Mon, 23 Jun 2014 13:33:23 +0200
8046289: compiler/6340864/TestLongVect.java timeout with
Reviewed-by: iveresov, vlivanov
1 /*
2 * Copyright (c) 1998, 2013, 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 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/output.hpp"
38 #include "opto/regalloc.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/type.hpp"
42 #include "runtime/handles.inline.hpp"
43 #include "utilities/xmlstream.hpp"
45 #ifndef PRODUCT
46 #define DEBUG_ARG(x) , x
47 #else
48 #define DEBUG_ARG(x)
49 #endif
51 // Convert Nodes to instruction bits and pass off to the VM
52 void Compile::Output() {
53 // RootNode goes
54 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
56 // The number of new nodes (mostly MachNop) is proportional to
57 // the number of java calls and inner loops which are aligned.
58 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
59 C->inner_loops()*(OptoLoopAlignment-1)),
60 "out of nodes before code generation" ) ) {
61 return;
62 }
63 // Make sure I can find the Start Node
64 Block *entry = _cfg->get_block(1);
65 Block *broot = _cfg->get_root_block();
67 const StartNode *start = entry->head()->as_Start();
69 // Replace StartNode with prolog
70 MachPrologNode *prolog = new (this) MachPrologNode();
71 entry->map_node(prolog, 0);
72 _cfg->map_node_to_block(prolog, entry);
73 _cfg->unmap_node_from_block(start); // start is no longer in any block
75 // Virtual methods need an unverified entry point
77 if( is_osr_compilation() ) {
78 if( PoisonOSREntry ) {
79 // TODO: Should use a ShouldNotReachHereNode...
80 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
81 }
82 } else {
83 if( _method && !_method->flags().is_static() ) {
84 // Insert unvalidated entry point
85 _cfg->insert( broot, 0, new (this) MachUEPNode() );
86 }
88 }
91 // Break before main entry point
92 if( (_method && _method->break_at_execute())
93 #ifndef PRODUCT
94 ||(OptoBreakpoint && is_method_compilation())
95 ||(OptoBreakpointOSR && is_osr_compilation())
96 ||(OptoBreakpointC2R && !_method)
97 #endif
98 ) {
99 // checking for _method means that OptoBreakpoint does not apply to
100 // runtime stubs or frame converters
101 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
102 }
104 // Insert epilogs before every return
105 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
106 Block* block = _cfg->get_block(i);
107 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
108 Node* m = block->end();
109 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
110 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
111 block->add_inst(epilog);
112 _cfg->map_node_to_block(epilog, block);
113 }
114 }
115 }
117 # ifdef ENABLE_ZAP_DEAD_LOCALS
118 if (ZapDeadCompiledLocals) {
119 Insert_zap_nodes();
120 }
121 # endif
123 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
124 blk_starts[0] = 0;
126 // Initialize code buffer and process short branches.
127 CodeBuffer* cb = init_buffer(blk_starts);
129 if (cb == NULL || failing()) {
130 return;
131 }
133 ScheduleAndBundle();
135 #ifndef PRODUCT
136 if (trace_opto_output()) {
137 tty->print("\n---- After ScheduleAndBundle ----\n");
138 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
139 tty->print("\nBB#%03d:\n", i);
140 Block* block = _cfg->get_block(i);
141 for (uint j = 0; j < block->number_of_nodes(); j++) {
142 Node* n = block->get_node(j);
143 OptoReg::Name reg = _regalloc->get_reg_first(n);
144 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
145 n->dump();
146 }
147 }
148 }
149 #endif
151 if (failing()) {
152 return;
153 }
155 BuildOopMaps();
157 if (failing()) {
158 return;
159 }
161 fill_buffer(cb, blk_starts);
162 }
164 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
165 // Determine if we need to generate a stack overflow check.
166 // Do it if the method is not a stub function and
167 // has java calls or has frame size > vm_page_size/8.
168 // The debug VM checks that deoptimization doesn't trigger an
169 // unexpected stack overflow (compiled method stack banging should
170 // guarantee it doesn't happen) so we always need the stack bang in
171 // a debug VM.
172 return (UseStackBanging && stub_function() == NULL &&
173 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
174 DEBUG_ONLY(|| true)));
175 }
177 bool Compile::need_register_stack_bang() const {
178 // Determine if we need to generate a register stack overflow check.
179 // This is only used on architectures which have split register
180 // and memory stacks (ie. IA64).
181 // Bang if the method is not a stub function and has java calls
182 return (stub_function() == NULL && has_java_calls());
183 }
185 # ifdef ENABLE_ZAP_DEAD_LOCALS
188 // In order to catch compiler oop-map bugs, we have implemented
189 // a debugging mode called ZapDeadCompilerLocals.
190 // This mode causes the compiler to insert a call to a runtime routine,
191 // "zap_dead_locals", right before each place in compiled code
192 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
193 // The runtime routine checks that locations mapped as oops are really
194 // oops, that locations mapped as values do not look like oops,
195 // and that locations mapped as dead are not used later
196 // (by zapping them to an invalid address).
198 int Compile::_CompiledZap_count = 0;
200 void Compile::Insert_zap_nodes() {
201 bool skip = false;
204 // Dink with static counts because code code without the extra
205 // runtime calls is MUCH faster for debugging purposes
207 if ( CompileZapFirst == 0 ) ; // nothing special
208 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
209 else if ( CompileZapFirst == CompiledZap_count() )
210 warning("starting zap compilation after skipping");
212 if ( CompileZapLast == -1 ) ; // nothing special
213 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
214 else if ( CompileZapLast == CompiledZap_count() )
215 warning("about to compile last zap");
217 ++_CompiledZap_count; // counts skipped zaps, too
219 if ( skip ) return;
222 if ( _method == NULL )
223 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
225 // Insert call to zap runtime stub before every node with an oop map
226 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
227 Block *b = _cfg->get_block(i);
228 for ( uint j = 0; j < b->number_of_nodes(); ++j ) {
229 Node *n = b->get_node(j);
231 // Determining if we should insert a zap-a-lot node in output.
232 // We do that for all nodes that has oopmap info, except for calls
233 // to allocation. Calls to allocation passes in the old top-of-eden pointer
234 // and expect the C code to reset it. Hence, there can be no safepoints between
235 // the inlined-allocation and the call to new_Java, etc.
236 // We also cannot zap monitor calls, as they must hold the microlock
237 // during the call to Zap, which also wants to grab the microlock.
238 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
239 if ( insert ) { // it is MachSafePoint
240 if ( !n->is_MachCall() ) {
241 insert = false;
242 } else if ( n->is_MachCall() ) {
243 MachCallNode* call = n->as_MachCall();
244 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
245 call->entry_point() == OptoRuntime::new_array_Java() ||
246 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
247 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
248 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
249 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
250 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
251 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
252 ) {
253 insert = false;
254 }
255 }
256 if (insert) {
257 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
258 b->insert_node(zap, j);
259 _cfg->map_node_to_block(zap, b);
260 ++j;
261 }
262 }
263 }
264 }
265 }
268 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
269 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
270 CallStaticJavaNode* ideal_node =
271 new (this) CallStaticJavaNode( tf,
272 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
273 "call zap dead locals stub", 0, TypePtr::BOTTOM);
274 // We need to copy the OopMap from the site we're zapping at.
275 // We have to make a copy, because the zap site might not be
276 // a call site, and zap_dead is a call site.
277 OopMap* clone = node_to_check->oop_map()->deep_copy();
279 // Add the cloned OopMap to the zap node
280 ideal_node->set_oop_map(clone);
281 return _matcher->match_sfpt(ideal_node);
282 }
284 bool Compile::is_node_getting_a_safepoint( Node* n) {
285 // This code duplicates the logic prior to the call of add_safepoint
286 // below in this file.
287 if( n->is_MachSafePoint() ) return true;
288 return false;
289 }
291 # endif // ENABLE_ZAP_DEAD_LOCALS
293 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
294 // of a loop. When aligning a loop we need to provide enough instructions
295 // in cpu's fetch buffer to feed decoders. The loop alignment could be
296 // avoided if we have enough instructions in fetch buffer at the head of a loop.
297 // By default, the size is set to 999999 by Block's constructor so that
298 // a loop will be aligned if the size is not reset here.
299 //
300 // Note: Mach instructions could contain several HW instructions
301 // so the size is estimated only.
302 //
303 void Compile::compute_loop_first_inst_sizes() {
304 // The next condition is used to gate the loop alignment optimization.
305 // Don't aligned a loop if there are enough instructions at the head of a loop
306 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
307 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
308 // equal to 11 bytes which is the largest address NOP instruction.
309 if (MaxLoopPad < OptoLoopAlignment - 1) {
310 uint last_block = _cfg->number_of_blocks() - 1;
311 for (uint i = 1; i <= last_block; i++) {
312 Block* block = _cfg->get_block(i);
313 // Check the first loop's block which requires an alignment.
314 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
315 uint sum_size = 0;
316 uint inst_cnt = NumberOfLoopInstrToAlign;
317 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
319 // Check subsequent fallthrough blocks if the loop's first
320 // block(s) does not have enough instructions.
321 Block *nb = block;
322 while(inst_cnt > 0 &&
323 i < last_block &&
324 !_cfg->get_block(i + 1)->has_loop_alignment() &&
325 !nb->has_successor(block)) {
326 i++;
327 nb = _cfg->get_block(i);
328 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
329 } // while( inst_cnt > 0 && i < last_block )
331 block->set_first_inst_size(sum_size);
332 } // f( b->head()->is_Loop() )
333 } // for( i <= last_block )
334 } // if( MaxLoopPad < OptoLoopAlignment-1 )
335 }
337 // The architecture description provides short branch variants for some long
338 // branch instructions. Replace eligible long branches with short branches.
339 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
340 // Compute size of each block, method size, and relocation information size
341 uint nblocks = _cfg->number_of_blocks();
343 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
344 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
345 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
347 // Collect worst case block paddings
348 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
349 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
351 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
352 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
354 bool has_short_branch_candidate = false;
356 // Initialize the sizes to 0
357 code_size = 0; // Size in bytes of generated code
358 stub_size = 0; // Size in bytes of all stub entries
359 // Size in bytes of all relocation entries, including those in local stubs.
360 // Start with 2-bytes of reloc info for the unvalidated entry point
361 reloc_size = 1; // Number of relocation entries
363 // Make three passes. The first computes pessimistic blk_starts,
364 // relative jmp_offset and reloc_size information. The second performs
365 // short branch substitution using the pessimistic sizing. The
366 // third inserts nops where needed.
368 // Step one, perform a pessimistic sizing pass.
369 uint last_call_adr = max_uint;
370 uint last_avoid_back_to_back_adr = max_uint;
371 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
372 for (uint i = 0; i < nblocks; i++) { // For all blocks
373 Block* block = _cfg->get_block(i);
375 // During short branch replacement, we store the relative (to blk_starts)
376 // offset of jump in jmp_offset, rather than the absolute offset of jump.
377 // This is so that we do not need to recompute sizes of all nodes when
378 // we compute correct blk_starts in our next sizing pass.
379 jmp_offset[i] = 0;
380 jmp_size[i] = 0;
381 jmp_nidx[i] = -1;
382 DEBUG_ONLY( jmp_target[i] = 0; )
383 DEBUG_ONLY( jmp_rule[i] = 0; )
385 // Sum all instruction sizes to compute block size
386 uint last_inst = block->number_of_nodes();
387 uint blk_size = 0;
388 for (uint j = 0; j < last_inst; j++) {
389 Node* nj = block->get_node(j);
390 // Handle machine instruction nodes
391 if (nj->is_Mach()) {
392 MachNode *mach = nj->as_Mach();
393 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
394 reloc_size += mach->reloc();
395 if (mach->is_MachCall()) {
396 // add size information for trampoline stub
397 // class CallStubImpl is platform-specific and defined in the *.ad files.
398 stub_size += CallStubImpl::size_call_trampoline();
399 reloc_size += CallStubImpl::reloc_call_trampoline();
401 MachCallNode *mcall = mach->as_MachCall();
402 // This destination address is NOT PC-relative
404 mcall->method_set((intptr_t)mcall->entry_point());
406 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
407 stub_size += CompiledStaticCall::to_interp_stub_size();
408 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
409 }
410 } else if (mach->is_MachSafePoint()) {
411 // If call/safepoint are adjacent, account for possible
412 // nop to disambiguate the two safepoints.
413 // ScheduleAndBundle() can rearrange nodes in a block,
414 // check for all offsets inside this block.
415 if (last_call_adr >= blk_starts[i]) {
416 blk_size += nop_size;
417 }
418 }
419 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
420 // Nop is inserted between "avoid back to back" instructions.
421 // ScheduleAndBundle() can rearrange nodes in a block,
422 // check for all offsets inside this block.
423 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
424 blk_size += nop_size;
425 }
426 }
427 if (mach->may_be_short_branch()) {
428 if (!nj->is_MachBranch()) {
429 #ifndef PRODUCT
430 nj->dump(3);
431 #endif
432 Unimplemented();
433 }
434 assert(jmp_nidx[i] == -1, "block should have only one branch");
435 jmp_offset[i] = blk_size;
436 jmp_size[i] = nj->size(_regalloc);
437 jmp_nidx[i] = j;
438 has_short_branch_candidate = true;
439 }
440 }
441 blk_size += nj->size(_regalloc);
442 // Remember end of call offset
443 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
444 last_call_adr = blk_starts[i]+blk_size;
445 }
446 // Remember end of avoid_back_to_back offset
447 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
448 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
449 }
450 }
452 // When the next block starts a loop, we may insert pad NOP
453 // instructions. Since we cannot know our future alignment,
454 // assume the worst.
455 if (i < nblocks - 1) {
456 Block* nb = _cfg->get_block(i + 1);
457 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
458 if (max_loop_pad > 0) {
459 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
460 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
461 // If either is the last instruction in this block, bump by
462 // max_loop_pad in lock-step with blk_size, so sizing
463 // calculations in subsequent blocks still can conservatively
464 // detect that it may the last instruction in this block.
465 if (last_call_adr == blk_starts[i]+blk_size) {
466 last_call_adr += max_loop_pad;
467 }
468 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
469 last_avoid_back_to_back_adr += max_loop_pad;
470 }
471 blk_size += max_loop_pad;
472 block_worst_case_pad[i + 1] = max_loop_pad;
473 }
474 }
476 // Save block size; update total method size
477 blk_starts[i+1] = blk_starts[i]+blk_size;
478 }
480 // Step two, replace eligible long jumps.
481 bool progress = true;
482 uint last_may_be_short_branch_adr = max_uint;
483 while (has_short_branch_candidate && progress) {
484 progress = false;
485 has_short_branch_candidate = false;
486 int adjust_block_start = 0;
487 for (uint i = 0; i < nblocks; i++) {
488 Block* block = _cfg->get_block(i);
489 int idx = jmp_nidx[i];
490 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
491 if (mach != NULL && mach->may_be_short_branch()) {
492 #ifdef ASSERT
493 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
494 int j;
495 // Find the branch; ignore trailing NOPs.
496 for (j = block->number_of_nodes()-1; j>=0; j--) {
497 Node* n = block->get_node(j);
498 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
499 break;
500 }
501 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
502 #endif
503 int br_size = jmp_size[i];
504 int br_offs = blk_starts[i] + jmp_offset[i];
506 // This requires the TRUE branch target be in succs[0]
507 uint bnum = block->non_connector_successor(0)->_pre_order;
508 int offset = blk_starts[bnum] - br_offs;
509 if (bnum > i) { // adjust following block's offset
510 offset -= adjust_block_start;
511 }
513 // This block can be a loop header, account for the padding
514 // in the previous block.
515 int block_padding = block_worst_case_pad[i];
516 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
517 // In the following code a nop could be inserted before
518 // the branch which will increase the backward distance.
519 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
520 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
522 if (needs_padding && offset <= 0)
523 offset -= nop_size;
525 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
526 // We've got a winner. Replace this branch.
527 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
529 // Update the jmp_size.
530 int new_size = replacement->size(_regalloc);
531 int diff = br_size - new_size;
532 assert(diff >= (int)nop_size, "short_branch size should be smaller");
533 // Conservatively take into account padding between
534 // avoid_back_to_back branches. Previous branch could be
535 // converted into avoid_back_to_back branch during next
536 // rounds.
537 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
538 jmp_offset[i] += nop_size;
539 diff -= nop_size;
540 }
541 adjust_block_start += diff;
542 block->map_node(replacement, idx);
543 mach->subsume_by(replacement, C);
544 mach = replacement;
545 progress = true;
547 jmp_size[i] = new_size;
548 DEBUG_ONLY( jmp_target[i] = bnum; );
549 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
550 } else {
551 // The jump distance is not short, try again during next iteration.
552 has_short_branch_candidate = true;
553 }
554 } // (mach->may_be_short_branch())
555 if (mach != NULL && (mach->may_be_short_branch() ||
556 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
557 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
558 }
559 blk_starts[i+1] -= adjust_block_start;
560 }
561 }
563 #ifdef ASSERT
564 for (uint i = 0; i < nblocks; i++) { // For all blocks
565 if (jmp_target[i] != 0) {
566 int br_size = jmp_size[i];
567 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
568 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
569 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]);
570 }
571 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
572 }
573 }
574 #endif
576 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
577 // after ScheduleAndBundle().
579 // ------------------
580 // Compute size for code buffer
581 code_size = blk_starts[nblocks];
583 // Relocation records
584 reloc_size += 1; // Relo entry for exception handler
586 // Adjust reloc_size to number of record of relocation info
587 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
588 // a relocation index.
589 // The CodeBuffer will expand the locs array if this estimate is too low.
590 reloc_size *= 10 / sizeof(relocInfo);
591 }
593 //------------------------------FillLocArray-----------------------------------
594 // Create a bit of debug info and append it to the array. The mapping is from
595 // Java local or expression stack to constant, register or stack-slot. For
596 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
597 // entry has been taken care of and caller should skip it).
598 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
599 // This should never have accepted Bad before
600 assert(OptoReg::is_valid(regnum), "location must be valid");
601 return (OptoReg::is_reg(regnum))
602 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
603 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
604 }
607 ObjectValue*
608 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
609 for (int i = 0; i < objs->length(); i++) {
610 assert(objs->at(i)->is_object(), "corrupt object cache");
611 ObjectValue* sv = (ObjectValue*) objs->at(i);
612 if (sv->id() == id) {
613 return sv;
614 }
615 }
616 // Otherwise..
617 return NULL;
618 }
620 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
621 ObjectValue* sv ) {
622 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
623 objs->append(sv);
624 }
627 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
628 GrowableArray<ScopeValue*> *array,
629 GrowableArray<ScopeValue*> *objs ) {
630 assert( local, "use _top instead of null" );
631 if (array->length() != idx) {
632 assert(array->length() == idx + 1, "Unexpected array count");
633 // Old functionality:
634 // return
635 // New functionality:
636 // Assert if the local is not top. In product mode let the new node
637 // override the old entry.
638 assert(local == top(), "LocArray collision");
639 if (local == top()) {
640 return;
641 }
642 array->pop();
643 }
644 const Type *t = local->bottom_type();
646 // Is it a safepoint scalar object node?
647 if (local->is_SafePointScalarObject()) {
648 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
650 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
651 if (sv == NULL) {
652 ciKlass* cik = t->is_oopptr()->klass();
653 assert(cik->is_instance_klass() ||
654 cik->is_array_klass(), "Not supported allocation.");
655 sv = new ObjectValue(spobj->_idx,
656 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
657 Compile::set_sv_for_object_node(objs, sv);
659 uint first_ind = spobj->first_index(sfpt->jvms());
660 for (uint i = 0; i < spobj->n_fields(); i++) {
661 Node* fld_node = sfpt->in(first_ind+i);
662 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
663 }
664 }
665 array->append(sv);
666 return;
667 }
669 // Grab the register number for the local
670 OptoReg::Name regnum = _regalloc->get_reg_first(local);
671 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
672 // Record the double as two float registers.
673 // The register mask for such a value always specifies two adjacent
674 // float registers, with the lower register number even.
675 // Normally, the allocation of high and low words to these registers
676 // is irrelevant, because nearly all operations on register pairs
677 // (e.g., StoreD) treat them as a single unit.
678 // Here, we assume in addition that the words in these two registers
679 // stored "naturally" (by operations like StoreD and double stores
680 // within the interpreter) such that the lower-numbered register
681 // is written to the lower memory address. This may seem like
682 // a machine dependency, but it is not--it is a requirement on
683 // the author of the <arch>.ad file to ensure that, for every
684 // even/odd double-register pair to which a double may be allocated,
685 // the word in the even single-register is stored to the first
686 // memory word. (Note that register numbers are completely
687 // arbitrary, and are not tied to any machine-level encodings.)
688 #ifdef _LP64
689 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
690 array->append(new ConstantIntValue(0));
691 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
692 } else if ( t->base() == Type::Long ) {
693 array->append(new ConstantIntValue(0));
694 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
695 } else if ( t->base() == Type::RawPtr ) {
696 // jsr/ret return address which must be restored into a the full
697 // width 64-bit stack slot.
698 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
699 }
700 #else //_LP64
701 #ifdef SPARC
702 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
703 // For SPARC we have to swap high and low words for
704 // long values stored in a single-register (g0-g7).
705 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
706 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
707 } else
708 #endif //SPARC
709 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
710 // Repack the double/long as two jints.
711 // The convention the interpreter uses is that the second local
712 // holds the first raw word of the native double representation.
713 // This is actually reasonable, since locals and stack arrays
714 // grow downwards in all implementations.
715 // (If, on some machine, the interpreter's Java locals or stack
716 // were to grow upwards, the embedded doubles would be word-swapped.)
717 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
718 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
719 }
720 #endif //_LP64
721 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
722 OptoReg::is_reg(regnum) ) {
723 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
724 ? Location::float_in_dbl : Location::normal ));
725 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
726 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
727 ? Location::int_in_long : Location::normal ));
728 } else if( t->base() == Type::NarrowOop ) {
729 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
730 } else {
731 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
732 }
733 return;
734 }
736 // No register. It must be constant data.
737 switch (t->base()) {
738 case Type::Half: // Second half of a double
739 ShouldNotReachHere(); // Caller should skip 2nd halves
740 break;
741 case Type::AnyPtr:
742 array->append(new ConstantOopWriteValue(NULL));
743 break;
744 case Type::AryPtr:
745 case Type::InstPtr: // fall through
746 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
747 break;
748 case Type::NarrowOop:
749 if (t == TypeNarrowOop::NULL_PTR) {
750 array->append(new ConstantOopWriteValue(NULL));
751 } else {
752 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
753 }
754 break;
755 case Type::Int:
756 array->append(new ConstantIntValue(t->is_int()->get_con()));
757 break;
758 case Type::RawPtr:
759 // A return address (T_ADDRESS).
760 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
761 #ifdef _LP64
762 // Must be restored to the full-width 64-bit stack slot.
763 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
764 #else
765 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
766 #endif
767 break;
768 case Type::FloatCon: {
769 float f = t->is_float_constant()->getf();
770 array->append(new ConstantIntValue(jint_cast(f)));
771 break;
772 }
773 case Type::DoubleCon: {
774 jdouble d = t->is_double_constant()->getd();
775 #ifdef _LP64
776 array->append(new ConstantIntValue(0));
777 array->append(new ConstantDoubleValue(d));
778 #else
779 // Repack the double as two jints.
780 // The convention the interpreter uses is that the second local
781 // holds the first raw word of the native double representation.
782 // This is actually reasonable, since locals and stack arrays
783 // grow downwards in all implementations.
784 // (If, on some machine, the interpreter's Java locals or stack
785 // were to grow upwards, the embedded doubles would be word-swapped.)
786 jlong_accessor acc;
787 acc.long_value = jlong_cast(d);
788 array->append(new ConstantIntValue(acc.words[1]));
789 array->append(new ConstantIntValue(acc.words[0]));
790 #endif
791 break;
792 }
793 case Type::Long: {
794 jlong d = t->is_long()->get_con();
795 #ifdef _LP64
796 array->append(new ConstantIntValue(0));
797 array->append(new ConstantLongValue(d));
798 #else
799 // Repack the long as two jints.
800 // The convention the interpreter uses is that the second local
801 // holds the first raw word of the native double representation.
802 // This is actually reasonable, since locals and stack arrays
803 // grow downwards in all implementations.
804 // (If, on some machine, the interpreter's Java locals or stack
805 // were to grow upwards, the embedded doubles would be word-swapped.)
806 jlong_accessor acc;
807 acc.long_value = d;
808 array->append(new ConstantIntValue(acc.words[1]));
809 array->append(new ConstantIntValue(acc.words[0]));
810 #endif
811 break;
812 }
813 case Type::Top: // Add an illegal value here
814 array->append(new LocationValue(Location()));
815 break;
816 default:
817 ShouldNotReachHere();
818 break;
819 }
820 }
822 // Determine if this node starts a bundle
823 bool Compile::starts_bundle(const Node *n) const {
824 return (_node_bundling_limit > n->_idx &&
825 _node_bundling_base[n->_idx].starts_bundle());
826 }
828 //--------------------------Process_OopMap_Node--------------------------------
829 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
831 // Handle special safepoint nodes for synchronization
832 MachSafePointNode *sfn = mach->as_MachSafePoint();
833 MachCallNode *mcall;
835 #ifdef ENABLE_ZAP_DEAD_LOCALS
836 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
837 #endif
839 int safepoint_pc_offset = current_offset;
840 bool is_method_handle_invoke = false;
841 bool return_oop = false;
843 // Add the safepoint in the DebugInfoRecorder
844 if( !mach->is_MachCall() ) {
845 mcall = NULL;
846 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
847 } else {
848 mcall = mach->as_MachCall();
850 // Is the call a MethodHandle call?
851 if (mcall->is_MachCallJava()) {
852 if (mcall->as_MachCallJava()->_method_handle_invoke) {
853 assert(has_method_handle_invokes(), "must have been set during call generation");
854 is_method_handle_invoke = true;
855 }
856 }
858 // Check if a call returns an object.
859 if (mcall->return_value_is_used() &&
860 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
861 return_oop = true;
862 }
863 safepoint_pc_offset += mcall->ret_addr_offset();
864 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
865 }
867 // Loop over the JVMState list to add scope information
868 // Do not skip safepoints with a NULL method, they need monitor info
869 JVMState* youngest_jvms = sfn->jvms();
870 int max_depth = youngest_jvms->depth();
872 // Allocate the object pool for scalar-replaced objects -- the map from
873 // small-integer keys (which can be recorded in the local and ostack
874 // arrays) to descriptions of the object state.
875 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
877 // Visit scopes from oldest to youngest.
878 for (int depth = 1; depth <= max_depth; depth++) {
879 JVMState* jvms = youngest_jvms->of_depth(depth);
880 int idx;
881 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
882 // Safepoints that do not have method() set only provide oop-map and monitor info
883 // to support GC; these do not support deoptimization.
884 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
885 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
886 int num_mon = jvms->nof_monitors();
887 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
888 "JVMS local count must match that of the method");
890 // Add Local and Expression Stack Information
892 // Insert locals into the locarray
893 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
894 for( idx = 0; idx < num_locs; idx++ ) {
895 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
896 }
898 // Insert expression stack entries into the exparray
899 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
900 for( idx = 0; idx < num_exps; idx++ ) {
901 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
902 }
904 // Add in mappings of the monitors
905 assert( !method ||
906 !method->is_synchronized() ||
907 method->is_native() ||
908 num_mon > 0 ||
909 !GenerateSynchronizationCode,
910 "monitors must always exist for synchronized methods");
912 // Build the growable array of ScopeValues for exp stack
913 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
915 // Loop over monitors and insert into array
916 for (idx = 0; idx < num_mon; idx++) {
917 // Grab the node that defines this monitor
918 Node* box_node = sfn->monitor_box(jvms, idx);
919 Node* obj_node = sfn->monitor_obj(jvms, idx);
921 // Create ScopeValue for object
922 ScopeValue *scval = NULL;
924 if (obj_node->is_SafePointScalarObject()) {
925 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
926 scval = Compile::sv_for_node_id(objs, spobj->_idx);
927 if (scval == NULL) {
928 const Type *t = spobj->bottom_type();
929 ciKlass* cik = t->is_oopptr()->klass();
930 assert(cik->is_instance_klass() ||
931 cik->is_array_klass(), "Not supported allocation.");
932 ObjectValue* sv = new ObjectValue(spobj->_idx,
933 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
934 Compile::set_sv_for_object_node(objs, sv);
936 uint first_ind = spobj->first_index(youngest_jvms);
937 for (uint i = 0; i < spobj->n_fields(); i++) {
938 Node* fld_node = sfn->in(first_ind+i);
939 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
940 }
941 scval = sv;
942 }
943 } else if (!obj_node->is_Con()) {
944 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
945 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
946 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
947 } else {
948 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
949 }
950 } else {
951 const TypePtr *tp = obj_node->get_ptr_type();
952 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
953 }
955 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
956 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
957 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
958 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
959 }
961 // We dump the object pool first, since deoptimization reads it in first.
962 debug_info()->dump_object_pool(objs);
964 // Build first class objects to pass to scope
965 DebugToken *locvals = debug_info()->create_scope_values(locarray);
966 DebugToken *expvals = debug_info()->create_scope_values(exparray);
967 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
969 // Make method available for all Safepoints
970 ciMethod* scope_method = method ? method : _method;
971 // Describe the scope here
972 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
973 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
974 // Now we can describe the scope.
975 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
976 } // End jvms loop
978 // Mark the end of the scope set.
979 debug_info()->end_safepoint(safepoint_pc_offset);
980 }
984 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
985 class NonSafepointEmitter {
986 Compile* C;
987 JVMState* _pending_jvms;
988 int _pending_offset;
990 void emit_non_safepoint();
992 public:
993 NonSafepointEmitter(Compile* compile) {
994 this->C = compile;
995 _pending_jvms = NULL;
996 _pending_offset = 0;
997 }
999 void observe_instruction(Node* n, int pc_offset) {
1000 if (!C->debug_info()->recording_non_safepoints()) return;
1002 Node_Notes* nn = C->node_notes_at(n->_idx);
1003 if (nn == NULL || nn->jvms() == NULL) return;
1004 if (_pending_jvms != NULL &&
1005 _pending_jvms->same_calls_as(nn->jvms())) {
1006 // Repeated JVMS? Stretch it up here.
1007 _pending_offset = pc_offset;
1008 } else {
1009 if (_pending_jvms != NULL &&
1010 _pending_offset < pc_offset) {
1011 emit_non_safepoint();
1012 }
1013 _pending_jvms = NULL;
1014 if (pc_offset > C->debug_info()->last_pc_offset()) {
1015 // This is the only way _pending_jvms can become non-NULL:
1016 _pending_jvms = nn->jvms();
1017 _pending_offset = pc_offset;
1018 }
1019 }
1020 }
1022 // Stay out of the way of real safepoints:
1023 void observe_safepoint(JVMState* jvms, int pc_offset) {
1024 if (_pending_jvms != NULL &&
1025 !_pending_jvms->same_calls_as(jvms) &&
1026 _pending_offset < pc_offset) {
1027 emit_non_safepoint();
1028 }
1029 _pending_jvms = NULL;
1030 }
1032 void flush_at_end() {
1033 if (_pending_jvms != NULL) {
1034 emit_non_safepoint();
1035 }
1036 _pending_jvms = NULL;
1037 }
1038 };
1040 void NonSafepointEmitter::emit_non_safepoint() {
1041 JVMState* youngest_jvms = _pending_jvms;
1042 int pc_offset = _pending_offset;
1044 // Clear it now:
1045 _pending_jvms = NULL;
1047 DebugInformationRecorder* debug_info = C->debug_info();
1048 assert(debug_info->recording_non_safepoints(), "sanity");
1050 debug_info->add_non_safepoint(pc_offset);
1051 int max_depth = youngest_jvms->depth();
1053 // Visit scopes from oldest to youngest.
1054 for (int depth = 1; depth <= max_depth; depth++) {
1055 JVMState* jvms = youngest_jvms->of_depth(depth);
1056 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1057 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1058 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1059 }
1061 // Mark the end of the scope set.
1062 debug_info->end_non_safepoint(pc_offset);
1063 }
1065 //------------------------------init_buffer------------------------------------
1066 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1068 // Set the initially allocated size
1069 int code_req = initial_code_capacity;
1070 int locs_req = initial_locs_capacity;
1071 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1072 int const_req = initial_const_capacity;
1074 int pad_req = NativeCall::instruction_size;
1075 // The extra spacing after the code is necessary on some platforms.
1076 // Sometimes we need to patch in a jump after the last instruction,
1077 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1079 // Compute the byte offset where we can store the deopt pc.
1080 if (fixed_slots() != 0) {
1081 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1082 }
1084 // Compute prolog code size
1085 _method_size = 0;
1086 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1087 #if defined(IA64) && !defined(AIX)
1088 if (save_argument_registers()) {
1089 // 4815101: this is a stub with implicit and unknown precision fp args.
1090 // The usual spill mechanism can only generate stfd's in this case, which
1091 // doesn't work if the fp reg to spill contains a single-precision denorm.
1092 // Instead, we hack around the normal spill mechanism using stfspill's and
1093 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1094 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1095 //
1096 // If we ever implement 16-byte 'registers' == stack slots, we can
1097 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1098 // instead of stfd/stfs/ldfd/ldfs.
1099 _frame_slots += 8*(16/BytesPerInt);
1100 }
1101 #endif
1102 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1104 if (has_mach_constant_base_node()) {
1105 uint add_size = 0;
1106 // Fill the constant table.
1107 // Note: This must happen before shorten_branches.
1108 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1109 Block* b = _cfg->get_block(i);
1111 for (uint j = 0; j < b->number_of_nodes(); j++) {
1112 Node* n = b->get_node(j);
1114 // If the node is a MachConstantNode evaluate the constant
1115 // value section.
1116 if (n->is_MachConstant()) {
1117 MachConstantNode* machcon = n->as_MachConstant();
1118 machcon->eval_constant(C);
1119 } else if (n->is_Mach()) {
1120 // On Power there are more nodes that issue constants.
1121 add_size += (n->as_Mach()->ins_num_consts() * 8);
1122 }
1123 }
1124 }
1126 // Calculate the offsets of the constants and the size of the
1127 // constant table (including the padding to the next section).
1128 constant_table().calculate_offsets_and_size();
1129 const_req = constant_table().size() + add_size;
1130 }
1132 // Initialize the space for the BufferBlob used to find and verify
1133 // instruction size in MachNode::emit_size()
1134 init_scratch_buffer_blob(const_req);
1135 if (failing()) return NULL; // Out of memory
1137 // Pre-compute the length of blocks and replace
1138 // long branches with short if machine supports it.
1139 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1141 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1142 // class HandlerImpl is platform-specific and defined in the *.ad files.
1143 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1144 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1145 stub_req += MAX_stubs_size; // ensure per-stub margin
1146 code_req += MAX_inst_size; // ensure per-instruction margin
1148 if (StressCodeBuffers)
1149 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1151 int total_req =
1152 const_req +
1153 code_req +
1154 pad_req +
1155 stub_req +
1156 exception_handler_req +
1157 deopt_handler_req; // deopt handler
1159 if (has_method_handle_invokes())
1160 total_req += deopt_handler_req; // deopt MH handler
1162 CodeBuffer* cb = code_buffer();
1163 cb->initialize(total_req, locs_req);
1165 // Have we run out of code space?
1166 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1167 C->record_failure("CodeCache is full");
1168 return NULL;
1169 }
1170 // Configure the code buffer.
1171 cb->initialize_consts_size(const_req);
1172 cb->initialize_stubs_size(stub_req);
1173 cb->initialize_oop_recorder(env()->oop_recorder());
1175 // fill in the nop array for bundling computations
1176 MachNode *_nop_list[Bundle::_nop_count];
1177 Bundle::initialize_nops(_nop_list, this);
1179 return cb;
1180 }
1182 //------------------------------fill_buffer------------------------------------
1183 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1184 // blk_starts[] contains offsets calculated during short branches processing,
1185 // offsets should not be increased during following steps.
1187 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1188 // of a loop. It is used to determine the padding for loop alignment.
1189 compute_loop_first_inst_sizes();
1191 // Create oopmap set.
1192 _oop_map_set = new OopMapSet();
1194 // !!!!! This preserves old handling of oopmaps for now
1195 debug_info()->set_oopmaps(_oop_map_set);
1197 uint nblocks = _cfg->number_of_blocks();
1198 // Count and start of implicit null check instructions
1199 uint inct_cnt = 0;
1200 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1202 // Count and start of calls
1203 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1205 uint return_offset = 0;
1206 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1208 int previous_offset = 0;
1209 int current_offset = 0;
1210 int last_call_offset = -1;
1211 int last_avoid_back_to_back_offset = -1;
1212 #ifdef ASSERT
1213 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1214 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1215 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1216 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1217 #endif
1219 // Create an array of unused labels, one for each basic block, if printing is enabled
1220 #ifndef PRODUCT
1221 int *node_offsets = NULL;
1222 uint node_offset_limit = unique();
1224 if (print_assembly())
1225 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1226 #endif
1228 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1230 // Emit the constant table.
1231 if (has_mach_constant_base_node()) {
1232 constant_table().emit(*cb);
1233 }
1235 // Create an array of labels, one for each basic block
1236 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1237 for (uint i=0; i <= nblocks; i++) {
1238 blk_labels[i].init();
1239 }
1241 // ------------------
1242 // Now fill in the code buffer
1243 Node *delay_slot = NULL;
1245 for (uint i = 0; i < nblocks; i++) {
1246 Block* block = _cfg->get_block(i);
1247 Node* head = block->head();
1249 // If this block needs to start aligned (i.e, can be reached other
1250 // than by falling-thru from the previous block), then force the
1251 // start of a new bundle.
1252 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1253 cb->flush_bundle(true);
1254 }
1256 #ifdef ASSERT
1257 if (!block->is_connector()) {
1258 stringStream st;
1259 block->dump_head(_cfg, &st);
1260 MacroAssembler(cb).block_comment(st.as_string());
1261 }
1262 jmp_target[i] = 0;
1263 jmp_offset[i] = 0;
1264 jmp_size[i] = 0;
1265 jmp_rule[i] = 0;
1266 #endif
1267 int blk_offset = current_offset;
1269 // Define the label at the beginning of the basic block
1270 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1272 uint last_inst = block->number_of_nodes();
1274 // Emit block normally, except for last instruction.
1275 // Emit means "dump code bits into code buffer".
1276 for (uint j = 0; j<last_inst; j++) {
1278 // Get the node
1279 Node* n = block->get_node(j);
1281 // See if delay slots are supported
1282 if (valid_bundle_info(n) &&
1283 node_bundling(n)->used_in_unconditional_delay()) {
1284 assert(delay_slot == NULL, "no use of delay slot node");
1285 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1287 delay_slot = n;
1288 continue;
1289 }
1291 // If this starts a new instruction group, then flush the current one
1292 // (but allow split bundles)
1293 if (Pipeline::requires_bundling() && starts_bundle(n))
1294 cb->flush_bundle(false);
1296 // The following logic is duplicated in the code ifdeffed for
1297 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1298 // should be factored out. Or maybe dispersed to the nodes?
1300 // Special handling for SafePoint/Call Nodes
1301 bool is_mcall = false;
1302 if (n->is_Mach()) {
1303 MachNode *mach = n->as_Mach();
1304 is_mcall = n->is_MachCall();
1305 bool is_sfn = n->is_MachSafePoint();
1307 // If this requires all previous instructions be flushed, then do so
1308 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1309 cb->flush_bundle(true);
1310 current_offset = cb->insts_size();
1311 }
1313 // A padding may be needed again since a previous instruction
1314 // could be moved to delay slot.
1316 // align the instruction if necessary
1317 int padding = mach->compute_padding(current_offset);
1318 // Make sure safepoint node for polling is distinct from a call's
1319 // return by adding a nop if needed.
1320 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1321 padding = nop_size;
1322 }
1323 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1324 current_offset == last_avoid_back_to_back_offset) {
1325 // Avoid back to back some instructions.
1326 padding = nop_size;
1327 }
1329 if(padding > 0) {
1330 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1331 int nops_cnt = padding / nop_size;
1332 MachNode *nop = new (this) MachNopNode(nops_cnt);
1333 block->insert_node(nop, j++);
1334 last_inst++;
1335 _cfg->map_node_to_block(nop, block);
1336 nop->emit(*cb, _regalloc);
1337 cb->flush_bundle(true);
1338 current_offset = cb->insts_size();
1339 }
1341 // Remember the start of the last call in a basic block
1342 if (is_mcall) {
1343 MachCallNode *mcall = mach->as_MachCall();
1345 // This destination address is NOT PC-relative
1346 mcall->method_set((intptr_t)mcall->entry_point());
1348 // Save the return address
1349 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1351 if (mcall->is_MachCallLeaf()) {
1352 is_mcall = false;
1353 is_sfn = false;
1354 }
1355 }
1357 // sfn will be valid whenever mcall is valid now because of inheritance
1358 if (is_sfn || is_mcall) {
1360 // Handle special safepoint nodes for synchronization
1361 if (!is_mcall) {
1362 MachSafePointNode *sfn = mach->as_MachSafePoint();
1363 // !!!!! Stubs only need an oopmap right now, so bail out
1364 if (sfn->jvms()->method() == NULL) {
1365 // Write the oopmap directly to the code blob??!!
1366 # ifdef ENABLE_ZAP_DEAD_LOCALS
1367 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1368 # endif
1369 continue;
1370 }
1371 } // End synchronization
1373 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1374 current_offset);
1375 Process_OopMap_Node(mach, current_offset);
1376 } // End if safepoint
1378 // If this is a null check, then add the start of the previous instruction to the list
1379 else if( mach->is_MachNullCheck() ) {
1380 inct_starts[inct_cnt++] = previous_offset;
1381 }
1383 // If this is a branch, then fill in the label with the target BB's label
1384 else if (mach->is_MachBranch()) {
1385 // This requires the TRUE branch target be in succs[0]
1386 uint block_num = block->non_connector_successor(0)->_pre_order;
1388 // Try to replace long branch if delay slot is not used,
1389 // it is mostly for back branches since forward branch's
1390 // distance is not updated yet.
1391 bool delay_slot_is_used = valid_bundle_info(n) &&
1392 node_bundling(n)->use_unconditional_delay();
1393 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1394 assert(delay_slot == NULL, "not expecting delay slot node");
1395 int br_size = n->size(_regalloc);
1396 int offset = blk_starts[block_num] - current_offset;
1397 if (block_num >= i) {
1398 // Current and following block's offset are not
1399 // finalized yet, adjust distance by the difference
1400 // between calculated and final offsets of current block.
1401 offset -= (blk_starts[i] - blk_offset);
1402 }
1403 // In the following code a nop could be inserted before
1404 // the branch which will increase the backward distance.
1405 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1406 if (needs_padding && offset <= 0)
1407 offset -= nop_size;
1409 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1410 // We've got a winner. Replace this branch.
1411 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1413 // Update the jmp_size.
1414 int new_size = replacement->size(_regalloc);
1415 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1416 // Insert padding between avoid_back_to_back branches.
1417 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1418 MachNode *nop = new (this) MachNopNode();
1419 block->insert_node(nop, j++);
1420 _cfg->map_node_to_block(nop, block);
1421 last_inst++;
1422 nop->emit(*cb, _regalloc);
1423 cb->flush_bundle(true);
1424 current_offset = cb->insts_size();
1425 }
1426 #ifdef ASSERT
1427 jmp_target[i] = block_num;
1428 jmp_offset[i] = current_offset - blk_offset;
1429 jmp_size[i] = new_size;
1430 jmp_rule[i] = mach->rule();
1431 #endif
1432 block->map_node(replacement, j);
1433 mach->subsume_by(replacement, C);
1434 n = replacement;
1435 mach = replacement;
1436 }
1437 }
1438 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1439 } else if (mach->ideal_Opcode() == Op_Jump) {
1440 for (uint h = 0; h < block->_num_succs; h++) {
1441 Block* succs_block = block->_succs[h];
1442 for (uint j = 1; j < succs_block->num_preds(); j++) {
1443 Node* jpn = succs_block->pred(j);
1444 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1445 uint block_num = succs_block->non_connector()->_pre_order;
1446 Label *blkLabel = &blk_labels[block_num];
1447 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1448 }
1449 }
1450 }
1451 }
1452 #ifdef ASSERT
1453 // Check that oop-store precedes the card-mark
1454 else if (mach->ideal_Opcode() == Op_StoreCM) {
1455 uint storeCM_idx = j;
1456 int count = 0;
1457 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1458 Node *oop_store = mach->in(prec); // Precedence edge
1459 if (oop_store == NULL) continue;
1460 count++;
1461 uint i4;
1462 for (i4 = 0; i4 < last_inst; ++i4) {
1463 if (block->get_node(i4) == oop_store) {
1464 break;
1465 }
1466 }
1467 // Note: This test can provide a false failure if other precedence
1468 // edges have been added to the storeCMNode.
1469 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1470 }
1471 assert(count > 0, "storeCM expects at least one precedence edge");
1472 }
1473 #endif
1474 else if (!n->is_Proj()) {
1475 // Remember the beginning of the previous instruction, in case
1476 // it's followed by a flag-kill and a null-check. Happens on
1477 // Intel all the time, with add-to-memory kind of opcodes.
1478 previous_offset = current_offset;
1479 }
1481 // Not an else-if!
1482 // If this is a trap based cmp then add its offset to the list.
1483 if (mach->is_TrapBasedCheckNode()) {
1484 inct_starts[inct_cnt++] = current_offset;
1485 }
1486 }
1488 // Verify that there is sufficient space remaining
1489 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1490 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1491 C->record_failure("CodeCache is full");
1492 return;
1493 }
1495 // Save the offset for the listing
1496 #ifndef PRODUCT
1497 if (node_offsets && n->_idx < node_offset_limit)
1498 node_offsets[n->_idx] = cb->insts_size();
1499 #endif
1501 // "Normal" instruction case
1502 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1503 n->emit(*cb, _regalloc);
1504 current_offset = cb->insts_size();
1506 #ifdef ASSERT
1507 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1508 n->dump();
1509 assert(false, "wrong size of mach node");
1510 }
1511 #endif
1512 non_safepoints.observe_instruction(n, current_offset);
1514 // mcall is last "call" that can be a safepoint
1515 // record it so we can see if a poll will directly follow it
1516 // in which case we'll need a pad to make the PcDesc sites unique
1517 // see 5010568. This can be slightly inaccurate but conservative
1518 // in the case that return address is not actually at current_offset.
1519 // This is a small price to pay.
1521 if (is_mcall) {
1522 last_call_offset = current_offset;
1523 }
1525 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1526 // Avoid back to back some instructions.
1527 last_avoid_back_to_back_offset = current_offset;
1528 }
1530 // See if this instruction has a delay slot
1531 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1532 assert(delay_slot != NULL, "expecting delay slot node");
1534 // Back up 1 instruction
1535 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1537 // Save the offset for the listing
1538 #ifndef PRODUCT
1539 if (node_offsets && delay_slot->_idx < node_offset_limit)
1540 node_offsets[delay_slot->_idx] = cb->insts_size();
1541 #endif
1543 // Support a SafePoint in the delay slot
1544 if (delay_slot->is_MachSafePoint()) {
1545 MachNode *mach = delay_slot->as_Mach();
1546 // !!!!! Stubs only need an oopmap right now, so bail out
1547 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1548 // Write the oopmap directly to the code blob??!!
1549 # ifdef ENABLE_ZAP_DEAD_LOCALS
1550 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1551 # endif
1552 delay_slot = NULL;
1553 continue;
1554 }
1556 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1557 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1558 adjusted_offset);
1559 // Generate an OopMap entry
1560 Process_OopMap_Node(mach, adjusted_offset);
1561 }
1563 // Insert the delay slot instruction
1564 delay_slot->emit(*cb, _regalloc);
1566 // Don't reuse it
1567 delay_slot = NULL;
1568 }
1570 } // End for all instructions in block
1572 // If the next block is the top of a loop, pad this block out to align
1573 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1574 if (i < nblocks-1) {
1575 Block *nb = _cfg->get_block(i + 1);
1576 int padding = nb->alignment_padding(current_offset);
1577 if( padding > 0 ) {
1578 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1579 block->insert_node(nop, block->number_of_nodes());
1580 _cfg->map_node_to_block(nop, block);
1581 nop->emit(*cb, _regalloc);
1582 current_offset = cb->insts_size();
1583 }
1584 }
1585 // Verify that the distance for generated before forward
1586 // short branches is still valid.
1587 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1589 // Save new block start offset
1590 blk_starts[i] = blk_offset;
1591 } // End of for all blocks
1592 blk_starts[nblocks] = current_offset;
1594 non_safepoints.flush_at_end();
1596 // Offset too large?
1597 if (failing()) return;
1599 // Define a pseudo-label at the end of the code
1600 MacroAssembler(cb).bind( blk_labels[nblocks] );
1602 // Compute the size of the first block
1603 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1605 assert(cb->insts_size() < 500000, "method is unreasonably large");
1607 #ifdef ASSERT
1608 for (uint i = 0; i < nblocks; i++) { // For all blocks
1609 if (jmp_target[i] != 0) {
1610 int br_size = jmp_size[i];
1611 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1612 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1613 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]);
1614 assert(false, "Displacement too large for short jmp");
1615 }
1616 }
1617 }
1618 #endif
1620 #ifndef PRODUCT
1621 // Information on the size of the method, without the extraneous code
1622 Scheduling::increment_method_size(cb->insts_size());
1623 #endif
1625 // ------------------
1626 // Fill in exception table entries.
1627 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1629 // Only java methods have exception handlers and deopt handlers
1630 // class HandlerImpl is platform-specific and defined in the *.ad files.
1631 if (_method) {
1632 // Emit the exception handler code.
1633 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1634 // Emit the deopt handler code.
1635 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1637 // Emit the MethodHandle deopt handler code (if required).
1638 if (has_method_handle_invokes()) {
1639 // We can use the same code as for the normal deopt handler, we
1640 // just need a different entry point address.
1641 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1642 }
1643 }
1645 // One last check for failed CodeBuffer::expand:
1646 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1647 C->record_failure("CodeCache is full");
1648 return;
1649 }
1651 #ifndef PRODUCT
1652 // Dump the assembly code, including basic-block numbers
1653 if (print_assembly()) {
1654 ttyLocker ttyl; // keep the following output all in one block
1655 if (!VMThread::should_terminate()) { // test this under the tty lock
1656 // This output goes directly to the tty, not the compiler log.
1657 // To enable tools to match it up with the compilation activity,
1658 // be sure to tag this tty output with the compile ID.
1659 if (xtty != NULL) {
1660 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1661 is_osr_compilation() ? " compile_kind='osr'" :
1662 "");
1663 }
1664 if (method() != NULL) {
1665 method()->print_metadata();
1666 }
1667 dump_asm(node_offsets, node_offset_limit);
1668 if (xtty != NULL) {
1669 xtty->tail("opto_assembly");
1670 }
1671 }
1672 }
1673 #endif
1675 }
1677 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1678 _inc_table.set_size(cnt);
1680 uint inct_cnt = 0;
1681 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1682 Block* block = _cfg->get_block(i);
1683 Node *n = NULL;
1684 int j;
1686 // Find the branch; ignore trailing NOPs.
1687 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1688 n = block->get_node(j);
1689 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1690 break;
1691 }
1692 }
1694 // If we didn't find anything, continue
1695 if (j < 0) {
1696 continue;
1697 }
1699 // Compute ExceptionHandlerTable subtable entry and add it
1700 // (skip empty blocks)
1701 if (n->is_Catch()) {
1703 // Get the offset of the return from the call
1704 uint call_return = call_returns[block->_pre_order];
1705 #ifdef ASSERT
1706 assert( call_return > 0, "no call seen for this basic block" );
1707 while (block->get_node(--j)->is_MachProj()) ;
1708 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1709 #endif
1710 // last instruction is a CatchNode, find it's CatchProjNodes
1711 int nof_succs = block->_num_succs;
1712 // allocate space
1713 GrowableArray<intptr_t> handler_bcis(nof_succs);
1714 GrowableArray<intptr_t> handler_pcos(nof_succs);
1715 // iterate through all successors
1716 for (int j = 0; j < nof_succs; j++) {
1717 Block* s = block->_succs[j];
1718 bool found_p = false;
1719 for (uint k = 1; k < s->num_preds(); k++) {
1720 Node* pk = s->pred(k);
1721 if (pk->is_CatchProj() && pk->in(0) == n) {
1722 const CatchProjNode* p = pk->as_CatchProj();
1723 found_p = true;
1724 // add the corresponding handler bci & pco information
1725 if (p->_con != CatchProjNode::fall_through_index) {
1726 // p leads to an exception handler (and is not fall through)
1727 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1728 // no duplicates, please
1729 if (!handler_bcis.contains(p->handler_bci())) {
1730 uint block_num = s->non_connector()->_pre_order;
1731 handler_bcis.append(p->handler_bci());
1732 handler_pcos.append(blk_labels[block_num].loc_pos());
1733 }
1734 }
1735 }
1736 }
1737 assert(found_p, "no matching predecessor found");
1738 // Note: Due to empty block removal, one block may have
1739 // several CatchProj inputs, from the same Catch.
1740 }
1742 // Set the offset of the return from the call
1743 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1744 continue;
1745 }
1747 // Handle implicit null exception table updates
1748 if (n->is_MachNullCheck()) {
1749 uint block_num = block->non_connector_successor(0)->_pre_order;
1750 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1751 continue;
1752 }
1753 // Handle implicit exception table updates: trap instructions.
1754 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1755 uint block_num = block->non_connector_successor(0)->_pre_order;
1756 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1757 continue;
1758 }
1759 } // End of for all blocks fill in exception table entries
1760 }
1762 // Static Variables
1763 #ifndef PRODUCT
1764 uint Scheduling::_total_nop_size = 0;
1765 uint Scheduling::_total_method_size = 0;
1766 uint Scheduling::_total_branches = 0;
1767 uint Scheduling::_total_unconditional_delays = 0;
1768 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1769 #endif
1771 // Initializer for class Scheduling
1773 Scheduling::Scheduling(Arena *arena, Compile &compile)
1774 : _arena(arena),
1775 _cfg(compile.cfg()),
1776 _regalloc(compile.regalloc()),
1777 _reg_node(arena),
1778 _bundle_instr_count(0),
1779 _bundle_cycle_number(0),
1780 _scheduled(arena),
1781 _available(arena),
1782 _next_node(NULL),
1783 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1784 _pinch_free_list(arena)
1785 #ifndef PRODUCT
1786 , _branches(0)
1787 , _unconditional_delays(0)
1788 #endif
1789 {
1790 // Create a MachNopNode
1791 _nop = new (&compile) MachNopNode();
1793 // Now that the nops are in the array, save the count
1794 // (but allow entries for the nops)
1795 _node_bundling_limit = compile.unique();
1796 uint node_max = _regalloc->node_regs_max_index();
1798 compile.set_node_bundling_limit(_node_bundling_limit);
1800 // This one is persistent within the Compile class
1801 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1803 // Allocate space for fixed-size arrays
1804 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1805 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1806 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1808 // Clear the arrays
1809 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1810 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1811 memset(_uses, 0, node_max * sizeof(short));
1812 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1814 // Clear the bundling information
1815 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1817 // Get the last node
1818 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1820 _next_node = block->get_node(block->number_of_nodes() - 1);
1821 }
1823 #ifndef PRODUCT
1824 // Scheduling destructor
1825 Scheduling::~Scheduling() {
1826 _total_branches += _branches;
1827 _total_unconditional_delays += _unconditional_delays;
1828 }
1829 #endif
1831 // Step ahead "i" cycles
1832 void Scheduling::step(uint i) {
1834 Bundle *bundle = node_bundling(_next_node);
1835 bundle->set_starts_bundle();
1837 // Update the bundle record, but leave the flags information alone
1838 if (_bundle_instr_count > 0) {
1839 bundle->set_instr_count(_bundle_instr_count);
1840 bundle->set_resources_used(_bundle_use.resourcesUsed());
1841 }
1843 // Update the state information
1844 _bundle_instr_count = 0;
1845 _bundle_cycle_number += i;
1846 _bundle_use.step(i);
1847 }
1849 void Scheduling::step_and_clear() {
1850 Bundle *bundle = node_bundling(_next_node);
1851 bundle->set_starts_bundle();
1853 // Update the bundle record
1854 if (_bundle_instr_count > 0) {
1855 bundle->set_instr_count(_bundle_instr_count);
1856 bundle->set_resources_used(_bundle_use.resourcesUsed());
1858 _bundle_cycle_number += 1;
1859 }
1861 // Clear the bundling information
1862 _bundle_instr_count = 0;
1863 _bundle_use.reset();
1865 memcpy(_bundle_use_elements,
1866 Pipeline_Use::elaborated_elements,
1867 sizeof(Pipeline_Use::elaborated_elements));
1868 }
1870 // Perform instruction scheduling and bundling over the sequence of
1871 // instructions in backwards order.
1872 void Compile::ScheduleAndBundle() {
1874 // Don't optimize this if it isn't a method
1875 if (!_method)
1876 return;
1878 // Don't optimize this if scheduling is disabled
1879 if (!do_scheduling())
1880 return;
1882 // Scheduling code works only with pairs (8 bytes) maximum.
1883 if (max_vector_size() > 8)
1884 return;
1886 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1888 // Create a data structure for all the scheduling information
1889 Scheduling scheduling(Thread::current()->resource_area(), *this);
1891 // Walk backwards over each basic block, computing the needed alignment
1892 // Walk over all the basic blocks
1893 scheduling.DoScheduling();
1894 }
1896 // Compute the latency of all the instructions. This is fairly simple,
1897 // because we already have a legal ordering. Walk over the instructions
1898 // from first to last, and compute the latency of the instruction based
1899 // on the latency of the preceding instruction(s).
1900 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1901 #ifndef PRODUCT
1902 if (_cfg->C->trace_opto_output())
1903 tty->print("# -> ComputeLocalLatenciesForward\n");
1904 #endif
1906 // Walk over all the schedulable instructions
1907 for( uint j=_bb_start; j < _bb_end; j++ ) {
1909 // This is a kludge, forcing all latency calculations to start at 1.
1910 // Used to allow latency 0 to force an instruction to the beginning
1911 // of the bb
1912 uint latency = 1;
1913 Node *use = bb->get_node(j);
1914 uint nlen = use->len();
1916 // Walk over all the inputs
1917 for ( uint k=0; k < nlen; k++ ) {
1918 Node *def = use->in(k);
1919 if (!def)
1920 continue;
1922 uint l = _node_latency[def->_idx] + use->latency(k);
1923 if (latency < l)
1924 latency = l;
1925 }
1927 _node_latency[use->_idx] = latency;
1929 #ifndef PRODUCT
1930 if (_cfg->C->trace_opto_output()) {
1931 tty->print("# latency %4d: ", latency);
1932 use->dump();
1933 }
1934 #endif
1935 }
1937 #ifndef PRODUCT
1938 if (_cfg->C->trace_opto_output())
1939 tty->print("# <- ComputeLocalLatenciesForward\n");
1940 #endif
1942 } // end ComputeLocalLatenciesForward
1944 // See if this node fits into the present instruction bundle
1945 bool Scheduling::NodeFitsInBundle(Node *n) {
1946 uint n_idx = n->_idx;
1948 // If this is the unconditional delay instruction, then it fits
1949 if (n == _unconditional_delay_slot) {
1950 #ifndef PRODUCT
1951 if (_cfg->C->trace_opto_output())
1952 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1953 #endif
1954 return (true);
1955 }
1957 // If the node cannot be scheduled this cycle, skip it
1958 if (_current_latency[n_idx] > _bundle_cycle_number) {
1959 #ifndef PRODUCT
1960 if (_cfg->C->trace_opto_output())
1961 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1962 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1963 #endif
1964 return (false);
1965 }
1967 const Pipeline *node_pipeline = n->pipeline();
1969 uint instruction_count = node_pipeline->instructionCount();
1970 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1971 instruction_count = 0;
1972 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1973 instruction_count++;
1975 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1976 #ifndef PRODUCT
1977 if (_cfg->C->trace_opto_output())
1978 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1979 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1980 #endif
1981 return (false);
1982 }
1984 // Don't allow non-machine nodes to be handled this way
1985 if (!n->is_Mach() && instruction_count == 0)
1986 return (false);
1988 // See if there is any overlap
1989 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1991 if (delay > 0) {
1992 #ifndef PRODUCT
1993 if (_cfg->C->trace_opto_output())
1994 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1995 #endif
1996 return false;
1997 }
1999 #ifndef PRODUCT
2000 if (_cfg->C->trace_opto_output())
2001 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
2002 #endif
2004 return true;
2005 }
2007 Node * Scheduling::ChooseNodeToBundle() {
2008 uint siz = _available.size();
2010 if (siz == 0) {
2012 #ifndef PRODUCT
2013 if (_cfg->C->trace_opto_output())
2014 tty->print("# ChooseNodeToBundle: NULL\n");
2015 #endif
2016 return (NULL);
2017 }
2019 // Fast path, if only 1 instruction in the bundle
2020 if (siz == 1) {
2021 #ifndef PRODUCT
2022 if (_cfg->C->trace_opto_output()) {
2023 tty->print("# ChooseNodeToBundle (only 1): ");
2024 _available[0]->dump();
2025 }
2026 #endif
2027 return (_available[0]);
2028 }
2030 // Don't bother, if the bundle is already full
2031 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2032 for ( uint i = 0; i < siz; i++ ) {
2033 Node *n = _available[i];
2035 // Skip projections, we'll handle them another way
2036 if (n->is_Proj())
2037 continue;
2039 // This presupposed that instructions are inserted into the
2040 // available list in a legality order; i.e. instructions that
2041 // must be inserted first are at the head of the list
2042 if (NodeFitsInBundle(n)) {
2043 #ifndef PRODUCT
2044 if (_cfg->C->trace_opto_output()) {
2045 tty->print("# ChooseNodeToBundle: ");
2046 n->dump();
2047 }
2048 #endif
2049 return (n);
2050 }
2051 }
2052 }
2054 // Nothing fits in this bundle, choose the highest priority
2055 #ifndef PRODUCT
2056 if (_cfg->C->trace_opto_output()) {
2057 tty->print("# ChooseNodeToBundle: ");
2058 _available[0]->dump();
2059 }
2060 #endif
2062 return _available[0];
2063 }
2065 void Scheduling::AddNodeToAvailableList(Node *n) {
2066 assert( !n->is_Proj(), "projections never directly made available" );
2067 #ifndef PRODUCT
2068 if (_cfg->C->trace_opto_output()) {
2069 tty->print("# AddNodeToAvailableList: ");
2070 n->dump();
2071 }
2072 #endif
2074 int latency = _current_latency[n->_idx];
2076 // Insert in latency order (insertion sort)
2077 uint i;
2078 for ( i=0; i < _available.size(); i++ )
2079 if (_current_latency[_available[i]->_idx] > latency)
2080 break;
2082 // Special Check for compares following branches
2083 if( n->is_Mach() && _scheduled.size() > 0 ) {
2084 int op = n->as_Mach()->ideal_Opcode();
2085 Node *last = _scheduled[0];
2086 if( last->is_MachIf() && last->in(1) == n &&
2087 ( op == Op_CmpI ||
2088 op == Op_CmpU ||
2089 op == Op_CmpP ||
2090 op == Op_CmpF ||
2091 op == Op_CmpD ||
2092 op == Op_CmpL ) ) {
2094 // Recalculate position, moving to front of same latency
2095 for ( i=0 ; i < _available.size(); i++ )
2096 if (_current_latency[_available[i]->_idx] >= latency)
2097 break;
2098 }
2099 }
2101 // Insert the node in the available list
2102 _available.insert(i, n);
2104 #ifndef PRODUCT
2105 if (_cfg->C->trace_opto_output())
2106 dump_available();
2107 #endif
2108 }
2110 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2111 for ( uint i=0; i < n->len(); i++ ) {
2112 Node *def = n->in(i);
2113 if (!def) continue;
2114 if( def->is_Proj() ) // If this is a machine projection, then
2115 def = def->in(0); // propagate usage thru to the base instruction
2117 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2118 continue;
2119 }
2121 // Compute the latency
2122 uint l = _bundle_cycle_number + n->latency(i);
2123 if (_current_latency[def->_idx] < l)
2124 _current_latency[def->_idx] = l;
2126 // If this does not have uses then schedule it
2127 if ((--_uses[def->_idx]) == 0)
2128 AddNodeToAvailableList(def);
2129 }
2130 }
2132 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2133 #ifndef PRODUCT
2134 if (_cfg->C->trace_opto_output()) {
2135 tty->print("# AddNodeToBundle: ");
2136 n->dump();
2137 }
2138 #endif
2140 // Remove this from the available list
2141 uint i;
2142 for (i = 0; i < _available.size(); i++)
2143 if (_available[i] == n)
2144 break;
2145 assert(i < _available.size(), "entry in _available list not found");
2146 _available.remove(i);
2148 // See if this fits in the current bundle
2149 const Pipeline *node_pipeline = n->pipeline();
2150 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2152 // Check for instructions to be placed in the delay slot. We
2153 // do this before we actually schedule the current instruction,
2154 // because the delay slot follows the current instruction.
2155 if (Pipeline::_branch_has_delay_slot &&
2156 node_pipeline->hasBranchDelay() &&
2157 !_unconditional_delay_slot) {
2159 uint siz = _available.size();
2161 // Conditional branches can support an instruction that
2162 // is unconditionally executed and not dependent by the
2163 // branch, OR a conditionally executed instruction if
2164 // the branch is taken. In practice, this means that
2165 // the first instruction at the branch target is
2166 // copied to the delay slot, and the branch goes to
2167 // the instruction after that at the branch target
2168 if ( n->is_MachBranch() ) {
2170 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2171 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2173 #ifndef PRODUCT
2174 _branches++;
2175 #endif
2177 // At least 1 instruction is on the available list
2178 // that is not dependent on the branch
2179 for (uint i = 0; i < siz; i++) {
2180 Node *d = _available[i];
2181 const Pipeline *avail_pipeline = d->pipeline();
2183 // Don't allow safepoints in the branch shadow, that will
2184 // cause a number of difficulties
2185 if ( avail_pipeline->instructionCount() == 1 &&
2186 !avail_pipeline->hasMultipleBundles() &&
2187 !avail_pipeline->hasBranchDelay() &&
2188 Pipeline::instr_has_unit_size() &&
2189 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2190 NodeFitsInBundle(d) &&
2191 !node_bundling(d)->used_in_delay()) {
2193 if (d->is_Mach() && !d->is_MachSafePoint()) {
2194 // A node that fits in the delay slot was found, so we need to
2195 // set the appropriate bits in the bundle pipeline information so
2196 // that it correctly indicates resource usage. Later, when we
2197 // attempt to add this instruction to the bundle, we will skip
2198 // setting the resource usage.
2199 _unconditional_delay_slot = d;
2200 node_bundling(n)->set_use_unconditional_delay();
2201 node_bundling(d)->set_used_in_unconditional_delay();
2202 _bundle_use.add_usage(avail_pipeline->resourceUse());
2203 _current_latency[d->_idx] = _bundle_cycle_number;
2204 _next_node = d;
2205 ++_bundle_instr_count;
2206 #ifndef PRODUCT
2207 _unconditional_delays++;
2208 #endif
2209 break;
2210 }
2211 }
2212 }
2213 }
2215 // No delay slot, add a nop to the usage
2216 if (!_unconditional_delay_slot) {
2217 // See if adding an instruction in the delay slot will overflow
2218 // the bundle.
2219 if (!NodeFitsInBundle(_nop)) {
2220 #ifndef PRODUCT
2221 if (_cfg->C->trace_opto_output())
2222 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2223 #endif
2224 step(1);
2225 }
2227 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2228 _next_node = _nop;
2229 ++_bundle_instr_count;
2230 }
2232 // See if the instruction in the delay slot requires a
2233 // step of the bundles
2234 if (!NodeFitsInBundle(n)) {
2235 #ifndef PRODUCT
2236 if (_cfg->C->trace_opto_output())
2237 tty->print("# *** STEP(branch won't fit) ***\n");
2238 #endif
2239 // Update the state information
2240 _bundle_instr_count = 0;
2241 _bundle_cycle_number += 1;
2242 _bundle_use.step(1);
2243 }
2244 }
2246 // Get the number of instructions
2247 uint instruction_count = node_pipeline->instructionCount();
2248 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2249 instruction_count = 0;
2251 // Compute the latency information
2252 uint delay = 0;
2254 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2255 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2256 if (relative_latency < 0)
2257 relative_latency = 0;
2259 delay = _bundle_use.full_latency(relative_latency, node_usage);
2261 // Does not fit in this bundle, start a new one
2262 if (delay > 0) {
2263 step(delay);
2265 #ifndef PRODUCT
2266 if (_cfg->C->trace_opto_output())
2267 tty->print("# *** STEP(%d) ***\n", delay);
2268 #endif
2269 }
2270 }
2272 // If this was placed in the delay slot, ignore it
2273 if (n != _unconditional_delay_slot) {
2275 if (delay == 0) {
2276 if (node_pipeline->hasMultipleBundles()) {
2277 #ifndef PRODUCT
2278 if (_cfg->C->trace_opto_output())
2279 tty->print("# *** STEP(multiple instructions) ***\n");
2280 #endif
2281 step(1);
2282 }
2284 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2285 #ifndef PRODUCT
2286 if (_cfg->C->trace_opto_output())
2287 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2288 instruction_count + _bundle_instr_count,
2289 Pipeline::_max_instrs_per_cycle);
2290 #endif
2291 step(1);
2292 }
2293 }
2295 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2296 _bundle_instr_count++;
2298 // Set the node's latency
2299 _current_latency[n->_idx] = _bundle_cycle_number;
2301 // Now merge the functional unit information
2302 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2303 _bundle_use.add_usage(node_usage);
2305 // Increment the number of instructions in this bundle
2306 _bundle_instr_count += instruction_count;
2308 // Remember this node for later
2309 if (n->is_Mach())
2310 _next_node = n;
2311 }
2313 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2314 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2315 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2316 // into the block. All other scheduled nodes get put in the schedule here.
2317 int op = n->Opcode();
2318 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2319 (op != Op_Node && // Not an unused antidepedence node and
2320 // not an unallocated boxlock
2321 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2323 // Push any trailing projections
2324 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2325 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2326 Node *foi = n->fast_out(i);
2327 if( foi->is_Proj() )
2328 _scheduled.push(foi);
2329 }
2330 }
2332 // Put the instruction in the schedule list
2333 _scheduled.push(n);
2334 }
2336 #ifndef PRODUCT
2337 if (_cfg->C->trace_opto_output())
2338 dump_available();
2339 #endif
2341 // Walk all the definitions, decrementing use counts, and
2342 // if a definition has a 0 use count, place it in the available list.
2343 DecrementUseCounts(n,bb);
2344 }
2346 // This method sets the use count within a basic block. We will ignore all
2347 // uses outside the current basic block. As we are doing a backwards walk,
2348 // any node we reach that has a use count of 0 may be scheduled. This also
2349 // avoids the problem of cyclic references from phi nodes, as long as phi
2350 // nodes are at the front of the basic block. This method also initializes
2351 // the available list to the set of instructions that have no uses within this
2352 // basic block.
2353 void Scheduling::ComputeUseCount(const Block *bb) {
2354 #ifndef PRODUCT
2355 if (_cfg->C->trace_opto_output())
2356 tty->print("# -> ComputeUseCount\n");
2357 #endif
2359 // Clear the list of available and scheduled instructions, just in case
2360 _available.clear();
2361 _scheduled.clear();
2363 // No delay slot specified
2364 _unconditional_delay_slot = NULL;
2366 #ifdef ASSERT
2367 for( uint i=0; i < bb->number_of_nodes(); i++ )
2368 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2369 #endif
2371 // Force the _uses count to never go to zero for unscheduable pieces
2372 // of the block
2373 for( uint k = 0; k < _bb_start; k++ )
2374 _uses[bb->get_node(k)->_idx] = 1;
2375 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2376 _uses[bb->get_node(l)->_idx] = 1;
2378 // Iterate backwards over the instructions in the block. Don't count the
2379 // branch projections at end or the block header instructions.
2380 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2381 Node *n = bb->get_node(j);
2382 if( n->is_Proj() ) continue; // Projections handled another way
2384 // Account for all uses
2385 for ( uint k = 0; k < n->len(); k++ ) {
2386 Node *inp = n->in(k);
2387 if (!inp) continue;
2388 assert(inp != n, "no cycles allowed" );
2389 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2390 if (inp->is_Proj()) { // Skip through Proj's
2391 inp = inp->in(0);
2392 }
2393 ++_uses[inp->_idx]; // Count 1 block-local use
2394 }
2395 }
2397 // If this instruction has a 0 use count, then it is available
2398 if (!_uses[n->_idx]) {
2399 _current_latency[n->_idx] = _bundle_cycle_number;
2400 AddNodeToAvailableList(n);
2401 }
2403 #ifndef PRODUCT
2404 if (_cfg->C->trace_opto_output()) {
2405 tty->print("# uses: %3d: ", _uses[n->_idx]);
2406 n->dump();
2407 }
2408 #endif
2409 }
2411 #ifndef PRODUCT
2412 if (_cfg->C->trace_opto_output())
2413 tty->print("# <- ComputeUseCount\n");
2414 #endif
2415 }
2417 // This routine performs scheduling on each basic block in reverse order,
2418 // using instruction latencies and taking into account function unit
2419 // availability.
2420 void Scheduling::DoScheduling() {
2421 #ifndef PRODUCT
2422 if (_cfg->C->trace_opto_output())
2423 tty->print("# -> DoScheduling\n");
2424 #endif
2426 Block *succ_bb = NULL;
2427 Block *bb;
2429 // Walk over all the basic blocks in reverse order
2430 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2431 bb = _cfg->get_block(i);
2433 #ifndef PRODUCT
2434 if (_cfg->C->trace_opto_output()) {
2435 tty->print("# Schedule BB#%03d (initial)\n", i);
2436 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2437 bb->get_node(j)->dump();
2438 }
2439 }
2440 #endif
2442 // On the head node, skip processing
2443 if (bb == _cfg->get_root_block()) {
2444 continue;
2445 }
2447 // Skip empty, connector blocks
2448 if (bb->is_connector())
2449 continue;
2451 // If the following block is not the sole successor of
2452 // this one, then reset the pipeline information
2453 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2454 #ifndef PRODUCT
2455 if (_cfg->C->trace_opto_output()) {
2456 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2457 _next_node->_idx, _bundle_instr_count);
2458 }
2459 #endif
2460 step_and_clear();
2461 }
2463 // Leave untouched the starting instruction, any Phis, a CreateEx node
2464 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2465 _bb_end = bb->number_of_nodes()-1;
2466 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2467 Node *n = bb->get_node(_bb_start);
2468 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2469 // Also, MachIdealNodes do not get scheduled
2470 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2471 MachNode *mach = n->as_Mach();
2472 int iop = mach->ideal_Opcode();
2473 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2474 if( iop == Op_Con ) continue; // Do not schedule Top
2475 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2476 mach->pipeline() == MachNode::pipeline_class() &&
2477 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2478 continue;
2479 break; // Funny loop structure to be sure...
2480 }
2481 // Compute last "interesting" instruction in block - last instruction we
2482 // might schedule. _bb_end points just after last schedulable inst. We
2483 // normally schedule conditional branches (despite them being forced last
2484 // in the block), because they have delay slots we can fill. Calls all
2485 // have their delay slots filled in the template expansions, so we don't
2486 // bother scheduling them.
2487 Node *last = bb->get_node(_bb_end);
2488 // Ignore trailing NOPs.
2489 while (_bb_end > 0 && last->is_Mach() &&
2490 last->as_Mach()->ideal_Opcode() == Op_Con) {
2491 last = bb->get_node(--_bb_end);
2492 }
2493 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2494 if( last->is_Catch() ||
2495 // Exclude unreachable path case when Halt node is in a separate block.
2496 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2497 // There must be a prior call. Skip it.
2498 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2499 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2500 }
2501 } else if( last->is_MachNullCheck() ) {
2502 // Backup so the last null-checked memory instruction is
2503 // outside the schedulable range. Skip over the nullcheck,
2504 // projection, and the memory nodes.
2505 Node *mem = last->in(1);
2506 do {
2507 _bb_end--;
2508 } while (mem != bb->get_node(_bb_end));
2509 } else {
2510 // Set _bb_end to point after last schedulable inst.
2511 _bb_end++;
2512 }
2514 assert( _bb_start <= _bb_end, "inverted block ends" );
2516 // Compute the register antidependencies for the basic block
2517 ComputeRegisterAntidependencies(bb);
2518 if (_cfg->C->failing()) return; // too many D-U pinch points
2520 // Compute intra-bb latencies for the nodes
2521 ComputeLocalLatenciesForward(bb);
2523 // Compute the usage within the block, and set the list of all nodes
2524 // in the block that have no uses within the block.
2525 ComputeUseCount(bb);
2527 // Schedule the remaining instructions in the block
2528 while ( _available.size() > 0 ) {
2529 Node *n = ChooseNodeToBundle();
2530 guarantee(n != NULL, "no nodes available");
2531 AddNodeToBundle(n,bb);
2532 }
2534 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2535 #ifdef ASSERT
2536 for( uint l = _bb_start; l < _bb_end; l++ ) {
2537 Node *n = bb->get_node(l);
2538 uint m;
2539 for( m = 0; m < _bb_end-_bb_start; m++ )
2540 if( _scheduled[m] == n )
2541 break;
2542 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2543 }
2544 #endif
2546 // Now copy the instructions (in reverse order) back to the block
2547 for ( uint k = _bb_start; k < _bb_end; k++ )
2548 bb->map_node(_scheduled[_bb_end-k-1], k);
2550 #ifndef PRODUCT
2551 if (_cfg->C->trace_opto_output()) {
2552 tty->print("# Schedule BB#%03d (final)\n", i);
2553 uint current = 0;
2554 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2555 Node *n = bb->get_node(j);
2556 if( valid_bundle_info(n) ) {
2557 Bundle *bundle = node_bundling(n);
2558 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2559 tty->print("*** Bundle: ");
2560 bundle->dump();
2561 }
2562 n->dump();
2563 }
2564 }
2565 }
2566 #endif
2567 #ifdef ASSERT
2568 verify_good_schedule(bb,"after block local scheduling");
2569 #endif
2570 }
2572 #ifndef PRODUCT
2573 if (_cfg->C->trace_opto_output())
2574 tty->print("# <- DoScheduling\n");
2575 #endif
2577 // Record final node-bundling array location
2578 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2580 } // end DoScheduling
2582 // Verify that no live-range used in the block is killed in the block by a
2583 // wrong DEF. This doesn't verify live-ranges that span blocks.
2585 // Check for edge existence. Used to avoid adding redundant precedence edges.
2586 static bool edge_from_to( Node *from, Node *to ) {
2587 for( uint i=0; i<from->len(); i++ )
2588 if( from->in(i) == to )
2589 return true;
2590 return false;
2591 }
2593 #ifdef ASSERT
2594 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2595 // Check for bad kills
2596 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2597 Node *prior_use = _reg_node[def];
2598 if( prior_use && !edge_from_to(prior_use,n) ) {
2599 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2600 n->dump();
2601 tty->print_cr("...");
2602 prior_use->dump();
2603 assert(edge_from_to(prior_use,n),msg);
2604 }
2605 _reg_node.map(def,NULL); // Kill live USEs
2606 }
2607 }
2609 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2611 // Zap to something reasonable for the verify code
2612 _reg_node.clear();
2614 // Walk over the block backwards. Check to make sure each DEF doesn't
2615 // kill a live value (other than the one it's supposed to). Add each
2616 // USE to the live set.
2617 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2618 Node *n = b->get_node(i);
2619 int n_op = n->Opcode();
2620 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2621 // Fat-proj kills a slew of registers
2622 RegMask rm = n->out_RegMask();// Make local copy
2623 while( rm.is_NotEmpty() ) {
2624 OptoReg::Name kill = rm.find_first_elem();
2625 rm.Remove(kill);
2626 verify_do_def( n, kill, msg );
2627 }
2628 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2629 // Get DEF'd registers the normal way
2630 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2631 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2632 }
2634 // Now make all USEs live
2635 for( uint i=1; i<n->req(); i++ ) {
2636 Node *def = n->in(i);
2637 assert(def != 0, "input edge required");
2638 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2639 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2640 if( OptoReg::is_valid(reg_lo) ) {
2641 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2642 _reg_node.map(reg_lo,n);
2643 }
2644 if( OptoReg::is_valid(reg_hi) ) {
2645 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2646 _reg_node.map(reg_hi,n);
2647 }
2648 }
2650 }
2652 // Zap to something reasonable for the Antidependence code
2653 _reg_node.clear();
2654 }
2655 #endif
2657 // Conditionally add precedence edges. Avoid putting edges on Projs.
2658 static void add_prec_edge_from_to( Node *from, Node *to ) {
2659 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2660 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2661 from = from->in(0);
2662 }
2663 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2664 !edge_from_to( from, to ) ) // Avoid duplicate edge
2665 from->add_prec(to);
2666 }
2668 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2669 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2670 return;
2672 Node *pinch = _reg_node[def_reg]; // Get pinch point
2673 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2674 is_def ) { // Check for a true def (not a kill)
2675 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2676 return;
2677 }
2679 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2680 debug_only( def = (Node*)0xdeadbeef; )
2682 // After some number of kills there _may_ be a later def
2683 Node *later_def = NULL;
2685 // Finding a kill requires a real pinch-point.
2686 // Check for not already having a pinch-point.
2687 // Pinch points are Op_Node's.
2688 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2689 later_def = pinch; // Must be def/kill as optimistic pinch-point
2690 if ( _pinch_free_list.size() > 0) {
2691 pinch = _pinch_free_list.pop();
2692 } else {
2693 pinch = new (_cfg->C) Node(1); // Pinch point to-be
2694 }
2695 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2696 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2697 return;
2698 }
2699 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2700 _reg_node.map(def_reg,pinch); // Record pinch-point
2701 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2702 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2703 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2704 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2705 later_def = NULL; // and no later def
2706 }
2707 pinch->set_req(0,later_def); // Hook later def so we can find it
2708 } else { // Else have valid pinch point
2709 if( pinch->in(0) ) // If there is a later-def
2710 later_def = pinch->in(0); // Get it
2711 }
2713 // Add output-dependence edge from later def to kill
2714 if( later_def ) // If there is some original def
2715 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2717 // See if current kill is also a use, and so is forced to be the pinch-point.
2718 if( pinch->Opcode() == Op_Node ) {
2719 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2720 for( uint i=1; i<uses->req(); i++ ) {
2721 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2722 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2723 // Yes, found a use/kill pinch-point
2724 pinch->set_req(0,NULL); //
2725 pinch->replace_by(kill); // Move anti-dep edges up
2726 pinch = kill;
2727 _reg_node.map(def_reg,pinch);
2728 return;
2729 }
2730 }
2731 }
2733 // Add edge from kill to pinch-point
2734 add_prec_edge_from_to(kill,pinch);
2735 }
2737 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2738 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2739 return;
2740 Node *pinch = _reg_node[use_reg]; // Get pinch point
2741 // Check for no later def_reg/kill in block
2742 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2743 // Use has to be block-local as well
2744 _cfg->get_block_for_node(use) == b) {
2745 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2746 pinch->req() == 1 ) { // pinch not yet in block?
2747 pinch->del_req(0); // yank pointer to later-def, also set flag
2748 // Insert the pinch-point in the block just after the last use
2749 b->insert_node(pinch, b->find_node(use) + 1);
2750 _bb_end++; // Increase size scheduled region in block
2751 }
2753 add_prec_edge_from_to(pinch,use);
2754 }
2755 }
2757 // We insert antidependences between the reads and following write of
2758 // allocated registers to prevent illegal code motion. Hopefully, the
2759 // number of added references should be fairly small, especially as we
2760 // are only adding references within the current basic block.
2761 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2763 #ifdef ASSERT
2764 verify_good_schedule(b,"before block local scheduling");
2765 #endif
2767 // A valid schedule, for each register independently, is an endless cycle
2768 // of: a def, then some uses (connected to the def by true dependencies),
2769 // then some kills (defs with no uses), finally the cycle repeats with a new
2770 // def. The uses are allowed to float relative to each other, as are the
2771 // kills. No use is allowed to slide past a kill (or def). This requires
2772 // antidependencies between all uses of a single def and all kills that
2773 // follow, up to the next def. More edges are redundant, because later defs
2774 // & kills are already serialized with true or antidependencies. To keep
2775 // the edge count down, we add a 'pinch point' node if there's more than
2776 // one use or more than one kill/def.
2778 // We add dependencies in one bottom-up pass.
2780 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2782 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2783 // register. If not, we record the DEF/KILL in _reg_node, the
2784 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2785 // "pinch point", a new Node that's in the graph but not in the block.
2786 // We put edges from the prior and current DEF/KILLs to the pinch point.
2787 // We put the pinch point in _reg_node. If there's already a pinch point
2788 // we merely add an edge from the current DEF/KILL to the pinch point.
2790 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2791 // put an edge from the pinch point to the USE.
2793 // To be expedient, the _reg_node array is pre-allocated for the whole
2794 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2795 // or a valid def/kill/pinch-point, or a leftover node from some prior
2796 // block. Leftover node from some prior block is treated like a NULL (no
2797 // prior def, so no anti-dependence needed). Valid def is distinguished by
2798 // it being in the current block.
2799 bool fat_proj_seen = false;
2800 uint last_safept = _bb_end-1;
2801 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2802 Node* last_safept_node = end_node;
2803 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2804 Node *n = b->get_node(i);
2805 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2806 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2807 // Fat-proj kills a slew of registers
2808 // This can add edges to 'n' and obscure whether or not it was a def,
2809 // hence the is_def flag.
2810 fat_proj_seen = true;
2811 RegMask rm = n->out_RegMask();// Make local copy
2812 while( rm.is_NotEmpty() ) {
2813 OptoReg::Name kill = rm.find_first_elem();
2814 rm.Remove(kill);
2815 anti_do_def( b, n, kill, is_def );
2816 }
2817 } else {
2818 // Get DEF'd registers the normal way
2819 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2820 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2821 }
2823 // Kill projections on a branch should appear to occur on the
2824 // branch, not afterwards, so grab the masks from the projections
2825 // and process them.
2826 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2827 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2828 Node* use = n->fast_out(i);
2829 if (use->is_Proj()) {
2830 RegMask rm = use->out_RegMask();// Make local copy
2831 while( rm.is_NotEmpty() ) {
2832 OptoReg::Name kill = rm.find_first_elem();
2833 rm.Remove(kill);
2834 anti_do_def( b, n, kill, false );
2835 }
2836 }
2837 }
2838 }
2840 // Check each register used by this instruction for a following DEF/KILL
2841 // that must occur afterward and requires an anti-dependence edge.
2842 for( uint j=0; j<n->req(); j++ ) {
2843 Node *def = n->in(j);
2844 if( def ) {
2845 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2846 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2847 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2848 }
2849 }
2850 // Do not allow defs of new derived values to float above GC
2851 // points unless the base is definitely available at the GC point.
2853 Node *m = b->get_node(i);
2855 // Add precedence edge from following safepoint to use of derived pointer
2856 if( last_safept_node != end_node &&
2857 m != last_safept_node) {
2858 for (uint k = 1; k < m->req(); k++) {
2859 const Type *t = m->in(k)->bottom_type();
2860 if( t->isa_oop_ptr() &&
2861 t->is_ptr()->offset() != 0 ) {
2862 last_safept_node->add_prec( m );
2863 break;
2864 }
2865 }
2866 }
2868 if( n->jvms() ) { // Precedence edge from derived to safept
2869 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2870 if( b->get_node(last_safept) != last_safept_node ) {
2871 last_safept = b->find_node(last_safept_node);
2872 }
2873 for( uint j=last_safept; j > i; j-- ) {
2874 Node *mach = b->get_node(j);
2875 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2876 mach->add_prec( n );
2877 }
2878 last_safept = i;
2879 last_safept_node = m;
2880 }
2881 }
2883 if (fat_proj_seen) {
2884 // Garbage collect pinch nodes that were not consumed.
2885 // They are usually created by a fat kill MachProj for a call.
2886 garbage_collect_pinch_nodes();
2887 }
2888 }
2890 // Garbage collect pinch nodes for reuse by other blocks.
2891 //
2892 // The block scheduler's insertion of anti-dependence
2893 // edges creates many pinch nodes when the block contains
2894 // 2 or more Calls. A pinch node is used to prevent a
2895 // combinatorial explosion of edges. If a set of kills for a
2896 // register is anti-dependent on a set of uses (or defs), rather
2897 // than adding an edge in the graph between each pair of kill
2898 // and use (or def), a pinch is inserted between them:
2899 //
2900 // use1 use2 use3
2901 // \ | /
2902 // \ | /
2903 // pinch
2904 // / | \
2905 // / | \
2906 // kill1 kill2 kill3
2907 //
2908 // One pinch node is created per register killed when
2909 // the second call is encountered during a backwards pass
2910 // over the block. Most of these pinch nodes are never
2911 // wired into the graph because the register is never
2912 // used or def'ed in the block.
2913 //
2914 void Scheduling::garbage_collect_pinch_nodes() {
2915 #ifndef PRODUCT
2916 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2917 #endif
2918 int trace_cnt = 0;
2919 for (uint k = 0; k < _reg_node.Size(); k++) {
2920 Node* pinch = _reg_node[k];
2921 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2922 // no predecence input edges
2923 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2924 cleanup_pinch(pinch);
2925 _pinch_free_list.push(pinch);
2926 _reg_node.map(k, NULL);
2927 #ifndef PRODUCT
2928 if (_cfg->C->trace_opto_output()) {
2929 trace_cnt++;
2930 if (trace_cnt > 40) {
2931 tty->print("\n");
2932 trace_cnt = 0;
2933 }
2934 tty->print(" %d", pinch->_idx);
2935 }
2936 #endif
2937 }
2938 }
2939 #ifndef PRODUCT
2940 if (_cfg->C->trace_opto_output()) tty->print("\n");
2941 #endif
2942 }
2944 // Clean up a pinch node for reuse.
2945 void Scheduling::cleanup_pinch( Node *pinch ) {
2946 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2948 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2949 Node* use = pinch->last_out(i);
2950 uint uses_found = 0;
2951 for (uint j = use->req(); j < use->len(); j++) {
2952 if (use->in(j) == pinch) {
2953 use->rm_prec(j);
2954 uses_found++;
2955 }
2956 }
2957 assert(uses_found > 0, "must be a precedence edge");
2958 i -= uses_found; // we deleted 1 or more copies of this edge
2959 }
2960 // May have a later_def entry
2961 pinch->set_req(0, NULL);
2962 }
2964 #ifndef PRODUCT
2966 void Scheduling::dump_available() const {
2967 tty->print("#Availist ");
2968 for (uint i = 0; i < _available.size(); i++)
2969 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2970 tty->cr();
2971 }
2973 // Print Scheduling Statistics
2974 void Scheduling::print_statistics() {
2975 // Print the size added by nops for bundling
2976 tty->print("Nops added %d bytes to total of %d bytes",
2977 _total_nop_size, _total_method_size);
2978 if (_total_method_size > 0)
2979 tty->print(", for %.2f%%",
2980 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2981 tty->print("\n");
2983 // Print the number of branch shadows filled
2984 if (Pipeline::_branch_has_delay_slot) {
2985 tty->print("Of %d branches, %d had unconditional delay slots filled",
2986 _total_branches, _total_unconditional_delays);
2987 if (_total_branches > 0)
2988 tty->print(", for %.2f%%",
2989 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2990 tty->print("\n");
2991 }
2993 uint total_instructions = 0, total_bundles = 0;
2995 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2996 uint bundle_count = _total_instructions_per_bundle[i];
2997 total_instructions += bundle_count * i;
2998 total_bundles += bundle_count;
2999 }
3001 if (total_bundles > 0)
3002 tty->print("Average ILP (excluding nops) is %.2f\n",
3003 ((double)total_instructions) / ((double)total_bundles));
3004 }
3005 #endif