
144 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
145 tty->print("\nBB#%03d:\n", i);
146 Block* block = _cfg->get_block(i);
147 for (uint j = 0; j < block->number_of_nodes(); j++) {
148 Node* n = block->get_node(j);
149 OptoReg::Name reg = _regalloc->get_reg_first(n);
150 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
151 n->dump();
152 }
153 }
154 }
155 #endif
157 if (failing()) {
158 return;
159 }
161 BuildOopMaps();
163 if (failing()) {
164 return;
165 }
167 fill_buffer(cb, blk_starts);
168 }
170 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
171 // Determine if we need to generate a stack overflow check.
172 // Do it if the method is not a stub function and
173 // has java calls or has frame size > vm_page_size/8.
174 // The debug VM checks that deoptimization doesn't trigger an
175 // unexpected stack overflow (compiled method stack banging should
176 // guarantee it doesn't happen) so we always need the stack bang in
177 // a debug VM.
178 return (UseStackBanging && stub_function() == NULL &&
179 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
180 DEBUG_ONLY(|| true)));
181 }
183 bool Compile::need_register_stack_bang() const {
184 // Determine if we need to generate a register stack overflow check.
185 // This is only used on architectures which have split register
186 // and memory stacks (ie. IA64).
187 // Bang if the method is not a stub function and has java calls
188 return (stub_function() == NULL && has_java_calls());
189 }
191 # ifdef ENABLE_ZAP_DEAD_LOCALS
194 // In order to catch compiler oop-map bugs, we have implemented
195 // a debugging mode called ZapDeadCompilerLocals.
196 // This mode causes the compiler to insert a call to a runtime routine,
197 // "zap_dead_locals", right before each place in compiled code
198 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
199 // The runtime routine checks that locations mapped as oops are really
200 // oops, that locations mapped as values do not look like oops,
201 // and that locations mapped as dead are not used later
202 // (by zapping them to an invalid address).
204 int Compile::_CompiledZap_count = 0;
206 void Compile::Insert_zap_nodes() {
207 bool skip = false;
210 // Dink with static counts because code code without the extra
211 // runtime calls is MUCH faster for debugging purposes
213 if ( CompileZapFirst == 0 ) ; // nothing special
214 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
215 else if ( CompileZapFirst == CompiledZap_count() )
216 warning("starting zap compilation after skipping");
218 if ( CompileZapLast == -1 ) ; // nothing special
219 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
220 else if ( CompileZapLast == CompiledZap_count() )
221 warning("about to compile last zap");
223 ++_CompiledZap_count; // counts skipped zaps, too
225 if ( skip ) return;
228 if ( _method == NULL )
229 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
231 // Insert call to zap runtime stub before every node with an oop map
232 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
233 Block *b = _cfg->get_block(i);
234 for ( uint j = 0; j < b->number_of_nodes(); ++j ) {
235 Node *n = b->get_node(j);
237 // Determining if we should insert a zap-a-lot node in output.
238 // We do that for all nodes that has oopmap info, except for calls
239 // to allocation. Calls to allocation passes in the old top-of-eden pointer
240 // and expect the C code to reset it. Hence, there can be no safepoints between
241 // the inlined-allocation and the call to new_Java, etc.
242 // We also cannot zap monitor calls, as they must hold the microlock
243 // during the call to Zap, which also wants to grab the microlock.
244 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
245 if ( insert ) { // it is MachSafePoint
246 if ( !n->is_MachCall() ) {
247 insert = false;
248 } else if ( n->is_MachCall() ) {
249 MachCallNode* call = n->as_MachCall();
250 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
251 call->entry_point() == OptoRuntime::new_array_Java() ||
252 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
253 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
254 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
255 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
256 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
257 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
258 ) {
259 insert = false;
260 }
261 }
262 if (insert) {
263 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
264 b->insert_node(zap, j);
265 _cfg->map_node_to_block(zap, b);
266 ++j;
267 }
268 }
269 }
270 }
271 }
274 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
275 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
276 CallStaticJavaNode* ideal_node =
277 new (this) CallStaticJavaNode( tf,
278 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
279 "call zap dead locals stub", 0, TypePtr::BOTTOM);
280 // We need to copy the OopMap from the site we're zapping at.
281 // We have to make a copy, because the zap site might not be
282 // a call site, and zap_dead is a call site.
283 OopMap* clone = node_to_check->oop_map()->deep_copy();
285 // Add the cloned OopMap to the zap node
286 ideal_node->set_oop_map(clone);
287 return _matcher->match_sfpt(ideal_node);
288 }
290 bool Compile::is_node_getting_a_safepoint( Node* n) {
291 // This code duplicates the logic prior to the call of add_safepoint
292 // below in this file.
293 if( n->is_MachSafePoint() ) return true;
294 return false;
295 }
297 # endif // ENABLE_ZAP_DEAD_LOCALS
299 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
300 // of a loop. When aligning a loop we need to provide enough instructions
301 // in cpu's fetch buffer to feed decoders. The loop alignment could be
302 // avoided if we have enough instructions in fetch buffer at the head of a loop.
303 // By default, the size is set to 999999 by Block's constructor so that
304 // a loop will be aligned if the size is not reset here.
305 //
306 // Note: Mach instructions could contain several HW instructions
307 // so the size is estimated only.
308 //
309 void Compile::compute_loop_first_inst_sizes() {
310 // The next condition is used to gate the loop alignment optimization.
311 // Don't aligned a loop if there are enough instructions at the head of a loop
312 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
313 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
314 // equal to 11 bytes which is the largest address NOP instruction.
315 if (MaxLoopPad < OptoLoopAlignment - 1) {
316 uint last_block = _cfg->number_of_blocks() - 1;
317 for (uint i = 1; i <= last_block; i++) {
318 Block* block = _cfg->get_block(i);
319 // Check the first loop's block which requires an alignment.
320 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
321 uint sum_size = 0;
322 uint inst_cnt = NumberOfLoopInstrToAlign;
323 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
325 // Check subsequent fallthrough blocks if the loop's first
326 // block(s) does not have enough instructions.
327 Block *nb = block;
328 while(inst_cnt > 0 &&
329 i < last_block &&
330 !_cfg->get_block(i + 1)->has_loop_alignment() &&
331 !nb->has_successor(block)) {
332 i++;
333 nb = _cfg->get_block(i);
334 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
335 } // while( inst_cnt > 0 && i < last_block )
337 block->set_first_inst_size(sum_size);
338 } // f( b->head()->is_Loop() )
339 } // for( i <= last_block )
340 } // if( MaxLoopPad < OptoLoopAlignment-1 )
341 }
343 // The architecture description provides short branch variants for some long
344 // branch instructions. Replace eligible long branches with short branches.
345 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
346 // Compute size of each block, method size, and relocation information size
347 uint nblocks = _cfg->number_of_blocks();
349 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
350 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
351 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
353 // Collect worst case block paddings
354 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
355 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
357 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
358 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
360 bool has_short_branch_candidate = false;
362 // Initialize the sizes to 0
363 code_size = 0; // Size in bytes of generated code
364 stub_size = 0; // Size in bytes of all stub entries
365 // Size in bytes of all relocation entries, including those in local stubs.
366 // Start with 2-bytes of reloc info for the unvalidated entry point
367 reloc_size = 1; // Number of relocation entries
369 // Make three passes. The first computes pessimistic blk_starts,
370 // relative jmp_offset and reloc_size information. The second performs
371 // short branch substitution using the pessimistic sizing. The
372 // third inserts nops where needed.
374 // Step one, perform a pessimistic sizing pass.
375 uint last_call_adr = max_uint;
376 uint last_avoid_back_to_back_adr = max_uint;
377 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
378 for (uint i = 0; i < nblocks; i++) { // For all blocks
379 Block* block = _cfg->get_block(i);
381 // During short branch replacement, we store the relative (to blk_starts)
382 // offset of jump in jmp_offset, rather than the absolute offset of jump.
383 // This is so that we do not need to recompute sizes of all nodes when
384 // we compute correct blk_starts in our next sizing pass.
385 jmp_offset[i] = 0;
386 jmp_size[i] = 0;
387 jmp_nidx[i] = -1;
388 DEBUG_ONLY( jmp_target[i] = 0; )
389 DEBUG_ONLY( jmp_rule[i] = 0; )
391 // Sum all instruction sizes to compute block size
392 uint last_inst = block->number_of_nodes();
393 uint blk_size = 0;
394 for (uint j = 0; j < last_inst; j++) {
395 Node* nj = block->get_node(j);
396 // Handle machine instruction nodes
397 if (nj->is_Mach()) {
398 MachNode *mach = nj->as_Mach();
399 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
400 reloc_size += mach->reloc();
401 if (mach->is_MachCall()) {
402 // add size information for trampoline stub
403 // class CallStubImpl is platform-specific and defined in the *.ad files.
404 stub_size += CallStubImpl::size_call_trampoline();
405 reloc_size += CallStubImpl::reloc_call_trampoline();
407 MachCallNode *mcall = mach->as_MachCall();
408 // This destination address is NOT PC-relative
410 mcall->method_set((intptr_t)mcall->entry_point());
412 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
413 stub_size += CompiledStaticCall::to_interp_stub_size();
414 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
415 }
416 } else if (mach->is_MachSafePoint()) {
417 // If call/safepoint are adjacent, account for possible
418 // nop to disambiguate the two safepoints.
419 // ScheduleAndBundle() can rearrange nodes in a block,
420 // check for all offsets inside this block.
421 if (last_call_adr >= blk_starts[i]) {
422 blk_size += nop_size;
423 }
424 }
425 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
426 // Nop is inserted between "avoid back to back" instructions.
427 // ScheduleAndBundle() can rearrange nodes in a block,
428 // check for all offsets inside this block.
429 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
430 blk_size += nop_size;
431 }
432 }
433 if (mach->may_be_short_branch()) {
434 if (!nj->is_MachBranch()) {
435 #ifndef PRODUCT
436 nj->dump(3);
437 #endif
438 Unimplemented();
439 }
440 assert(jmp_nidx[i] == -1, "block should have only one branch");
441 jmp_offset[i] = blk_size;
442 jmp_size[i] = nj->size(_regalloc);
443 jmp_nidx[i] = j;
444 has_short_branch_candidate = true;
445 }
446 }
447 blk_size += nj->size(_regalloc);
448 // Remember end of call offset
449 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
450 last_call_adr = blk_starts[i]+blk_size;
451 }
452 // Remember end of avoid_back_to_back offset
453 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
454 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
455 }
456 }
458 // When the next block starts a loop, we may insert pad NOP
459 // instructions. Since we cannot know our future alignment,
460 // assume the worst.
461 if (i < nblocks - 1) {
462 Block* nb = _cfg->get_block(i + 1);
463 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
464 if (max_loop_pad > 0) {
465 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
466 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
467 // If either is the last instruction in this block, bump by
468 // max_loop_pad in lock-step with blk_size, so sizing
469 // calculations in subsequent blocks still can conservatively
470 // detect that it may the last instruction in this block.
471 if (last_call_adr == blk_starts[i]+blk_size) {
472 last_call_adr += max_loop_pad;
473 }
474 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
475 last_avoid_back_to_back_adr += max_loop_pad;
476 }
477 blk_size += max_loop_pad;
478 block_worst_case_pad[i + 1] = max_loop_pad;
479 }
480 }
482 // Save block size; update total method size
483 blk_starts[i+1] = blk_starts[i]+blk_size;
484 }
486 // Step two, replace eligible long jumps.
487 bool progress = true;
488 uint last_may_be_short_branch_adr = max_uint;
489 while (has_short_branch_candidate && progress) {
490 progress = false;
491 has_short_branch_candidate = false;
492 int adjust_block_start = 0;
493 for (uint i = 0; i < nblocks; i++) {
494 Block* block = _cfg->get_block(i);
495 int idx = jmp_nidx[i];
496 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
497 if (mach != NULL && mach->may_be_short_branch()) {
498 #ifdef ASSERT
499 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
500 int j;
501 // Find the branch; ignore trailing NOPs.
502 for (j = block->number_of_nodes()-1; j>=0; j--) {
503 Node* n = block->get_node(j);
504 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
505 break;
506 }
507 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
508 #endif
509 int br_size = jmp_size[i];
510 int br_offs = blk_starts[i] + jmp_offset[i];
512 // This requires the TRUE branch target be in succs[0]
513 uint bnum = block->non_connector_successor(0)->_pre_order;
514 int offset = blk_starts[bnum] - br_offs;
515 if (bnum > i) { // adjust following block's offset
516 offset -= adjust_block_start;
517 }
519 // This block can be a loop header, account for the padding
520 // in the previous block.
521 int block_padding = block_worst_case_pad[i];
522 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
523 // In the following code a nop could be inserted before
524 // the branch which will increase the backward distance.
525 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
526 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
528 if (needs_padding && offset <= 0)
529 offset -= nop_size;
531 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
532 // We've got a winner. Replace this branch.
533 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
535 // Update the jmp_size.
536 int new_size = replacement->size(_regalloc);
537 int diff = br_size - new_size;
538 assert(diff >= (int)nop_size, "short_branch size should be smaller");
539 // Conservatively take into account padding between
540 // avoid_back_to_back branches. Previous branch could be
541 // converted into avoid_back_to_back branch during next
542 // rounds.
543 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
544 jmp_offset[i] += nop_size;
545 diff -= nop_size;
546 }
547 adjust_block_start += diff;
548 block->map_node(replacement, idx);
549 mach->subsume_by(replacement, C);
550 mach = replacement;
551 progress = true;
553 jmp_size[i] = new_size;
554 DEBUG_ONLY( jmp_target[i] = bnum; );
555 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
556 } else {
557 #ifdef MIPS64
558 env()->record_method_not_compilable("Branch out of range for MIPS");
559 return;
560 #endif
561 // The jump distance is not short, try again during next iteration.
562 has_short_branch_candidate = true;
563 }
564 } // (mach->may_be_short_branch())
565 if (mach != NULL && (mach->may_be_short_branch() ||
566 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
567 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
568 }
569 blk_starts[i+1] -= adjust_block_start;
570 }
571 }
573 #ifdef ASSERT
574 for (uint i = 0; i < nblocks; i++) { // For all blocks
575 if (jmp_target[i] != 0) {
576 int br_size = jmp_size[i];
577 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
578 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
579 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]);
580 }
581 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
582 }
583 }
584 #endif
586 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
587 // after ScheduleAndBundle().
589 // ------------------
590 // Compute size for code buffer
591 code_size = blk_starts[nblocks];
593 // Relocation records
594 reloc_size += 1; // Relo entry for exception handler
596 // Adjust reloc_size to number of record of relocation info
597 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
598 // a relocation index.
599 // The CodeBuffer will expand the locs array if this estimate is too low.
600 reloc_size *= 10 / sizeof(relocInfo);
601 }
603 //------------------------------FillLocArray-----------------------------------
604 // Create a bit of debug info and append it to the array. The mapping is from
605 // Java local or expression stack to constant, register or stack-slot. For
606 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
607 // entry has been taken care of and caller should skip it).
608 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
609 // This should never have accepted Bad before
610 assert(OptoReg::is_valid(regnum), "location must be valid");
611 return (OptoReg::is_reg(regnum))
612 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
613 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
614 }
617 ObjectValue*
618 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
619 for (int i = 0; i < objs->length(); i++) {
620 assert(objs->at(i)->is_object(), "corrupt object cache");
621 ObjectValue* sv = (ObjectValue*) objs->at(i);
622 if (sv->id() == id) {
623 return sv;
624 }
625 }
626 // Otherwise..
627 return NULL;
628 }
630 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
631 ObjectValue* sv ) {
632 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
633 objs->append(sv);
634 }
637 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
638 GrowableArray<ScopeValue*> *array,
639 GrowableArray<ScopeValue*> *objs ) {
640 assert( local, "use _top instead of null" );
641 if (array->length() != idx) {
642 assert(array->length() == idx + 1, "Unexpected array count");
643 // Old functionality:
644 // return
645 // New functionality:
646 // Assert if the local is not top. In product mode let the new node
647 // override the old entry.
648 assert(local == top(), "LocArray collision");
649 if (local == top()) {
650 return;
651 }
652 array->pop();
653 }
654 const Type *t = local->bottom_type();
656 // Is it a safepoint scalar object node?
657 if (local->is_SafePointScalarObject()) {
658 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
660 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
661 if (sv == NULL) {
662 ciKlass* cik = t->is_oopptr()->klass();
663 assert(cik->is_instance_klass() ||
664 cik->is_array_klass(), "Not supported allocation.");
665 sv = new ObjectValue(spobj->_idx,
666 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
667 Compile::set_sv_for_object_node(objs, sv);
669 uint first_ind = spobj->first_index(sfpt->jvms());
670 for (uint i = 0; i < spobj->n_fields(); i++) {
671 Node* fld_node = sfpt->in(first_ind+i);
672 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
673 }
674 }
675 array->append(sv);
676 return;
677 }
679 // Grab the register number for the local
680 OptoReg::Name regnum = _regalloc->get_reg_first(local);
681 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
682 // Record the double as two float registers.
683 // The register mask for such a value always specifies two adjacent
684 // float registers, with the lower register number even.
685 // Normally, the allocation of high and low words to these registers
686 // is irrelevant, because nearly all operations on register pairs
687 // (e.g., StoreD) treat them as a single unit.
688 // Here, we assume in addition that the words in these two registers
689 // stored "naturally" (by operations like StoreD and double stores
690 // within the interpreter) such that the lower-numbered register
691 // is written to the lower memory address. This may seem like
692 // a machine dependency, but it is not--it is a requirement on
693 // the author of the <arch>.ad file to ensure that, for every
694 // even/odd double-register pair to which a double may be allocated,
695 // the word in the even single-register is stored to the first
696 // memory word. (Note that register numbers are completely
697 // arbitrary, and are not tied to any machine-level encodings.)
698 #ifdef _LP64
699 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
700 array->append(new ConstantIntValue(0));
701 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
702 } else if ( t->base() == Type::Long ) {
703 array->append(new ConstantIntValue(0));
704 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
705 } else if ( t->base() == Type::RawPtr ) {
706 // jsr/ret return address which must be restored into a the full
707 // width 64-bit stack slot.
708 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
709 }
710 #else //_LP64
711 #ifdef SPARC
712 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
713 // For SPARC we have to swap high and low words for
714 // long values stored in a single-register (g0-g7).
715 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
716 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
717 } else
718 #endif //SPARC
719 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
720 // Repack the double/long as two jints.
721 // The convention the interpreter uses is that the second local
722 // holds the first raw word of the native double representation.
723 // This is actually reasonable, since locals and stack arrays
724 // grow downwards in all implementations.
725 // (If, on some machine, the interpreter's Java locals or stack
726 // were to grow upwards, the embedded doubles would be word-swapped.)
727 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
728 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
729 }
730 #endif //_LP64
731 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
732 OptoReg::is_reg(regnum) ) {
733 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
734 ? Location::float_in_dbl : Location::normal ));
735 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
736 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
737 ? Location::int_in_long : Location::normal ));
738 } else if( t->base() == Type::NarrowOop ) {
739 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
740 } else {
741 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
742 }
743 return;
744 }
746 // No register. It must be constant data.
747 switch (t->base()) {
748 case Type::Half: // Second half of a double
749 ShouldNotReachHere(); // Caller should skip 2nd halves
750 break;
751 case Type::AnyPtr:
752 array->append(new ConstantOopWriteValue(NULL));
753 break;
754 case Type::AryPtr:
755 case Type::InstPtr: // fall through
756 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
757 break;
758 case Type::NarrowOop:
759 if (t == TypeNarrowOop::NULL_PTR) {
760 array->append(new ConstantOopWriteValue(NULL));
761 } else {
762 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
763 }
764 break;
765 case Type::Int:
766 array->append(new ConstantIntValue(t->is_int()->get_con()));
767 break;
768 case Type::RawPtr:
769 // A return address (T_ADDRESS).
770 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
771 #ifdef _LP64
772 // Must be restored to the full-width 64-bit stack slot.
773 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
774 #else
775 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
776 #endif
777 break;
778 case Type::FloatCon: {
779 float f = t->is_float_constant()->getf();
780 array->append(new ConstantIntValue(jint_cast(f)));
781 break;
782 }
783 case Type::DoubleCon: {
784 jdouble d = t->is_double_constant()->getd();
785 #ifdef _LP64
786 array->append(new ConstantIntValue(0));
787 array->append(new ConstantDoubleValue(d));
788 #else
789 // Repack the double as two jints.
790 // The convention the interpreter uses is that the second local
791 // holds the first raw word of the native double representation.
792 // This is actually reasonable, since locals and stack arrays
793 // grow downwards in all implementations.
794 // (If, on some machine, the interpreter's Java locals or stack
795 // were to grow upwards, the embedded doubles would be word-swapped.)
796 jlong_accessor acc;
797 acc.long_value = jlong_cast(d);
798 array->append(new ConstantIntValue(acc.words[1]));
799 array->append(new ConstantIntValue(acc.words[0]));
800 #endif
801 break;
802 }
803 case Type::Long: {
804 jlong d = t->is_long()->get_con();
805 #ifdef _LP64
806 array->append(new ConstantIntValue(0));
807 array->append(new ConstantLongValue(d));
808 #else
809 // Repack the long as two jints.
810 // The convention the interpreter uses is that the second local
811 // holds the first raw word of the native double representation.
812 // This is actually reasonable, since locals and stack arrays
813 // grow downwards in all implementations.
814 // (If, on some machine, the interpreter's Java locals or stack
815 // were to grow upwards, the embedded doubles would be word-swapped.)
816 jlong_accessor acc;
817 acc.long_value = d;
818 array->append(new ConstantIntValue(acc.words[1]));
819 array->append(new ConstantIntValue(acc.words[0]));
820 #endif
821 break;
822 }
823 case Type::Top: // Add an illegal value here
824 array->append(new LocationValue(Location()));
825 break;
826 default:
827 ShouldNotReachHere();
828 break;
829 }
830 }
832 // Determine if this node starts a bundle
833 bool Compile::starts_bundle(const Node *n) const {
834 return (_node_bundling_limit > n->_idx &&
835 _node_bundling_base[n->_idx].starts_bundle());
836 }
838 //--------------------------Process_OopMap_Node--------------------------------
839 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
841 // Handle special safepoint nodes for synchronization
842 MachSafePointNode *sfn = mach->as_MachSafePoint();
843 MachCallNode *mcall;
845 #ifdef ENABLE_ZAP_DEAD_LOCALS
846 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
847 #endif
849 int safepoint_pc_offset = current_offset;
850 bool is_method_handle_invoke = false;
851 bool return_oop = false;
853 // Add the safepoint in the DebugInfoRecorder
854 if( !mach->is_MachCall() ) {
855 mcall = NULL;
856 #ifdef MIPS64
857 /*
858 2013/10/30 Jin: safepoint_pc_offset should point to tha last instruction in safePoint.
859 In X86 and sparc, their safePoints only contain one instruction.
860 However, we should add current_offset with the size of safePoint in MIPS.
861 0x2d6ff22c: lw s2, 0x14(s2)
862 last_pd->pc_offset()=308, pc_offset=304, bci=64
863 last_pd->pc_offset()=312, pc_offset=312, bci=64
864 /mnt/openjdk6/hotspot/src/share/vm/code/debugInfoRec.cpp, 289 , assert(last_pd->pc_offset() == pc_offset,"must be last pc")
866 ;; Safepoint:
867 ---> pc_offset=304
868 0x2d6ff230: lui at, 0x2b7a ; OopMap{s2=Oop s5=Oop t4=Oop off=308}
869 ;*goto
870 ; - java.util.Hashtable::get@64 (line 353)
871 ---> last_pd(308)
872 0x2d6ff234: lw at, 0xffffc100(at) ;*goto
873 ; - java.util.Hashtable::get@64 (line 353)
874 ; {poll}
875 0x2d6ff238: addiu s0, zero, 0x0
876 */
877 safepoint_pc_offset += sfn->size(_regalloc) - 4;
878 #endif
879 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
880 } else {
881 mcall = mach->as_MachCall();
883 // Is the call a MethodHandle call?
884 if (mcall->is_MachCallJava()) {
885 if (mcall->as_MachCallJava()->_method_handle_invoke) {
886 assert(has_method_handle_invokes(), "must have been set during call generation");
887 is_method_handle_invoke = true;
888 }
889 }
891 // Check if a call returns an object.
892 if (mcall->returns_pointer()) {
893 return_oop = true;
894 }
895 safepoint_pc_offset += mcall->ret_addr_offset();
896 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
897 }
899 // Loop over the JVMState list to add scope information
900 // Do not skip safepoints with a NULL method, they need monitor info
901 JVMState* youngest_jvms = sfn->jvms();
902 int max_depth = youngest_jvms->depth();
904 // Allocate the object pool for scalar-replaced objects -- the map from
905 // small-integer keys (which can be recorded in the local and ostack
906 // arrays) to descriptions of the object state.
907 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
909 // Visit scopes from oldest to youngest.
910 for (int depth = 1; depth <= max_depth; depth++) {
911 JVMState* jvms = youngest_jvms->of_depth(depth);
912 int idx;
913 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
914 // Safepoints that do not have method() set only provide oop-map and monitor info
915 // to support GC; these do not support deoptimization.
916 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
917 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
918 int num_mon = jvms->nof_monitors();
919 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
920 "JVMS local count must match that of the method");
922 // Add Local and Expression Stack Information
924 // Insert locals into the locarray
925 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
926 for( idx = 0; idx < num_locs; idx++ ) {
927 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
928 }
930 // Insert expression stack entries into the exparray
931 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
932 for( idx = 0; idx < num_exps; idx++ ) {
933 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
934 }
936 // Add in mappings of the monitors
937 assert( !method ||
938 !method->is_synchronized() ||
939 method->is_native() ||
940 num_mon > 0 ||
941 !GenerateSynchronizationCode,
942 "monitors must always exist for synchronized methods");
944 // Build the growable array of ScopeValues for exp stack
945 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
947 // Loop over monitors and insert into array
948 for (idx = 0; idx < num_mon; idx++) {
949 // Grab the node that defines this monitor
950 Node* box_node = sfn->monitor_box(jvms, idx);
951 Node* obj_node = sfn->monitor_obj(jvms, idx);
953 // Create ScopeValue for object
954 ScopeValue *scval = NULL;
956 if (obj_node->is_SafePointScalarObject()) {
957 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
958 scval = Compile::sv_for_node_id(objs, spobj->_idx);
959 if (scval == NULL) {
960 const Type *t = spobj->bottom_type();
961 ciKlass* cik = t->is_oopptr()->klass();
962 assert(cik->is_instance_klass() ||
963 cik->is_array_klass(), "Not supported allocation.");
964 ObjectValue* sv = new ObjectValue(spobj->_idx,
965 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
966 Compile::set_sv_for_object_node(objs, sv);
968 uint first_ind = spobj->first_index(youngest_jvms);
969 for (uint i = 0; i < spobj->n_fields(); i++) {
970 Node* fld_node = sfn->in(first_ind+i);
971 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
972 }
973 scval = sv;
974 }
975 } else if (!obj_node->is_Con()) {
976 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
977 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
978 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
979 } else {
980 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
981 }
982 } else {
983 const TypePtr *tp = obj_node->get_ptr_type();
984 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
985 }
987 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
988 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
989 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
990 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
991 }
993 // We dump the object pool first, since deoptimization reads it in first.
994 debug_info()->dump_object_pool(objs);
996 // Build first class objects to pass to scope
997 DebugToken *locvals = debug_info()->create_scope_values(locarray);
998 DebugToken *expvals = debug_info()->create_scope_values(exparray);
999 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
1001 // Make method available for all Safepoints
1002 ciMethod* scope_method = method ? method : _method;
1003 // Describe the scope here
1004 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
1005 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
1006 // Now we can describe the scope.
1007 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
1008 } // End jvms loop
1010 // Mark the end of the scope set.
1011 debug_info()->end_safepoint(safepoint_pc_offset);
1012 }
1016 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
1017 class NonSafepointEmitter {
1018 Compile* C;
1019 JVMState* _pending_jvms;
1020 int _pending_offset;
1022 void emit_non_safepoint();
1024 public:
1025 NonSafepointEmitter(Compile* compile) {
1026 this->C = compile;
1027 _pending_jvms = NULL;
1028 _pending_offset = 0;
1029 }
1031 void observe_instruction(Node* n, int pc_offset) {
1032 if (!C->debug_info()->recording_non_safepoints()) return;
1034 Node_Notes* nn = C->node_notes_at(n->_idx);
1035 if (nn == NULL || nn->jvms() == NULL) return;
1036 if (_pending_jvms != NULL &&
1037 _pending_jvms->same_calls_as(nn->jvms())) {
1038 // Repeated JVMS? Stretch it up here.
1039 _pending_offset = pc_offset;
1040 } else {
1041 if (_pending_jvms != NULL &&
1042 _pending_offset < pc_offset) {
1043 emit_non_safepoint();
1044 }
1045 _pending_jvms = NULL;
1046 if (pc_offset > C->debug_info()->last_pc_offset()) {
1047 // This is the only way _pending_jvms can become non-NULL:
1048 _pending_jvms = nn->jvms();
1049 _pending_offset = pc_offset;
1050 }
1051 }
1052 }
1054 // Stay out of the way of real safepoints:
1055 void observe_safepoint(JVMState* jvms, int pc_offset) {
1056 if (_pending_jvms != NULL &&
1057 !_pending_jvms->same_calls_as(jvms) &&
1058 _pending_offset < pc_offset) {
1059 emit_non_safepoint();
1060 }
1061 _pending_jvms = NULL;
1062 }
1064 void flush_at_end() {
1065 if (_pending_jvms != NULL) {
1066 emit_non_safepoint();
1067 }
1068 _pending_jvms = NULL;
1069 }
1070 };
1072 void NonSafepointEmitter::emit_non_safepoint() {
1073 JVMState* youngest_jvms = _pending_jvms;
1074 int pc_offset = _pending_offset;
1076 // Clear it now:
1077 _pending_jvms = NULL;
1079 DebugInformationRecorder* debug_info = C->debug_info();
1080 assert(debug_info->recording_non_safepoints(), "sanity");
1082 debug_info->add_non_safepoint(pc_offset);
1083 int max_depth = youngest_jvms->depth();
1085 // Visit scopes from oldest to youngest.
1086 for (int depth = 1; depth <= max_depth; depth++) {
1087 JVMState* jvms = youngest_jvms->of_depth(depth);
1088 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1089 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1090 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1091 }
1093 // Mark the end of the scope set.
1094 debug_info->end_non_safepoint(pc_offset);
1095 }
1097 //------------------------------init_buffer------------------------------------
1098 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1100 // Set the initially allocated size
1101 int code_req = initial_code_capacity;
1102 int locs_req = initial_locs_capacity;
1103 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1104 int const_req = initial_const_capacity;
1106 int pad_req = NativeCall::instruction_size;
1107 // The extra spacing after the code is necessary on some platforms.
1108 // Sometimes we need to patch in a jump after the last instruction,
1109 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1111 // Compute the byte offset where we can store the deopt pc.
1112 if (fixed_slots() != 0) {
1113 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1114 }
1116 // Compute prolog code size
1117 _method_size = 0;
1118 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1119 #if defined(IA64) && !defined(AIX)
1120 if (save_argument_registers()) {
1121 // 4815101: this is a stub with implicit and unknown precision fp args.
1122 // The usual spill mechanism can only generate stfd's in this case, which
1123 // doesn't work if the fp reg to spill contains a single-precision denorm.
1124 // Instead, we hack around the normal spill mechanism using stfspill's and
1125 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1126 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1127 //
1128 // If we ever implement 16-byte 'registers' == stack slots, we can
1129 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1130 // instead of stfd/stfs/ldfd/ldfs.
1131 _frame_slots += 8*(16/BytesPerInt);
1132 }
1133 #endif
1134 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1136 if (has_mach_constant_base_node()) {
1137 uint add_size = 0;
1138 // Fill the constant table.
1139 // Note: This must happen before shorten_branches.
1140 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1141 Block* b = _cfg->get_block(i);
1143 for (uint j = 0; j < b->number_of_nodes(); j++) {
1144 Node* n = b->get_node(j);
1146 // If the node is a MachConstantNode evaluate the constant
1147 // value section.
1148 if (n->is_MachConstant()) {
1149 MachConstantNode* machcon = n->as_MachConstant();
1150 machcon->eval_constant(C);
1151 } else if (n->is_Mach()) {
1152 // On Power there are more nodes that issue constants.
1153 add_size += (n->as_Mach()->ins_num_consts() * 8);
1154 }
1155 }
1156 }
1158 // Calculate the offsets of the constants and the size of the
1159 // constant table (including the padding to the next section).
1160 constant_table().calculate_offsets_and_size();
1161 const_req = constant_table().size() + add_size;
1162 }
1164 // Initialize the space for the BufferBlob used to find and verify
1165 // instruction size in MachNode::emit_size()
1166 init_scratch_buffer_blob(const_req);
1167 if (failing()) return NULL; // Out of memory
1169 // Pre-compute the length of blocks and replace
1170 // long branches with short if machine supports it.
1171 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1172 #ifdef MIPS64
1173 if (failing()) return NULL; // Branch out of range for MIPS
1174 #endif
1176 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1177 // class HandlerImpl is platform-specific and defined in the *.ad files.
1178 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1179 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1180 stub_req += MAX_stubs_size; // ensure per-stub margin
1181 code_req += MAX_inst_size; // ensure per-instruction margin
1183 if (StressCodeBuffers)
1184 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1186 int total_req =
1187 const_req +
1188 code_req +
1189 pad_req +
1190 stub_req +
1191 exception_handler_req +
1192 deopt_handler_req; // deopt handler
1194 if (has_method_handle_invokes())
1195 total_req += deopt_handler_req; // deopt MH handler
1197 CodeBuffer* cb = code_buffer();
1198 cb->initialize(total_req, locs_req);
1200 // Have we run out of code space?
1201 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1202 C->record_failure("CodeCache is full");
1203 return NULL;
1204 }
1205 // Configure the code buffer.
1206 cb->initialize_consts_size(const_req);
1207 cb->initialize_stubs_size(stub_req);
1208 cb->initialize_oop_recorder(env()->oop_recorder());
1210 // fill in the nop array for bundling computations
1211 MachNode *_nop_list[Bundle::_nop_count];
1212 Bundle::initialize_nops(_nop_list, this);
1214 return cb;
1215 }
1217 //------------------------------fill_buffer------------------------------------
1218 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1219 // blk_starts[] contains offsets calculated during short branches processing,
1220 // offsets should not be increased during following steps.
1222 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1223 // of a loop. It is used to determine the padding for loop alignment.
1224 compute_loop_first_inst_sizes();
1226 // Create oopmap set.
1227 _oop_map_set = new OopMapSet();
1229 // !!!!! This preserves old handling of oopmaps for now
1230 debug_info()->set_oopmaps(_oop_map_set);
1232 uint nblocks = _cfg->number_of_blocks();
1233 // Count and start of implicit null check instructions
1234 uint inct_cnt = 0;
1235 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1237 // Count and start of calls
1238 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1240 uint return_offset = 0;
1241 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1243 int previous_offset = 0;
1244 int current_offset = 0;
1245 int last_call_offset = -1;
1246 int last_avoid_back_to_back_offset = -1;
1247 #ifdef ASSERT
1248 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1249 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1250 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1251 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1252 #endif
1254 // Create an array of unused labels, one for each basic block, if printing is enabled
1255 #ifndef PRODUCT
1256 int *node_offsets = NULL;
1257 uint node_offset_limit = unique();
1259 if (print_assembly())
1260 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1261 #endif
1263 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1265 // Emit the constant table.
1266 if (has_mach_constant_base_node()) {
1267 constant_table().emit(*cb);
1268 }
1270 // Create an array of labels, one for each basic block
1271 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1272 for (uint i=0; i <= nblocks; i++) {
1273 blk_labels[i].init();
1274 }
1276 // ------------------
1277 // Now fill in the code buffer
1278 Node *delay_slot = NULL;
1280 for (uint i = 0; i < nblocks; i++) {
1281 Block* block = _cfg->get_block(i);
1282 Node* head = block->head();
1284 // If this block needs to start aligned (i.e, can be reached other
1285 // than by falling-thru from the previous block), then force the
1286 // start of a new bundle.
1287 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1288 cb->flush_bundle(true);
1289 }
1291 #ifdef ASSERT
1292 if (!block->is_connector()) {
1293 stringStream st;
1294 block->dump_head(_cfg, &st);
1295 MacroAssembler(cb).block_comment(st.as_string());
1296 }
1297 jmp_target[i] = 0;
1298 jmp_offset[i] = 0;
1299 jmp_size[i] = 0;
1300 jmp_rule[i] = 0;
1301 #endif
1302 int blk_offset = current_offset;
1304 // Define the label at the beginning of the basic block
1305 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1307 uint last_inst = block->number_of_nodes();
1309 // Emit block normally, except for last instruction.
1310 // Emit means "dump code bits into code buffer".
1311 for (uint j = 0; j<last_inst; j++) {
1313 // Get the node
1314 Node* n = block->get_node(j);
1316 // See if delay slots are supported
1317 if (valid_bundle_info(n) &&
1318 node_bundling(n)->used_in_unconditional_delay()) {
1319 assert(delay_slot == NULL, "no use of delay slot node");
1320 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1322 delay_slot = n;
1323 continue;
1324 }
1326 // If this starts a new instruction group, then flush the current one
1327 // (but allow split bundles)
1328 if (Pipeline::requires_bundling() && starts_bundle(n))
1329 cb->flush_bundle(false);
1331 // The following logic is duplicated in the code ifdeffed for
1332 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1333 // should be factored out. Or maybe dispersed to the nodes?
1335 // Special handling for SafePoint/Call Nodes
1336 bool is_mcall = false;
1337 if (n->is_Mach()) {
1338 MachNode *mach = n->as_Mach();
1339 is_mcall = n->is_MachCall();
1340 bool is_sfn = n->is_MachSafePoint();
1342 // If this requires all previous instructions be flushed, then do so
1343 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1344 cb->flush_bundle(true);
1345 current_offset = cb->insts_size();
1346 }
1348 // A padding may be needed again since a previous instruction
1349 // could be moved to delay slot.
1351 // align the instruction if necessary
1352 int padding = mach->compute_padding(current_offset);
1353 // Make sure safepoint node for polling is distinct from a call's
1354 // return by adding a nop if needed.
1355 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1356 padding = nop_size;
1357 }
1358 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1359 current_offset == last_avoid_back_to_back_offset) {
1360 // Avoid back to back some instructions.
1361 padding = nop_size;
1362 }
1364 if(padding > 0) {
1365 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1366 int nops_cnt = padding / nop_size;
1367 MachNode *nop = new (this) MachNopNode(nops_cnt);
1368 block->insert_node(nop, j++);
1369 last_inst++;
1370 _cfg->map_node_to_block(nop, block);
1371 nop->emit(*cb, _regalloc);
1372 cb->flush_bundle(true);
1373 current_offset = cb->insts_size();
1374 }
1376 // Remember the start of the last call in a basic block
1377 if (is_mcall) {
1378 MachCallNode *mcall = mach->as_MachCall();
1380 // This destination address is NOT PC-relative
1381 mcall->method_set((intptr_t)mcall->entry_point());
1383 // Save the return address
1384 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1386 if (mcall->is_MachCallLeaf()) {
1387 is_mcall = false;
1388 is_sfn = false;
1389 }
1390 }
1392 // sfn will be valid whenever mcall is valid now because of inheritance
1393 if (is_sfn || is_mcall) {
1395 // Handle special safepoint nodes for synchronization
1396 if (!is_mcall) {
1397 MachSafePointNode *sfn = mach->as_MachSafePoint();
1398 // !!!!! Stubs only need an oopmap right now, so bail out
1399 if (sfn->jvms()->method() == NULL) {
1400 // Write the oopmap directly to the code blob??!!
1401 # ifdef ENABLE_ZAP_DEAD_LOCALS
1402 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1403 # endif
1404 continue;
1405 }
1406 } // End synchronization
1408 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1409 current_offset);
1410 Process_OopMap_Node(mach, current_offset);
1411 } // End if safepoint
1413 // If this is a null check, then add the start of the previous instruction to the list
1414 else if( mach->is_MachNullCheck() ) {
1415 inct_starts[inct_cnt++] = previous_offset;
1416 }
1418 // If this is a branch, then fill in the label with the target BB's label
1419 else if (mach->is_MachBranch()) {
1420 // This requires the TRUE branch target be in succs[0]
1421 uint block_num = block->non_connector_successor(0)->_pre_order;
1423 // Try to replace long branch if delay slot is not used,
1424 // it is mostly for back branches since forward branch's
1425 // distance is not updated yet.
1426 bool delay_slot_is_used = valid_bundle_info(n) &&
1427 node_bundling(n)->use_unconditional_delay();
1428 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1429 assert(delay_slot == NULL, "not expecting delay slot node");
1430 int br_size = n->size(_regalloc);
1431 int offset = blk_starts[block_num] - current_offset;
1432 if (block_num >= i) {
1433 // Current and following block's offset are not
1434 // finalized yet, adjust distance by the difference
1435 // between calculated and final offsets of current block.
1436 offset -= (blk_starts[i] - blk_offset);
1437 }
1438 // In the following code a nop could be inserted before
1439 // the branch which will increase the backward distance.
1440 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1441 if (needs_padding && offset <= 0)
1442 offset -= nop_size;
1444 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1445 // We've got a winner. Replace this branch.
1446 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1448 // Update the jmp_size.
1449 int new_size = replacement->size(_regalloc);
1450 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1451 // Insert padding between avoid_back_to_back branches.
1452 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1453 MachNode *nop = new (this) MachNopNode();
1454 block->insert_node(nop, j++);
1455 _cfg->map_node_to_block(nop, block);
1456 last_inst++;
1457 nop->emit(*cb, _regalloc);
1458 cb->flush_bundle(true);
1459 current_offset = cb->insts_size();
1460 }
1461 #ifdef ASSERT
1462 jmp_target[i] = block_num;
1463 jmp_offset[i] = current_offset - blk_offset;
1464 jmp_size[i] = new_size;
1465 jmp_rule[i] = mach->rule();
1466 #endif
1467 block->map_node(replacement, j);
1468 mach->subsume_by(replacement, C);
1469 n = replacement;
1470 mach = replacement;
1471 }
1472 }
1473 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1474 } else if (mach->ideal_Opcode() == Op_Jump) {
1475 for (uint h = 0; h < block->_num_succs; h++) {
1476 Block* succs_block = block->_succs[h];
1477 for (uint j = 1; j < succs_block->num_preds(); j++) {
1478 Node* jpn = succs_block->pred(j);
1479 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1480 uint block_num = succs_block->non_connector()->_pre_order;
1481 Label *blkLabel = &blk_labels[block_num];
1482 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1483 }
1484 }
1485 }
1486 }
1487 #ifdef ASSERT
1488 // Check that oop-store precedes the card-mark
1489 else if (mach->ideal_Opcode() == Op_StoreCM) {
1490 uint storeCM_idx = j;
1491 int count = 0;
1492 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1493 Node *oop_store = mach->in(prec); // Precedence edge
1494 if (oop_store == NULL) continue;
1495 count++;
1496 uint i4;
1497 for (i4 = 0; i4 < last_inst; ++i4) {
1498 if (block->get_node(i4) == oop_store) {
1499 break;
1500 }
1501 }
1502 // Note: This test can provide a false failure if other precedence
1503 // edges have been added to the storeCMNode.
1504 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1505 }
1506 assert(count > 0, "storeCM expects at least one precedence edge");
1507 }
1508 #endif
1509 else if (!n->is_Proj()) {
1510 // Remember the beginning of the previous instruction, in case
1511 // it's followed by a flag-kill and a null-check. Happens on
1512 // Intel all the time, with add-to-memory kind of opcodes.
1513 previous_offset = current_offset;
1514 }
1516 // Not an else-if!
1517 // If this is a trap based cmp then add its offset to the list.
1518 if (mach->is_TrapBasedCheckNode()) {
1519 inct_starts[inct_cnt++] = current_offset;
1520 }
1521 }
1523 // Verify that there is sufficient space remaining
1524 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1525 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1526 C->record_failure("CodeCache is full");
1527 return;
1528 }
1530 // Save the offset for the listing
1531 #ifndef PRODUCT
1532 if (node_offsets && n->_idx < node_offset_limit)
1533 node_offsets[n->_idx] = cb->insts_size();
1534 #endif
1536 // "Normal" instruction case
1537 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1538 n->emit(*cb, _regalloc);
1539 current_offset = cb->insts_size();
1540 #ifdef MIPS64
1541 if (!n->is_Proj()) {
1542 // For MIPS, the first instruction of the previous node (usually a instruction sequence) sometime
1543 // is not the instruction which access memory. adjust is needed. previous_offset points to the
1544 // instruction which access memory. Instruction size is 4. cb->insts_size() and
1545 // cb->insts()->end() are the location of current instruction.
1546 int adjust = 4;
1547 NativeInstruction* inst = (NativeInstruction*) (cb->insts()->end() - 4);
1548 if (inst->is_sync()) {
1549 // a sync may be the last instruction, see store_B_immI_enc_sync
1550 adjust += 4;
1551 inst = (NativeInstruction*) (cb->insts()->end() - 8);
1552 }
1553 previous_offset = current_offset - adjust;
1554 }
1555 #endif
1557 // Above we only verified that there is enough space in the instruction section.
1558 // However, the instruction may emit stubs that cause code buffer expansion.
1559 // Bail out here if expansion failed due to a lack of code cache space.
1560 if (failing()) {
1561 return;
1562 }
1564 #ifdef ASSERT
1565 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1566 n->dump();
1567 assert(false, "wrong size of mach node");
1568 }
1569 #endif
1570 non_safepoints.observe_instruction(n, current_offset);
1572 // mcall is last "call" that can be a safepoint
1573 // record it so we can see if a poll will directly follow it
1574 // in which case we'll need a pad to make the PcDesc sites unique
1575 // see 5010568. This can be slightly inaccurate but conservative
1576 // in the case that return address is not actually at current_offset.
1577 // This is a small price to pay.
1579 if (is_mcall) {
1580 last_call_offset = current_offset;
1581 }
1583 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1584 // Avoid back to back some instructions.
1585 last_avoid_back_to_back_offset = current_offset;
1586 }
1588 // See if this instruction has a delay slot
1589 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1590 assert(delay_slot != NULL, "expecting delay slot node");
1592 // Back up 1 instruction
1593 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1595 // Save the offset for the listing
1596 #ifndef PRODUCT
1597 if (node_offsets && delay_slot->_idx < node_offset_limit)
1598 node_offsets[delay_slot->_idx] = cb->insts_size();
1599 #endif
1601 // Support a SafePoint in the delay slot
1602 if (delay_slot->is_MachSafePoint()) {
1603 MachNode *mach = delay_slot->as_Mach();
1604 // !!!!! Stubs only need an oopmap right now, so bail out
1605 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1606 // Write the oopmap directly to the code blob??!!
1607 # ifdef ENABLE_ZAP_DEAD_LOCALS
1608 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1609 # endif
1610 delay_slot = NULL;
1611 continue;
1612 }
1614 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1615 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1616 adjusted_offset);
1617 // Generate an OopMap entry
1618 Process_OopMap_Node(mach, adjusted_offset);
1619 }
1621 // Insert the delay slot instruction
1622 delay_slot->emit(*cb, _regalloc);
1624 // Don't reuse it
1625 delay_slot = NULL;
1626 }
1628 } // End for all instructions in block
1630 // If the next block is the top of a loop, pad this block out to align
1631 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1632 if (i < nblocks-1) {
1633 Block *nb = _cfg->get_block(i + 1);
1634 int padding = nb->alignment_padding(current_offset);
1635 if( padding > 0 ) {
1636 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1637 block->insert_node(nop, block->number_of_nodes());
1638 _cfg->map_node_to_block(nop, block);
1639 nop->emit(*cb, _regalloc);
1640 current_offset = cb->insts_size();
1641 }
1642 }
1643 // Verify that the distance for generated before forward
1644 // short branches is still valid.
1645 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1647 // Save new block start offset
1648 blk_starts[i] = blk_offset;
1649 } // End of for all blocks
1650 blk_starts[nblocks] = current_offset;
1652 non_safepoints.flush_at_end();
1654 // Offset too large?
1655 if (failing()) return;
1657 // Define a pseudo-label at the end of the code
1658 MacroAssembler(cb).bind( blk_labels[nblocks] );
1660 // Compute the size of the first block
1661 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1663 assert(cb->insts_size() < 500000, "method is unreasonably large");
1665 #ifdef ASSERT
1666 for (uint i = 0; i < nblocks; i++) { // For all blocks
1667 if (jmp_target[i] != 0) {
1668 int br_size = jmp_size[i];
1669 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1670 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1671 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]);
1672 assert(false, "Displacement too large for short jmp");
1673 }
1674 }
1675 }
1676 #endif
1678 #ifndef PRODUCT
1679 // Information on the size of the method, without the extraneous code
1680 Scheduling::increment_method_size(cb->insts_size());
1681 #endif
1683 // ------------------
1684 // Fill in exception table entries.
1685 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1687 // Only java methods have exception handlers and deopt handlers
1688 // class HandlerImpl is platform-specific and defined in the *.ad files.
1689 if (_method) {
1690 // Emit the exception handler code.
1691 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1692 if (failing()) {
1693 return; // CodeBuffer::expand failed
1694 }
1695 // Emit the deopt handler code.
1696 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1698 // Emit the MethodHandle deopt handler code (if required).
1699 if (has_method_handle_invokes() && !failing()) {
1700 // We can use the same code as for the normal deopt handler, we
1701 // just need a different entry point address.
1702 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1703 }
1704 }
1706 // One last check for failed CodeBuffer::expand:
1707 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1708 C->record_failure("CodeCache is full");
1709 return;
1710 }
1712 #ifndef PRODUCT
1713 // Dump the assembly code, including basic-block numbers
1714 if (print_assembly()) {
1715 ttyLocker ttyl; // keep the following output all in one block
1716 if (!VMThread::should_terminate()) { // test this under the tty lock
1717 // This output goes directly to the tty, not the compiler log.
1718 // To enable tools to match it up with the compilation activity,
1719 // be sure to tag this tty output with the compile ID.
1720 if (xtty != NULL) {
1721 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1722 is_osr_compilation() ? " compile_kind='osr'" :
1723 "");
1724 }
1725 if (method() != NULL) {
1726 method()->print_metadata();
1727 }
1728 dump_asm(node_offsets, node_offset_limit);
1729 if (xtty != NULL) {
1730 xtty->tail("opto_assembly");
1731 }
1732 }
1733 }
1734 #endif
1736 }
1738 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1739 _inc_table.set_size(cnt);
1741 uint inct_cnt = 0;
1742 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1743 Block* block = _cfg->get_block(i);
1744 Node *n = NULL;
1745 int j;
1747 // Find the branch; ignore trailing NOPs.
1748 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1749 n = block->get_node(j);
1750 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1751 break;
1752 }
1753 }
1755 // If we didn't find anything, continue
1756 if (j < 0) {
1757 continue;
1758 }
1760 // Compute ExceptionHandlerTable subtable entry and add it
1761 // (skip empty blocks)
1762 if (n->is_Catch()) {
1764 // Get the offset of the return from the call
1765 uint call_return = call_returns[block->_pre_order];
1766 #ifdef ASSERT
1767 assert( call_return > 0, "no call seen for this basic block" );
1768 while (block->get_node(--j)->is_MachProj()) ;
1769 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1770 #endif
1771 // last instruction is a CatchNode, find it's CatchProjNodes
1772 int nof_succs = block->_num_succs;
1773 // allocate space
1774 GrowableArray<intptr_t> handler_bcis(nof_succs);
1775 GrowableArray<intptr_t> handler_pcos(nof_succs);
1776 // iterate through all successors
1777 for (int j = 0; j < nof_succs; j++) {
1778 Block* s = block->_succs[j];
1779 bool found_p = false;
1780 for (uint k = 1; k < s->num_preds(); k++) {
1781 Node* pk = s->pred(k);
1782 if (pk->is_CatchProj() && pk->in(0) == n) {
1783 const CatchProjNode* p = pk->as_CatchProj();
1784 found_p = true;
1785 // add the corresponding handler bci & pco information
1786 if (p->_con != CatchProjNode::fall_through_index) {
1787 // p leads to an exception handler (and is not fall through)
1788 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1789 // no duplicates, please
1790 if (!handler_bcis.contains(p->handler_bci())) {
1791 uint block_num = s->non_connector()->_pre_order;
1792 handler_bcis.append(p->handler_bci());
1793 handler_pcos.append(blk_labels[block_num].loc_pos());
1794 }
1795 }
1796 }
1797 }
1798 assert(found_p, "no matching predecessor found");
1799 // Note: Due to empty block removal, one block may have
1800 // several CatchProj inputs, from the same Catch.
1801 }
1803 // Set the offset of the return from the call
1804 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1805 continue;
1806 }
1808 // Handle implicit null exception table updates
1809 if (n->is_MachNullCheck()) {
1810 uint block_num = block->non_connector_successor(0)->_pre_order;
1811 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1812 continue;
1813 }
1814 // Handle implicit exception table updates: trap instructions.
1815 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1816 uint block_num = block->non_connector_successor(0)->_pre_order;
1817 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1818 continue;
1819 }
1820 } // End of for all blocks fill in exception table entries
1821 }
1823 // Static Variables
1824 #ifndef PRODUCT
1825 uint Scheduling::_total_nop_size = 0;
1826 uint Scheduling::_total_method_size = 0;
1827 uint Scheduling::_total_branches = 0;
1828 uint Scheduling::_total_unconditional_delays = 0;
1829 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1830 #endif
1832 // Initializer for class Scheduling
1834 Scheduling::Scheduling(Arena *arena, Compile &compile)
1835 : _arena(arena),
1836 _cfg(compile.cfg()),
1837 _regalloc(compile.regalloc()),
1838 _reg_node(arena),
1839 _bundle_instr_count(0),
1840 _bundle_cycle_number(0),
1841 _scheduled(arena),
1842 _available(arena),
1843 _next_node(NULL),
1844 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1845 _pinch_free_list(arena)
1846 #ifndef PRODUCT
1847 , _branches(0)
1848 , _unconditional_delays(0)
1849 #endif
1850 {
1851 // Create a MachNopNode
1852 _nop = new (&compile) MachNopNode();
1854 // Now that the nops are in the array, save the count
1855 // (but allow entries for the nops)
1856 _node_bundling_limit = compile.unique();
1857 uint node_max = _regalloc->node_regs_max_index();
1859 compile.set_node_bundling_limit(_node_bundling_limit);
1861 // This one is persistent within the Compile class
1862 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1864 // Allocate space for fixed-size arrays
1865 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1866 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1867 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1869 // Clear the arrays
1870 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1871 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1872 memset(_uses, 0, node_max * sizeof(short));
1873 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1875 // Clear the bundling information
1876 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1878 // Get the last node
1879 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1881 _next_node = block->get_node(block->number_of_nodes() - 1);
1882 }
1884 #ifndef PRODUCT
1885 // Scheduling destructor
1886 Scheduling::~Scheduling() {
1887 _total_branches += _branches;
1888 _total_unconditional_delays += _unconditional_delays;
1889 }
1890 #endif
1892 // Step ahead "i" cycles
1893 void Scheduling::step(uint i) {
1895 Bundle *bundle = node_bundling(_next_node);
1896 bundle->set_starts_bundle();
1898 // Update the bundle record, but leave the flags information alone
1899 if (_bundle_instr_count > 0) {
1900 bundle->set_instr_count(_bundle_instr_count);
1901 bundle->set_resources_used(_bundle_use.resourcesUsed());
1902 }
1904 // Update the state information
1905 _bundle_instr_count = 0;
1906 _bundle_cycle_number += i;
1907 _bundle_use.step(i);
1908 }
1910 void Scheduling::step_and_clear() {
1911 Bundle *bundle = node_bundling(_next_node);
1912 bundle->set_starts_bundle();
1914 // Update the bundle record
1915 if (_bundle_instr_count > 0) {
1916 bundle->set_instr_count(_bundle_instr_count);
1917 bundle->set_resources_used(_bundle_use.resourcesUsed());
1919 _bundle_cycle_number += 1;
1920 }
1922 // Clear the bundling information
1923 _bundle_instr_count = 0;
1924 _bundle_use.reset();
1926 memcpy(_bundle_use_elements,
1927 Pipeline_Use::elaborated_elements,
1928 sizeof(Pipeline_Use::elaborated_elements));
1929 }
1931 // Perform instruction scheduling and bundling over the sequence of
1932 // instructions in backwards order.
1933 void Compile::ScheduleAndBundle() {
1935 // Don't optimize this if it isn't a method
1936 if (!_method)
1937 return;
1939 // Don't optimize this if scheduling is disabled
1940 if (!do_scheduling())
1941 return;
1943 // Scheduling code works only with pairs (8 bytes) maximum.
1944 if (max_vector_size() > 8)
1945 return;
1947 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1949 // Create a data structure for all the scheduling information
1950 Scheduling scheduling(Thread::current()->resource_area(), *this);
1952 // Walk backwards over each basic block, computing the needed alignment
1953 // Walk over all the basic blocks
1954 scheduling.DoScheduling();
1955 }
1957 // Compute the latency of all the instructions. This is fairly simple,
1958 // because we already have a legal ordering. Walk over the instructions
1959 // from first to last, and compute the latency of the instruction based
1960 // on the latency of the preceding instruction(s).
1961 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1962 #ifndef PRODUCT
1963 if (_cfg->C->trace_opto_output())
1964 tty->print("# -> ComputeLocalLatenciesForward\n");
1965 #endif
1967 // Walk over all the schedulable instructions
1968 for( uint j=_bb_start; j < _bb_end; j++ ) {
1970 // This is a kludge, forcing all latency calculations to start at 1.
1971 // Used to allow latency 0 to force an instruction to the beginning
1972 // of the bb
1973 uint latency = 1;
1974 Node *use = bb->get_node(j);
1975 uint nlen = use->len();
1977 // Walk over all the inputs
1978 for ( uint k=0; k < nlen; k++ ) {
1979 Node *def = use->in(k);
1980 if (!def)
1981 continue;
1983 uint l = _node_latency[def->_idx] + use->latency(k);
1984 if (latency < l)
1985 latency = l;
1986 }
1988 _node_latency[use->_idx] = latency;
1990 #ifndef PRODUCT
1991 if (_cfg->C->trace_opto_output()) {
1992 tty->print("# latency %4d: ", latency);
1993 use->dump();
1994 }
1995 #endif
1996 }
1998 #ifndef PRODUCT
1999 if (_cfg->C->trace_opto_output())
2000 tty->print("# <- ComputeLocalLatenciesForward\n");
2001 #endif
2003 } // end ComputeLocalLatenciesForward
2005 // See if this node fits into the present instruction bundle
2006 bool Scheduling::NodeFitsInBundle(Node *n) {
2007 uint n_idx = n->_idx;
2009 // If this is the unconditional delay instruction, then it fits
2010 if (n == _unconditional_delay_slot) {
2011 #ifndef PRODUCT
2012 if (_cfg->C->trace_opto_output())
2013 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
2014 #endif
2015 return (true);
2016 }
2018 // If the node cannot be scheduled this cycle, skip it
2019 if (_current_latency[n_idx] > _bundle_cycle_number) {
2020 #ifndef PRODUCT
2021 if (_cfg->C->trace_opto_output())
2022 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
2023 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
2024 #endif
2025 return (false);
2026 }
2028 const Pipeline *node_pipeline = n->pipeline();
2030 uint instruction_count = node_pipeline->instructionCount();
2031 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2032 instruction_count = 0;
2033 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2034 instruction_count++;
2036 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
2037 #ifndef PRODUCT
2038 if (_cfg->C->trace_opto_output())
2039 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
2040 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
2041 #endif
2042 return (false);
2043 }
2045 // Don't allow non-machine nodes to be handled this way
2046 if (!n->is_Mach() && instruction_count == 0)
2047 return (false);
2049 // See if there is any overlap
2050 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
2052 if (delay > 0) {
2053 #ifndef PRODUCT
2054 if (_cfg->C->trace_opto_output())
2055 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
2056 #endif
2057 return false;
2058 }
2060 #ifndef PRODUCT
2061 if (_cfg->C->trace_opto_output())
2062 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2063 #endif
2065 return true;
2066 }
2068 Node * Scheduling::ChooseNodeToBundle() {
2069 uint siz = _available.size();
2071 if (siz == 0) {
2073 #ifndef PRODUCT
2074 if (_cfg->C->trace_opto_output())
2075 tty->print("# ChooseNodeToBundle: NULL\n");
2076 #endif
2077 return (NULL);
2078 }
2080 // Fast path, if only 1 instruction in the bundle
2081 if (siz == 1) {
2082 #ifndef PRODUCT
2083 if (_cfg->C->trace_opto_output()) {
2084 tty->print("# ChooseNodeToBundle (only 1): ");
2085 _available[0]->dump();
2086 }
2087 #endif
2088 return (_available[0]);
2089 }
2091 // Don't bother, if the bundle is already full
2092 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2093 for ( uint i = 0; i < siz; i++ ) {
2094 Node *n = _available[i];
2096 // Skip projections, we'll handle them another way
2097 if (n->is_Proj())
2098 continue;
2100 // This presupposed that instructions are inserted into the
2101 // available list in a legality order; i.e. instructions that
2102 // must be inserted first are at the head of the list
2103 if (NodeFitsInBundle(n)) {
2104 #ifndef PRODUCT
2105 if (_cfg->C->trace_opto_output()) {
2106 tty->print("# ChooseNodeToBundle: ");
2107 n->dump();
2108 }
2109 #endif
2110 return (n);
2111 }
2112 }
2113 }
2115 // Nothing fits in this bundle, choose the highest priority
2116 #ifndef PRODUCT
2117 if (_cfg->C->trace_opto_output()) {
2118 tty->print("# ChooseNodeToBundle: ");
2119 _available[0]->dump();
2120 }
2121 #endif
2123 return _available[0];
2124 }
2126 void Scheduling::AddNodeToAvailableList(Node *n) {
2127 assert( !n->is_Proj(), "projections never directly made available" );
2128 #ifndef PRODUCT
2129 if (_cfg->C->trace_opto_output()) {
2130 tty->print("# AddNodeToAvailableList: ");
2131 n->dump();
2132 }
2133 #endif
2135 int latency = _current_latency[n->_idx];
2137 // Insert in latency order (insertion sort)
2138 uint i;
2139 for ( i=0; i < _available.size(); i++ )
2140 if (_current_latency[_available[i]->_idx] > latency)
2141 break;
2143 // Special Check for compares following branches
2144 if( n->is_Mach() && _scheduled.size() > 0 ) {
2145 int op = n->as_Mach()->ideal_Opcode();
2146 Node *last = _scheduled[0];
2147 if( last->is_MachIf() && last->in(1) == n &&
2148 ( op == Op_CmpI ||
2149 op == Op_CmpU ||
2150 op == Op_CmpUL ||
2151 op == Op_CmpP ||
2152 op == Op_CmpF ||
2153 op == Op_CmpD ||
2154 op == Op_CmpL ) ) {
2156 // Recalculate position, moving to front of same latency
2157 for ( i=0 ; i < _available.size(); i++ )
2158 if (_current_latency[_available[i]->_idx] >= latency)
2159 break;
2160 }
2161 }
2163 // Insert the node in the available list
2164 _available.insert(i, n);
2166 #ifndef PRODUCT
2167 if (_cfg->C->trace_opto_output())
2168 dump_available();
2169 #endif
2170 }
2172 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2173 for ( uint i=0; i < n->len(); i++ ) {
2174 Node *def = n->in(i);
2175 if (!def) continue;
2176 if( def->is_Proj() ) // If this is a machine projection, then
2177 def = def->in(0); // propagate usage thru to the base instruction
2179 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2180 continue;
2181 }
2183 // Compute the latency
2184 uint l = _bundle_cycle_number + n->latency(i);
2185 if (_current_latency[def->_idx] < l)
2186 _current_latency[def->_idx] = l;
2188 // If this does not have uses then schedule it
2189 if ((--_uses[def->_idx]) == 0)
2190 AddNodeToAvailableList(def);
2191 }
2192 }
2194 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2195 #ifndef PRODUCT
2196 if (_cfg->C->trace_opto_output()) {
2197 tty->print("# AddNodeToBundle: ");
2198 n->dump();
2199 }
2200 #endif
2202 // Remove this from the available list
2203 uint i;
2204 for (i = 0; i < _available.size(); i++)
2205 if (_available[i] == n)
2206 break;
2207 assert(i < _available.size(), "entry in _available list not found");
2208 _available.remove(i);
2210 // See if this fits in the current bundle
2211 const Pipeline *node_pipeline = n->pipeline();
2212 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2214 // Check for instructions to be placed in the delay slot. We
2215 // do this before we actually schedule the current instruction,
2216 // because the delay slot follows the current instruction.
2217 if (Pipeline::_branch_has_delay_slot &&
2218 node_pipeline->hasBranchDelay() &&
2219 !_unconditional_delay_slot) {
2221 uint siz = _available.size();
2223 // Conditional branches can support an instruction that
2224 // is unconditionally executed and not dependent by the
2225 // branch, OR a conditionally executed instruction if
2226 // the branch is taken. In practice, this means that
2227 // the first instruction at the branch target is
2228 // copied to the delay slot, and the branch goes to
2229 // the instruction after that at the branch target
2230 if ( n->is_MachBranch() ) {
2232 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2233 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2235 #ifndef PRODUCT
2236 _branches++;
2237 #endif
2239 // At least 1 instruction is on the available list
2240 // that is not dependent on the branch
2241 for (uint i = 0; i < siz; i++) {
2242 Node *d = _available[i];
2243 const Pipeline *avail_pipeline = d->pipeline();
2245 // Don't allow safepoints in the branch shadow, that will
2246 // cause a number of difficulties
2247 if ( avail_pipeline->instructionCount() == 1 &&
2248 !avail_pipeline->hasMultipleBundles() &&
2249 !avail_pipeline->hasBranchDelay() &&
2250 Pipeline::instr_has_unit_size() &&
2251 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2252 NodeFitsInBundle(d) &&
2253 !node_bundling(d)->used_in_delay()) {
2255 if (d->is_Mach() && !d->is_MachSafePoint()) {
2256 // A node that fits in the delay slot was found, so we need to
2257 // set the appropriate bits in the bundle pipeline information so
2258 // that it correctly indicates resource usage. Later, when we
2259 // attempt to add this instruction to the bundle, we will skip
2260 // setting the resource usage.
2261 _unconditional_delay_slot = d;
2262 node_bundling(n)->set_use_unconditional_delay();
2263 node_bundling(d)->set_used_in_unconditional_delay();
2264 _bundle_use.add_usage(avail_pipeline->resourceUse());
2265 _current_latency[d->_idx] = _bundle_cycle_number;
2266 _next_node = d;
2267 ++_bundle_instr_count;
2268 #ifndef PRODUCT
2269 _unconditional_delays++;
2270 #endif
2271 break;
2272 }
2273 }
2274 }
2275 }
2277 // No delay slot, add a nop to the usage
2278 if (!_unconditional_delay_slot) {
2279 // See if adding an instruction in the delay slot will overflow
2280 // the bundle.
2281 if (!NodeFitsInBundle(_nop)) {
2282 #ifndef PRODUCT
2283 if (_cfg->C->trace_opto_output())
2284 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2285 #endif
2286 step(1);
2287 }
2289 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2290 _next_node = _nop;
2291 ++_bundle_instr_count;
2292 }
2294 // See if the instruction in the delay slot requires a
2295 // step of the bundles
2296 if (!NodeFitsInBundle(n)) {
2297 #ifndef PRODUCT
2298 if (_cfg->C->trace_opto_output())
2299 tty->print("# *** STEP(branch won't fit) ***\n");
2300 #endif
2301 // Update the state information
2302 _bundle_instr_count = 0;
2303 _bundle_cycle_number += 1;
2304 _bundle_use.step(1);
2305 }
2306 }
2308 // Get the number of instructions
2309 uint instruction_count = node_pipeline->instructionCount();
2310 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2311 instruction_count = 0;
2313 // Compute the latency information
2314 uint delay = 0;
2316 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2317 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2318 if (relative_latency < 0)
2319 relative_latency = 0;
2321 delay = _bundle_use.full_latency(relative_latency, node_usage);
2323 // Does not fit in this bundle, start a new one
2324 if (delay > 0) {
2325 step(delay);
2327 #ifndef PRODUCT
2328 if (_cfg->C->trace_opto_output())
2329 tty->print("# *** STEP(%d) ***\n", delay);
2330 #endif
2331 }
2332 }
2334 // If this was placed in the delay slot, ignore it
2335 if (n != _unconditional_delay_slot) {
2337 if (delay == 0) {
2338 if (node_pipeline->hasMultipleBundles()) {
2339 #ifndef PRODUCT
2340 if (_cfg->C->trace_opto_output())
2341 tty->print("# *** STEP(multiple instructions) ***\n");
2342 #endif
2343 step(1);
2344 }
2346 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2347 #ifndef PRODUCT
2348 if (_cfg->C->trace_opto_output())
2349 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2350 instruction_count + _bundle_instr_count,
2351 Pipeline::_max_instrs_per_cycle);
2352 #endif
2353 step(1);
2354 }
2355 }
2357 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2358 _bundle_instr_count++;
2360 // Set the node's latency
2361 _current_latency[n->_idx] = _bundle_cycle_number;
2363 // Now merge the functional unit information
2364 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2365 _bundle_use.add_usage(node_usage);
2367 // Increment the number of instructions in this bundle
2368 _bundle_instr_count += instruction_count;
2370 // Remember this node for later
2371 if (n->is_Mach())
2372 _next_node = n;
2373 }
2375 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2376 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2377 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2378 // into the block. All other scheduled nodes get put in the schedule here.
2379 int op = n->Opcode();
2380 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2381 (op != Op_Node && // Not an unused antidepedence node and
2382 // not an unallocated boxlock
2383 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2385 // Push any trailing projections
2386 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2387 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2388 Node *foi = n->fast_out(i);
2389 if( foi->is_Proj() )
2390 _scheduled.push(foi);
2391 }
2392 }
2394 // Put the instruction in the schedule list
2395 _scheduled.push(n);
2396 }
2398 #ifndef PRODUCT
2399 if (_cfg->C->trace_opto_output())
2400 dump_available();
2401 #endif
2403 // Walk all the definitions, decrementing use counts, and
2404 // if a definition has a 0 use count, place it in the available list.
2405 DecrementUseCounts(n,bb);
2406 }
2408 // This method sets the use count within a basic block. We will ignore all
2409 // uses outside the current basic block. As we are doing a backwards walk,
2410 // any node we reach that has a use count of 0 may be scheduled. This also
2411 // avoids the problem of cyclic references from phi nodes, as long as phi
2412 // nodes are at the front of the basic block. This method also initializes
2413 // the available list to the set of instructions that have no uses within this
2414 // basic block.
2415 void Scheduling::ComputeUseCount(const Block *bb) {
2416 #ifndef PRODUCT
2417 if (_cfg->C->trace_opto_output())
2418 tty->print("# -> ComputeUseCount\n");
2419 #endif
2421 // Clear the list of available and scheduled instructions, just in case
2422 _available.clear();
2423 _scheduled.clear();
2425 // No delay slot specified
2426 _unconditional_delay_slot = NULL;
2428 #ifdef ASSERT
2429 for( uint i=0; i < bb->number_of_nodes(); i++ )
2430 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2431 #endif
2433 // Force the _uses count to never go to zero for unscheduable pieces
2434 // of the block
2435 for( uint k = 0; k < _bb_start; k++ )
2436 _uses[bb->get_node(k)->_idx] = 1;
2437 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2438 _uses[bb->get_node(l)->_idx] = 1;
2440 // Iterate backwards over the instructions in the block. Don't count the
2441 // branch projections at end or the block header instructions.
2442 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2443 Node *n = bb->get_node(j);
2444 if( n->is_Proj() ) continue; // Projections handled another way
2446 // Account for all uses
2447 for ( uint k = 0; k < n->len(); k++ ) {
2448 Node *inp = n->in(k);
2449 if (!inp) continue;
2450 assert(inp != n, "no cycles allowed" );
2451 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2452 if (inp->is_Proj()) { // Skip through Proj's
2453 inp = inp->in(0);
2454 }
2455 ++_uses[inp->_idx]; // Count 1 block-local use
2456 }
2457 }
2459 // If this instruction has a 0 use count, then it is available
2460 if (!_uses[n->_idx]) {
2461 _current_latency[n->_idx] = _bundle_cycle_number;
2462 AddNodeToAvailableList(n);
2463 }
2465 #ifndef PRODUCT
2466 if (_cfg->C->trace_opto_output()) {
2467 tty->print("# uses: %3d: ", _uses[n->_idx]);
2468 n->dump();
2469 }
2470 #endif
2471 }
2473 #ifndef PRODUCT
2474 if (_cfg->C->trace_opto_output())
2475 tty->print("# <- ComputeUseCount\n");
2476 #endif
2477 }
2479 // This routine performs scheduling on each basic block in reverse order,
2480 // using instruction latencies and taking into account function unit
2481 // availability.
2482 void Scheduling::DoScheduling() {
2483 #ifndef PRODUCT
2484 if (_cfg->C->trace_opto_output())
2485 tty->print("# -> DoScheduling\n");
2486 #endif
2488 Block *succ_bb = NULL;
2489 Block *bb;
2491 // Walk over all the basic blocks in reverse order
2492 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2493 bb = _cfg->get_block(i);
2495 #ifndef PRODUCT
2496 if (_cfg->C->trace_opto_output()) {
2497 tty->print("# Schedule BB#%03d (initial)\n", i);
2498 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2499 bb->get_node(j)->dump();
2500 }
2501 }
2502 #endif
2504 // On the head node, skip processing
2505 if (bb == _cfg->get_root_block()) {
2506 continue;
2507 }
2509 // Skip empty, connector blocks
2510 if (bb->is_connector())
2511 continue;
2513 // If the following block is not the sole successor of
2514 // this one, then reset the pipeline information
2515 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2516 #ifndef PRODUCT
2517 if (_cfg->C->trace_opto_output()) {
2518 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2519 _next_node->_idx, _bundle_instr_count);
2520 }
2521 #endif
2522 step_and_clear();
2523 }
2525 // Leave untouched the starting instruction, any Phis, a CreateEx node
2526 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2527 _bb_end = bb->number_of_nodes()-1;
2528 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2529 Node *n = bb->get_node(_bb_start);
2530 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2531 // Also, MachIdealNodes do not get scheduled
2532 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2533 MachNode *mach = n->as_Mach();
2534 int iop = mach->ideal_Opcode();
2535 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2536 if( iop == Op_Con ) continue; // Do not schedule Top
2537 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2538 mach->pipeline() == MachNode::pipeline_class() &&
2539 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2540 continue;
2541 break; // Funny loop structure to be sure...
2542 }
2543 // Compute last "interesting" instruction in block - last instruction we
2544 // might schedule. _bb_end points just after last schedulable inst. We
2545 // normally schedule conditional branches (despite them being forced last
2546 // in the block), because they have delay slots we can fill. Calls all
2547 // have their delay slots filled in the template expansions, so we don't
2548 // bother scheduling them.
2549 Node *last = bb->get_node(_bb_end);
2550 // Ignore trailing NOPs.
2551 while (_bb_end > 0 && last->is_Mach() &&
2552 last->as_Mach()->ideal_Opcode() == Op_Con) {
2553 last = bb->get_node(--_bb_end);
2554 }
2555 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2556 if( last->is_Catch() ||
2557 // Exclude unreachable path case when Halt node is in a separate block.
2558 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2559 // There must be a prior call. Skip it.
2560 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2561 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2562 }
2563 } else if( last->is_MachNullCheck() ) {
2564 // Backup so the last null-checked memory instruction is
2565 // outside the schedulable range. Skip over the nullcheck,
2566 // projection, and the memory nodes.
2567 Node *mem = last->in(1);
2568 do {
2569 _bb_end--;
2570 } while (mem != bb->get_node(_bb_end));
2571 } else {
2572 // Set _bb_end to point after last schedulable inst.
2573 _bb_end++;
2574 }
2576 assert( _bb_start <= _bb_end, "inverted block ends" );
2578 // Compute the register antidependencies for the basic block
2579 ComputeRegisterAntidependencies(bb);
2580 if (_cfg->C->failing()) return; // too many D-U pinch points
2582 // Compute intra-bb latencies for the nodes
2583 ComputeLocalLatenciesForward(bb);
2585 // Compute the usage within the block, and set the list of all nodes
2586 // in the block that have no uses within the block.
2587 ComputeUseCount(bb);
2589 // Schedule the remaining instructions in the block
2590 while ( _available.size() > 0 ) {
2591 Node *n = ChooseNodeToBundle();
2592 guarantee(n != NULL, "no nodes available");
2593 AddNodeToBundle(n,bb);
2594 }
2596 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2597 #ifdef ASSERT
2598 for( uint l = _bb_start; l < _bb_end; l++ ) {
2599 Node *n = bb->get_node(l);
2600 uint m;
2601 for( m = 0; m < _bb_end-_bb_start; m++ )
2602 if( _scheduled[m] == n )
2603 break;
2604 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2605 }
2606 #endif
2608 // Now copy the instructions (in reverse order) back to the block
2609 for ( uint k = _bb_start; k < _bb_end; k++ )
2610 bb->map_node(_scheduled[_bb_end-k-1], k);
2612 #ifndef PRODUCT
2613 if (_cfg->C->trace_opto_output()) {
2614 tty->print("# Schedule BB#%03d (final)\n", i);
2615 uint current = 0;
2616 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2617 Node *n = bb->get_node(j);
2618 if( valid_bundle_info(n) ) {
2619 Bundle *bundle = node_bundling(n);
2620 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2621 tty->print("*** Bundle: ");
2622 bundle->dump();
2623 }
2624 n->dump();
2625 }
2626 }
2627 }
2628 #endif
2629 #ifdef ASSERT
2630 verify_good_schedule(bb,"after block local scheduling");
2631 #endif
2632 }
2634 #ifndef PRODUCT
2635 if (_cfg->C->trace_opto_output())
2636 tty->print("# <- DoScheduling\n");
2637 #endif
2639 // Record final node-bundling array location
2640 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2642 } // end DoScheduling
2644 // Verify that no live-range used in the block is killed in the block by a
2645 // wrong DEF. This doesn't verify live-ranges that span blocks.
2647 // Check for edge existence. Used to avoid adding redundant precedence edges.
2648 static bool edge_from_to( Node *from, Node *to ) {
2649 for( uint i=0; i<from->len(); i++ )
2650 if( from->in(i) == to )
2651 return true;
2652 return false;
2653 }
2655 #ifdef ASSERT
2656 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2657 // Check for bad kills
2658 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2659 Node *prior_use = _reg_node[def];
2660 if( prior_use && !edge_from_to(prior_use,n) ) {
2661 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2662 n->dump();
2663 tty->print_cr("...");
2664 prior_use->dump();
2665 assert(edge_from_to(prior_use,n),msg);
2666 }
2667 _reg_node.map(def,NULL); // Kill live USEs
2668 }
2669 }
2671 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2673 // Zap to something reasonable for the verify code
2674 _reg_node.clear();
2676 // Walk over the block backwards. Check to make sure each DEF doesn't
2677 // kill a live value (other than the one it's supposed to). Add each
2678 // USE to the live set.
2679 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2680 Node *n = b->get_node(i);
2681 int n_op = n->Opcode();
2682 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2683 // Fat-proj kills a slew of registers
2684 RegMask rm = n->out_RegMask();// Make local copy
2685 while( rm.is_NotEmpty() ) {
2686 OptoReg::Name kill = rm.find_first_elem();
2687 rm.Remove(kill);
2688 verify_do_def( n, kill, msg );
2689 }
2690 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2691 // Get DEF'd registers the normal way
2692 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2693 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2694 }
2696 // Now make all USEs live
2697 for( uint i=1; i<n->req(); i++ ) {
2698 Node *def = n->in(i);
2699 assert(def != 0, "input edge required");
2700 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2701 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2702 if( OptoReg::is_valid(reg_lo) ) {
2703 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2704 _reg_node.map(reg_lo,n);
2705 }
2706 if( OptoReg::is_valid(reg_hi) ) {
2707 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2708 _reg_node.map(reg_hi,n);
2709 }
2710 }
2712 }
2714 // Zap to something reasonable for the Antidependence code
2715 _reg_node.clear();
2716 }
2717 #endif
2719 // Conditionally add precedence edges. Avoid putting edges on Projs.
2720 static void add_prec_edge_from_to( Node *from, Node *to ) {
2721 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2722 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2723 from = from->in(0);
2724 }
2725 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2726 !edge_from_to( from, to ) ) // Avoid duplicate edge
2727 from->add_prec(to);
2728 }
2730 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2731 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2732 return;
2734 Node *pinch = _reg_node[def_reg]; // Get pinch point
2735 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2736 is_def ) { // Check for a true def (not a kill)
2737 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2738 return;
2739 }
2741 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2742 debug_only( def = (Node*)0xdeadbeef; )
2744 // After some number of kills there _may_ be a later def
2745 Node *later_def = NULL;
2747 // Finding a kill requires a real pinch-point.
2748 // Check for not already having a pinch-point.
2749 // Pinch points are Op_Node's.
2750 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2751 later_def = pinch; // Must be def/kill as optimistic pinch-point
2752 if ( _pinch_free_list.size() > 0) {
2753 pinch = _pinch_free_list.pop();
2754 } else {
2755 pinch = new (_cfg->C) Node(1); // Pinch point to-be
2756 }
2757 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2758 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2759 return;
2760 }
2761 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2762 _reg_node.map(def_reg,pinch); // Record pinch-point
2763 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2764 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2765 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2766 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2767 later_def = NULL; // and no later def
2768 }
2769 pinch->set_req(0,later_def); // Hook later def so we can find it
2770 } else { // Else have valid pinch point
2771 if( pinch->in(0) ) // If there is a later-def
2772 later_def = pinch->in(0); // Get it
2773 }
2775 // Add output-dependence edge from later def to kill
2776 if( later_def ) // If there is some original def
2777 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2779 // See if current kill is also a use, and so is forced to be the pinch-point.
2780 if( pinch->Opcode() == Op_Node ) {
2781 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2782 for( uint i=1; i<uses->req(); i++ ) {
2783 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2784 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2785 // Yes, found a use/kill pinch-point
2786 pinch->set_req(0,NULL); //
2787 pinch->replace_by(kill); // Move anti-dep edges up
2788 pinch = kill;
2789 _reg_node.map(def_reg,pinch);
2790 return;
2791 }
2792 }
2793 }
2795 // Add edge from kill to pinch-point
2796 add_prec_edge_from_to(kill,pinch);
2797 }
2799 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2800 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2801 return;
2802 Node *pinch = _reg_node[use_reg]; // Get pinch point
2803 // Check for no later def_reg/kill in block
2804 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2805 // Use has to be block-local as well
2806 _cfg->get_block_for_node(use) == b) {
2807 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2808 pinch->req() == 1 ) { // pinch not yet in block?
2809 pinch->del_req(0); // yank pointer to later-def, also set flag
2810 // Insert the pinch-point in the block just after the last use
2811 b->insert_node(pinch, b->find_node(use) + 1);
2812 _bb_end++; // Increase size scheduled region in block
2813 }
2815 add_prec_edge_from_to(pinch,use);
2816 }
2817 }
2819 // We insert antidependences between the reads and following write of
2820 // allocated registers to prevent illegal code motion. Hopefully, the
2821 // number of added references should be fairly small, especially as we
2822 // are only adding references within the current basic block.
2823 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2825 #ifdef ASSERT
2826 verify_good_schedule(b,"before block local scheduling");
2827 #endif
2829 // A valid schedule, for each register independently, is an endless cycle
2830 // of: a def, then some uses (connected to the def by true dependencies),
2831 // then some kills (defs with no uses), finally the cycle repeats with a new
2832 // def. The uses are allowed to float relative to each other, as are the
2833 // kills. No use is allowed to slide past a kill (or def). This requires
2834 // antidependencies between all uses of a single def and all kills that
2835 // follow, up to the next def. More edges are redundant, because later defs
2836 // & kills are already serialized with true or antidependencies. To keep
2837 // the edge count down, we add a 'pinch point' node if there's more than
2838 // one use or more than one kill/def.
2840 // We add dependencies in one bottom-up pass.
2842 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2844 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2845 // register. If not, we record the DEF/KILL in _reg_node, the
2846 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2847 // "pinch point", a new Node that's in the graph but not in the block.
2848 // We put edges from the prior and current DEF/KILLs to the pinch point.
2849 // We put the pinch point in _reg_node. If there's already a pinch point
2850 // we merely add an edge from the current DEF/KILL to the pinch point.
2852 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2853 // put an edge from the pinch point to the USE.
2855 // To be expedient, the _reg_node array is pre-allocated for the whole
2856 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2857 // or a valid def/kill/pinch-point, or a leftover node from some prior
2858 // block. Leftover node from some prior block is treated like a NULL (no
2859 // prior def, so no anti-dependence needed). Valid def is distinguished by
2860 // it being in the current block.
2861 bool fat_proj_seen = false;
2862 uint last_safept = _bb_end-1;
2863 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2864 Node* last_safept_node = end_node;
2865 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2866 Node *n = b->get_node(i);
2867 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2868 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2869 // Fat-proj kills a slew of registers
2870 // This can add edges to 'n' and obscure whether or not it was a def,
2871 // hence the is_def flag.
2872 fat_proj_seen = true;
2873 RegMask rm = n->out_RegMask();// Make local copy
2874 while( rm.is_NotEmpty() ) {
2875 OptoReg::Name kill = rm.find_first_elem();
2876 rm.Remove(kill);
2877 anti_do_def( b, n, kill, is_def );
2878 }
2879 } else {
2880 // Get DEF'd registers the normal way
2881 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2882 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2883 }
2885 // Kill projections on a branch should appear to occur on the
2886 // branch, not afterwards, so grab the masks from the projections
2887 // and process them.
2888 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2889 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2890 Node* use = n->fast_out(i);
2891 if (use->is_Proj()) {
2892 RegMask rm = use->out_RegMask();// Make local copy
2893 while( rm.is_NotEmpty() ) {
2894 OptoReg::Name kill = rm.find_first_elem();
2895 rm.Remove(kill);
2896 anti_do_def( b, n, kill, false );
2897 }
2898 }
2899 }
2900 }
2902 // Check each register used by this instruction for a following DEF/KILL
2903 // that must occur afterward and requires an anti-dependence edge.
2904 for( uint j=0; j<n->req(); j++ ) {
2905 Node *def = n->in(j);
2906 if( def ) {
2907 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2908 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2909 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2910 }
2911 }
2912 // Do not allow defs of new derived values to float above GC
2913 // points unless the base is definitely available at the GC point.
2915 Node *m = b->get_node(i);
2917 // Add precedence edge from following safepoint to use of derived pointer
2918 if( last_safept_node != end_node &&
2919 m != last_safept_node) {
2920 for (uint k = 1; k < m->req(); k++) {
2921 const Type *t = m->in(k)->bottom_type();
2922 if( t->isa_oop_ptr() &&
2923 t->is_ptr()->offset() != 0 ) {
2924 last_safept_node->add_prec( m );
2925 break;
2926 }
2927 }
2928 }
2930 if( n->jvms() ) { // Precedence edge from derived to safept
2931 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2932 if( b->get_node(last_safept) != last_safept_node ) {
2933 last_safept = b->find_node(last_safept_node);
2934 }
2935 for( uint j=last_safept; j > i; j-- ) {
2936 Node *mach = b->get_node(j);
2937 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2938 mach->add_prec( n );
2939 }
2940 last_safept = i;
2941 last_safept_node = m;
2942 }
2943 }
2945 if (fat_proj_seen) {
2946 // Garbage collect pinch nodes that were not consumed.
2947 // They are usually created by a fat kill MachProj for a call.
2948 garbage_collect_pinch_nodes();
2949 }
2950 }
2952 // Garbage collect pinch nodes for reuse by other blocks.
2953 //
2954 // The block scheduler's insertion of anti-dependence
2955 // edges creates many pinch nodes when the block contains
2956 // 2 or more Calls. A pinch node is used to prevent a
2957 // combinatorial explosion of edges. If a set of kills for a
2958 // register is anti-dependent on a set of uses (or defs), rather
2959 // than adding an edge in the graph between each pair of kill
2960 // and use (or def), a pinch is inserted between them:
2961 //
2962 // use1 use2 use3
2963 // \ | /
2964 // \ | /
2965 // pinch
2966 // / | \
2967 // / | \
2968 // kill1 kill2 kill3
2969 //
2970 // One pinch node is created per register killed when
2971 // the second call is encountered during a backwards pass
2972 // over the block. Most of these pinch nodes are never
2973 // wired into the graph because the register is never
2974 // used or def'ed in the block.
2975 //
2976 void Scheduling::garbage_collect_pinch_nodes() {
2977 #ifndef PRODUCT
2978 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2979 #endif
2980 int trace_cnt = 0;
2981 for (uint k = 0; k < _reg_node.Size(); k++) {
2982 Node* pinch = _reg_node[k];
2983 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2984 // no predecence input edges
2985 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2986 cleanup_pinch(pinch);
2987 _pinch_free_list.push(pinch);
2988 _reg_node.map(k, NULL);
2989 #ifndef PRODUCT
2990 if (_cfg->C->trace_opto_output()) {
2991 trace_cnt++;
2992 if (trace_cnt > 40) {
2993 tty->print("\n");
2994 trace_cnt = 0;
2995 }
2996 tty->print(" %d", pinch->_idx);
2997 }
2998 #endif
2999 }
3000 }
3001 #ifndef PRODUCT
3002 if (_cfg->C->trace_opto_output()) tty->print("\n");
3003 #endif
3004 }
3006 // Clean up a pinch node for reuse.
3007 void Scheduling::cleanup_pinch( Node *pinch ) {
3008 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
3010 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
3011 Node* use = pinch->last_out(i);
3012 uint uses_found = 0;
3013 for (uint j = use->req(); j < use->len(); j++) {
3014 if (use->in(j) == pinch) {
3015 use->rm_prec(j);
3016 uses_found++;
3017 }
3018 }
3019 assert(uses_found > 0, "must be a precedence edge");
3020 i -= uses_found; // we deleted 1 or more copies of this edge
3021 }
3022 // May have a later_def entry
3023 pinch->set_req(0, NULL);
3024 }
3026 #ifndef PRODUCT
3028 void Scheduling::dump_available() const {
3029 tty->print("#Availist ");
3030 for (uint i = 0; i < _available.size(); i++)
3031 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
3032 tty->cr();
3033 }
3035 // Print Scheduling Statistics
3036 void Scheduling::print_statistics() {
3037 // Print the size added by nops for bundling
3038 tty->print("Nops added %d bytes to total of %d bytes",
3039 _total_nop_size, _total_method_size);
3040 if (_total_method_size > 0)
3041 tty->print(", for %.2f%%",
3042 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
3043 tty->print("\n");
3045 // Print the number of branch shadows filled
3046 if (Pipeline::_branch_has_delay_slot) {
3047 tty->print("Of %d branches, %d had unconditional delay slots filled",
3048 _total_branches, _total_unconditional_delays);
3049 if (_total_branches > 0)
3050 tty->print(", for %.2f%%",
3051 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
3052 tty->print("\n");
3053 }
3055 uint total_instructions = 0, total_bundles = 0;
3057 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
3058 uint bundle_count = _total_instructions_per_bundle[i];
3059 total_instructions += bundle_count * i;
3060 total_bundles += bundle_count;
3061 }
3063 if (total_bundles > 0)
3064 tty->print("Average ILP (excluding nops) is %.2f\n",
3065 ((double)total_instructions) / ((double)total_bundles));
3066 }
3067 #endif