1 /*
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/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"
44
45 extern uint size_exception_handler();
46 extern uint size_deopt_handler();
47
48 #ifndef PRODUCT
49 #define DEBUG_ARG(x) , x
50 #else
51 #define DEBUG_ARG(x)
52 #endif
53
54 extern int emit_exception_handler(CodeBuffer &cbuf);
55 extern int emit_deopt_handler(CodeBuffer &cbuf);
56
57 //------------------------------Output-----------------------------------------
58 // Convert Nodes to instruction bits and pass off to the VM
59 void Compile::Output() {
60 // RootNode goes
61 assert( _cfg->_broot->_nodes.size() == 0, "" );
62
63 // The number of new nodes (mostly MachNop) is proportional to
64 // the number of java calls and inner loops which are aligned.
65 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
66 C->inner_loops()*(OptoLoopAlignment-1)),
67 "out of nodes before code generation" ) ) {
68 return;
69 }
70 // Make sure I can find the Start Node
71 Block *entry = _cfg->_blocks[1];
72 Block *broot = _cfg->_broot;
73
74 const StartNode *start = entry->_nodes[0]->as_Start();
75
76 // Replace StartNode with prolog
77 MachPrologNode *prolog = new (this) MachPrologNode();
78 entry->_nodes.map( 0, prolog );
79 _cfg->map_node_to_block(prolog, entry);
80 _cfg->unmap_node_from_block(start); // start is no longer in any block
81
82 // Virtual methods need an unverified entry point
83
84 if( is_osr_compilation() ) {
85 if( PoisonOSREntry ) {
86 // TODO: Should use a ShouldNotReachHereNode...
87 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
88 }
89 } else {
90 if( _method && !_method->flags().is_static() ) {
91 // Insert unvalidated entry point
92 _cfg->insert( broot, 0, new (this) MachUEPNode() );
93 }
94
95 }
96
97
98 // Break before main entry point
99 if( (_method && _method->break_at_execute())
100 #ifndef PRODUCT
101 ||(OptoBreakpoint && is_method_compilation())
102 ||(OptoBreakpointOSR && is_osr_compilation())
103 ||(OptoBreakpointC2R && !_method)
104 #endif
105 ) {
106 // checking for _method means that OptoBreakpoint does not apply to
107 // runtime stubs or frame converters
108 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
109 }
110
111 // Insert epilogs before every return
112 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
113 Block *b = _cfg->_blocks[i];
114 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point?
115 Node *m = b->end();
116 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) {
117 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
118 b->add_inst( epilog );
119 _cfg->map_node_to_block(epilog, b);
120 }
121 }
122 }
123
124 # ifdef ENABLE_ZAP_DEAD_LOCALS
125 if ( ZapDeadCompiledLocals ) Insert_zap_nodes();
126 # endif
127
128 uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
129 blk_starts[0] = 0;
130
131 // Initialize code buffer and process short branches.
132 CodeBuffer* cb = init_buffer(blk_starts);
133
134 if (cb == NULL || failing()) return;
135
136 ScheduleAndBundle();
137
138 #ifndef PRODUCT
139 if (trace_opto_output()) {
140 tty->print("\n---- After ScheduleAndBundle ----\n");
141 for (uint i = 0; i < _cfg->_num_blocks; i++) {
142 tty->print("\nBB#%03d:\n", i);
143 Block *bb = _cfg->_blocks[i];
144 for (uint j = 0; j < bb->_nodes.size(); j++) {
145 Node *n = bb->_nodes[j];
146 OptoReg::Name reg = _regalloc->get_reg_first(n);
147 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
148 n->dump();
149 }
150 }
151 }
152 #endif
153
154 if (failing()) return;
155
156 BuildOopMaps();
157
158 if (failing()) return;
159
160 fill_buffer(cb, blk_starts);
161 }
162
163 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
164 // Determine if we need to generate a stack overflow check.
165 // Do it if the method is not a stub function and
166 // has java calls or has frame size > vm_page_size/8.
167 return (UseStackBanging && stub_function() == NULL &&
168 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
169 }
170
171 bool Compile::need_register_stack_bang() const {
172 // Determine if we need to generate a register stack overflow check.
173 // This is only used on architectures which have split register
174 // and memory stacks (ie. IA64).
175 // Bang if the method is not a stub function and has java calls
176 return (stub_function() == NULL && has_java_calls());
177 }
178
179 # ifdef ENABLE_ZAP_DEAD_LOCALS
180
181
182 // In order to catch compiler oop-map bugs, we have implemented
183 // a debugging mode called ZapDeadCompilerLocals.
184 // This mode causes the compiler to insert a call to a runtime routine,
185 // "zap_dead_locals", right before each place in compiled code
186 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
187 // The runtime routine checks that locations mapped as oops are really
188 // oops, that locations mapped as values do not look like oops,
189 // and that locations mapped as dead are not used later
190 // (by zapping them to an invalid address).
191
192 int Compile::_CompiledZap_count = 0;
193
194 void Compile::Insert_zap_nodes() {
195 bool skip = false;
196
197
198 // Dink with static counts because code code without the extra
199 // runtime calls is MUCH faster for debugging purposes
200
201 if ( CompileZapFirst == 0 ) ; // nothing special
202 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
203 else if ( CompileZapFirst == CompiledZap_count() )
204 warning("starting zap compilation after skipping");
205
206 if ( CompileZapLast == -1 ) ; // nothing special
207 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
208 else if ( CompileZapLast == CompiledZap_count() )
209 warning("about to compile last zap");
210
211 ++_CompiledZap_count; // counts skipped zaps, too
212
213 if ( skip ) return;
214
215
216 if ( _method == NULL )
217 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
218
219 // Insert call to zap runtime stub before every node with an oop map
220 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
221 Block *b = _cfg->_blocks[i];
222 for ( uint j = 0; j < b->_nodes.size(); ++j ) {
223 Node *n = b->_nodes[j];
224
225 // Determining if we should insert a zap-a-lot node in output.
226 // We do that for all nodes that has oopmap info, except for calls
227 // to allocation. Calls to allocation passes in the old top-of-eden pointer
228 // and expect the C code to reset it. Hence, there can be no safepoints between
229 // the inlined-allocation and the call to new_Java, etc.
230 // We also cannot zap monitor calls, as they must hold the microlock
231 // during the call to Zap, which also wants to grab the microlock.
232 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
233 if ( insert ) { // it is MachSafePoint
234 if ( !n->is_MachCall() ) {
235 insert = false;
236 } else if ( n->is_MachCall() ) {
237 MachCallNode* call = n->as_MachCall();
238 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
239 call->entry_point() == OptoRuntime::new_array_Java() ||
240 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
241 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
242 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
243 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
244 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
245 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
246 ) {
247 insert = false;
248 }
249 }
250 if (insert) {
251 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
252 b->_nodes.insert( j, zap );
253 _cfg->map_node_to_block(zap, b);
254 ++j;
255 }
256 }
257 }
258 }
259 }
260
261
262 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
263 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
264 CallStaticJavaNode* ideal_node =
265 new (this) CallStaticJavaNode( tf,
266 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
267 "call zap dead locals stub", 0, TypePtr::BOTTOM);
268 // We need to copy the OopMap from the site we're zapping at.
269 // We have to make a copy, because the zap site might not be
270 // a call site, and zap_dead is a call site.
271 OopMap* clone = node_to_check->oop_map()->deep_copy();
272
273 // Add the cloned OopMap to the zap node
274 ideal_node->set_oop_map(clone);
275 return _matcher->match_sfpt(ideal_node);
276 }
277
278 //------------------------------is_node_getting_a_safepoint--------------------
279 bool Compile::is_node_getting_a_safepoint( Node* n) {
280 // This code duplicates the logic prior to the call of add_safepoint
281 // below in this file.
282 if( n->is_MachSafePoint() ) return true;
283 return false;
284 }
285
286 # endif // ENABLE_ZAP_DEAD_LOCALS
287
288 //------------------------------compute_loop_first_inst_sizes------------------
289 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
290 // of a loop. When aligning a loop we need to provide enough instructions
291 // in cpu's fetch buffer to feed decoders. The loop alignment could be
292 // avoided if we have enough instructions in fetch buffer at the head of a loop.
293 // By default, the size is set to 999999 by Block's constructor so that
294 // a loop will be aligned if the size is not reset here.
295 //
296 // Note: Mach instructions could contain several HW instructions
297 // so the size is estimated only.
298 //
299 void Compile::compute_loop_first_inst_sizes() {
300 // The next condition is used to gate the loop alignment optimization.
301 // Don't aligned a loop if there are enough instructions at the head of a loop
302 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
303 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
304 // equal to 11 bytes which is the largest address NOP instruction.
305 if( MaxLoopPad < OptoLoopAlignment-1 ) {
306 uint last_block = _cfg->_num_blocks-1;
307 for( uint i=1; i <= last_block; i++ ) {
308 Block *b = _cfg->_blocks[i];
309 // Check the first loop's block which requires an alignment.
310 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {
311 uint sum_size = 0;
312 uint inst_cnt = NumberOfLoopInstrToAlign;
313 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
314
315 // Check subsequent fallthrough blocks if the loop's first
316 // block(s) does not have enough instructions.
317 Block *nb = b;
318 while( inst_cnt > 0 &&
319 i < last_block &&
320 !_cfg->_blocks[i+1]->has_loop_alignment() &&
321 !nb->has_successor(b) ) {
322 i++;
323 nb = _cfg->_blocks[i];
324 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
325 } // while( inst_cnt > 0 && i < last_block )
326
327 b->set_first_inst_size(sum_size);
328 } // f( b->head()->is_Loop() )
329 } // for( i <= last_block )
330 } // if( MaxLoopPad < OptoLoopAlignment-1 )
331 }
332
333 //----------------------shorten_branches---------------------------------------
334 // The architecture description provides short branch variants for some long
335 // branch instructions. Replace eligible long branches with short branches.
336 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
337
338 // ------------------
339 // Compute size of each block, method size, and relocation information size
340 uint nblocks = _cfg->_num_blocks;
341
342 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
343 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
344 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
345 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
346 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
347
348 bool has_short_branch_candidate = false;
349
350 // Initialize the sizes to 0
351 code_size = 0; // Size in bytes of generated code
352 stub_size = 0; // Size in bytes of all stub entries
353 // Size in bytes of all relocation entries, including those in local stubs.
354 // Start with 2-bytes of reloc info for the unvalidated entry point
355 reloc_size = 1; // Number of relocation entries
356
357 // Make three passes. The first computes pessimistic blk_starts,
358 // relative jmp_offset and reloc_size information. The second performs
359 // short branch substitution using the pessimistic sizing. The
360 // third inserts nops where needed.
361
362 // Step one, perform a pessimistic sizing pass.
363 uint last_call_adr = max_uint;
364 uint last_avoid_back_to_back_adr = max_uint;
365 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
366 for (uint i = 0; i < nblocks; i++) { // For all blocks
367 Block *b = _cfg->_blocks[i];
368
369 // During short branch replacement, we store the relative (to blk_starts)
370 // offset of jump in jmp_offset, rather than the absolute offset of jump.
371 // This is so that we do not need to recompute sizes of all nodes when
372 // we compute correct blk_starts in our next sizing pass.
373 jmp_offset[i] = 0;
374 jmp_size[i] = 0;
375 jmp_nidx[i] = -1;
376 DEBUG_ONLY( jmp_target[i] = 0; )
377 DEBUG_ONLY( jmp_rule[i] = 0; )
378
379 // Sum all instruction sizes to compute block size
380 uint last_inst = b->_nodes.size();
381 uint blk_size = 0;
382 for (uint j = 0; j < last_inst; j++) {
383 Node* nj = b->_nodes[j];
384 // Handle machine instruction nodes
385 if (nj->is_Mach()) {
386 MachNode *mach = nj->as_Mach();
387 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
388 reloc_size += mach->reloc();
389 if (mach->is_MachCall()) {
390 MachCallNode *mcall = mach->as_MachCall();
391 // This destination address is NOT PC-relative
392
393 mcall->method_set((intptr_t)mcall->entry_point());
394
395 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
396 stub_size += CompiledStaticCall::to_interp_stub_size();
397 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
398 }
399 } else if (mach->is_MachSafePoint()) {
400 // If call/safepoint are adjacent, account for possible
401 // nop to disambiguate the two safepoints.
402 // ScheduleAndBundle() can rearrange nodes in a block,
403 // check for all offsets inside this block.
404 if (last_call_adr >= blk_starts[i]) {
405 blk_size += nop_size;
406 }
407 }
408 if (mach->avoid_back_to_back()) {
409 // Nop is inserted between "avoid back to back" instructions.
410 // ScheduleAndBundle() can rearrange nodes in a block,
411 // check for all offsets inside this block.
412 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
413 blk_size += nop_size;
414 }
415 }
416 if (mach->may_be_short_branch()) {
417 if (!nj->is_MachBranch()) {
418 #ifndef PRODUCT
419 nj->dump(3);
420 #endif
421 Unimplemented();
422 }
423 assert(jmp_nidx[i] == -1, "block should have only one branch");
424 jmp_offset[i] = blk_size;
425 jmp_size[i] = nj->size(_regalloc);
426 jmp_nidx[i] = j;
427 has_short_branch_candidate = true;
428 }
429 }
430 blk_size += nj->size(_regalloc);
431 // Remember end of call offset
432 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
433 last_call_adr = blk_starts[i]+blk_size;
434 }
435 // Remember end of avoid_back_to_back offset
436 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
437 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
438 }
439 }
440
441 // When the next block starts a loop, we may insert pad NOP
442 // instructions. Since we cannot know our future alignment,
443 // assume the worst.
444 if (i< nblocks-1) {
445 Block *nb = _cfg->_blocks[i+1];
446 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
447 if (max_loop_pad > 0) {
448 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
449 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
450 // If either is the last instruction in this block, bump by
451 // max_loop_pad in lock-step with blk_size, so sizing
452 // calculations in subsequent blocks still can conservatively
453 // detect that it may the last instruction in this block.
454 if (last_call_adr == blk_starts[i]+blk_size) {
455 last_call_adr += max_loop_pad;
456 }
457 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
458 last_avoid_back_to_back_adr += max_loop_pad;
459 }
460 blk_size += max_loop_pad;
461 }
462 }
463
464 // Save block size; update total method size
465 blk_starts[i+1] = blk_starts[i]+blk_size;
466 }
467
468 // Step two, replace eligible long jumps.
469 bool progress = true;
470 uint last_may_be_short_branch_adr = max_uint;
471 while (has_short_branch_candidate && progress) {
472 progress = false;
473 has_short_branch_candidate = false;
474 int adjust_block_start = 0;
475 for (uint i = 0; i < nblocks; i++) {
476 Block *b = _cfg->_blocks[i];
477 int idx = jmp_nidx[i];
478 MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach();
479 if (mach != NULL && mach->may_be_short_branch()) {
480 #ifdef ASSERT
481 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
482 int j;
483 // Find the branch; ignore trailing NOPs.
484 for (j = b->_nodes.size()-1; j>=0; j--) {
485 Node* n = b->_nodes[j];
486 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
487 break;
488 }
489 assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity");
490 #endif
491 int br_size = jmp_size[i];
492 int br_offs = blk_starts[i] + jmp_offset[i];
493
494 // This requires the TRUE branch target be in succs[0]
495 uint bnum = b->non_connector_successor(0)->_pre_order;
496 int offset = blk_starts[bnum] - br_offs;
497 if (bnum > i) { // adjust following block's offset
498 offset -= adjust_block_start;
499 }
500 // In the following code a nop could be inserted before
501 // the branch which will increase the backward distance.
502 bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);
503 if (needs_padding && offset <= 0)
504 offset -= nop_size;
505
506 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
507 // We've got a winner. Replace this branch.
508 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
509
510 // Update the jmp_size.
511 int new_size = replacement->size(_regalloc);
512 int diff = br_size - new_size;
513 assert(diff >= (int)nop_size, "short_branch size should be smaller");
514 // Conservatively take into accound padding between
515 // avoid_back_to_back branches. Previous branch could be
516 // converted into avoid_back_to_back branch during next
517 // rounds.
518 if (needs_padding && replacement->avoid_back_to_back()) {
519 jmp_offset[i] += nop_size;
520 diff -= nop_size;
521 }
522 adjust_block_start += diff;
523 b->_nodes.map(idx, replacement);
524 mach->subsume_by(replacement, C);
525 mach = replacement;
526 progress = true;
527
528 jmp_size[i] = new_size;
529 DEBUG_ONLY( jmp_target[i] = bnum; );
530 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
531 } else {
532 // The jump distance is not short, try again during next iteration.
533 has_short_branch_candidate = true;
534 }
535 } // (mach->may_be_short_branch())
536 if (mach != NULL && (mach->may_be_short_branch() ||
537 mach->avoid_back_to_back())) {
538 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
539 }
540 blk_starts[i+1] -= adjust_block_start;
541 }
542 }
543
544 #ifdef ASSERT
545 for (uint i = 0; i < nblocks; i++) { // For all blocks
546 if (jmp_target[i] != 0) {
547 int br_size = jmp_size[i];
548 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
549 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
550 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]);
551 }
552 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
553 }
554 }
555 #endif
556
557 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
558 // after ScheduleAndBundle().
559
560 // ------------------
561 // Compute size for code buffer
562 code_size = blk_starts[nblocks];
563
564 // Relocation records
565 reloc_size += 1; // Relo entry for exception handler
566
567 // Adjust reloc_size to number of record of relocation info
568 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
569 // a relocation index.
570 // The CodeBuffer will expand the locs array if this estimate is too low.
571 reloc_size *= 10 / sizeof(relocInfo);
572 }
573
574 //------------------------------FillLocArray-----------------------------------
575 // Create a bit of debug info and append it to the array. The mapping is from
576 // Java local or expression stack to constant, register or stack-slot. For
577 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
578 // entry has been taken care of and caller should skip it).
579 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
580 // This should never have accepted Bad before
581 assert(OptoReg::is_valid(regnum), "location must be valid");
582 return (OptoReg::is_reg(regnum))
583 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
584 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
585 }
586
587
588 ObjectValue*
589 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
590 for (int i = 0; i < objs->length(); i++) {
591 assert(objs->at(i)->is_object(), "corrupt object cache");
592 ObjectValue* sv = (ObjectValue*) objs->at(i);
593 if (sv->id() == id) {
594 return sv;
595 }
596 }
597 // Otherwise..
598 return NULL;
599 }
600
601 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
602 ObjectValue* sv ) {
603 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
604 objs->append(sv);
605 }
606
607
608 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
609 GrowableArray<ScopeValue*> *array,
610 GrowableArray<ScopeValue*> *objs ) {
611 assert( local, "use _top instead of null" );
612 if (array->length() != idx) {
613 assert(array->length() == idx + 1, "Unexpected array count");
614 // Old functionality:
615 // return
616 // New functionality:
617 // Assert if the local is not top. In product mode let the new node
618 // override the old entry.
619 assert(local == top(), "LocArray collision");
620 if (local == top()) {
621 return;
622 }
623 array->pop();
624 }
625 const Type *t = local->bottom_type();
626
627 // Is it a safepoint scalar object node?
628 if (local->is_SafePointScalarObject()) {
629 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
630
631 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
632 if (sv == NULL) {
633 ciKlass* cik = t->is_oopptr()->klass();
634 assert(cik->is_instance_klass() ||
635 cik->is_array_klass(), "Not supported allocation.");
636 sv = new ObjectValue(spobj->_idx,
637 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
638 Compile::set_sv_for_object_node(objs, sv);
639
640 uint first_ind = spobj->first_index();
641 for (uint i = 0; i < spobj->n_fields(); i++) {
642 Node* fld_node = sfpt->in(first_ind+i);
643 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
644 }
645 }
646 array->append(sv);
647 return;
648 }
649
650 // Grab the register number for the local
651 OptoReg::Name regnum = _regalloc->get_reg_first(local);
652 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
653 // Record the double as two float registers.
654 // The register mask for such a value always specifies two adjacent
655 // float registers, with the lower register number even.
656 // Normally, the allocation of high and low words to these registers
657 // is irrelevant, because nearly all operations on register pairs
658 // (e.g., StoreD) treat them as a single unit.
659 // Here, we assume in addition that the words in these two registers
660 // stored "naturally" (by operations like StoreD and double stores
661 // within the interpreter) such that the lower-numbered register
662 // is written to the lower memory address. This may seem like
663 // a machine dependency, but it is not--it is a requirement on
664 // the author of the <arch>.ad file to ensure that, for every
665 // even/odd double-register pair to which a double may be allocated,
666 // the word in the even single-register is stored to the first
667 // memory word. (Note that register numbers are completely
668 // arbitrary, and are not tied to any machine-level encodings.)
669 #ifdef _LP64
670 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
671 array->append(new ConstantIntValue(0));
672 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
673 } else if ( t->base() == Type::Long ) {
674 array->append(new ConstantIntValue(0));
675 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
676 } else if ( t->base() == Type::RawPtr ) {
677 // jsr/ret return address which must be restored into a the full
678 // width 64-bit stack slot.
679 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
680 }
681 #else //_LP64
682 #ifdef SPARC
683 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
684 // For SPARC we have to swap high and low words for
685 // long values stored in a single-register (g0-g7).
686 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
687 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
688 } else
689 #endif //SPARC
690 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
691 // Repack the double/long as two jints.
692 // The convention the interpreter uses is that the second local
693 // holds the first raw word of the native double representation.
694 // This is actually reasonable, since locals and stack arrays
695 // grow downwards in all implementations.
696 // (If, on some machine, the interpreter's Java locals or stack
697 // were to grow upwards, the embedded doubles would be word-swapped.)
698 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
699 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
700 }
701 #endif //_LP64
702 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
703 OptoReg::is_reg(regnum) ) {
704 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
705 ? Location::float_in_dbl : Location::normal ));
706 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
707 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
708 ? Location::int_in_long : Location::normal ));
709 } else if( t->base() == Type::NarrowOop ) {
710 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
711 } else {
712 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
713 }
714 return;
715 }
716
717 // No register. It must be constant data.
718 switch (t->base()) {
719 case Type::Half: // Second half of a double
720 ShouldNotReachHere(); // Caller should skip 2nd halves
721 break;
722 case Type::AnyPtr:
723 array->append(new ConstantOopWriteValue(NULL));
724 break;
725 case Type::AryPtr:
726 case Type::InstPtr: // fall through
727 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
728 break;
729 case Type::NarrowOop:
730 if (t == TypeNarrowOop::NULL_PTR) {
731 array->append(new ConstantOopWriteValue(NULL));
732 } else {
733 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
734 }
735 break;
736 case Type::Int:
737 array->append(new ConstantIntValue(t->is_int()->get_con()));
738 break;
739 case Type::RawPtr:
740 // A return address (T_ADDRESS).
741 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
742 #ifdef _LP64
743 // Must be restored to the full-width 64-bit stack slot.
744 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
745 #else
746 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
747 #endif
748 break;
749 case Type::FloatCon: {
750 float f = t->is_float_constant()->getf();
751 array->append(new ConstantIntValue(jint_cast(f)));
752 break;
753 }
754 case Type::DoubleCon: {
755 jdouble d = t->is_double_constant()->getd();
756 #ifdef _LP64
757 array->append(new ConstantIntValue(0));
758 array->append(new ConstantDoubleValue(d));
759 #else
760 // Repack the double as two jints.
761 // The convention the interpreter uses is that the second local
762 // holds the first raw word of the native double representation.
763 // This is actually reasonable, since locals and stack arrays
764 // grow downwards in all implementations.
765 // (If, on some machine, the interpreter's Java locals or stack
766 // were to grow upwards, the embedded doubles would be word-swapped.)
767 jint *dp = (jint*)&d;
768 array->append(new ConstantIntValue(dp[1]));
769 array->append(new ConstantIntValue(dp[0]));
770 #endif
771 break;
772 }
773 case Type::Long: {
774 jlong d = t->is_long()->get_con();
775 #ifdef _LP64
776 array->append(new ConstantIntValue(0));
777 array->append(new ConstantLongValue(d));
778 #else
779 // Repack the long 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 jint *dp = (jint*)&d;
787 array->append(new ConstantIntValue(dp[1]));
788 array->append(new ConstantIntValue(dp[0]));
789 #endif
790 break;
791 }
792 case Type::Top: // Add an illegal value here
793 array->append(new LocationValue(Location()));
794 break;
795 default:
796 ShouldNotReachHere();
797 break;
798 }
799 }
800
801 // Determine if this node starts a bundle
802 bool Compile::starts_bundle(const Node *n) const {
803 return (_node_bundling_limit > n->_idx &&
804 _node_bundling_base[n->_idx].starts_bundle());
805 }
806
807 //--------------------------Process_OopMap_Node--------------------------------
808 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
809
810 // Handle special safepoint nodes for synchronization
811 MachSafePointNode *sfn = mach->as_MachSafePoint();
812 MachCallNode *mcall;
813
814 #ifdef ENABLE_ZAP_DEAD_LOCALS
815 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
816 #endif
817
818 int safepoint_pc_offset = current_offset;
819 bool is_method_handle_invoke = false;
820 bool return_oop = false;
821
822 // Add the safepoint in the DebugInfoRecorder
823 if( !mach->is_MachCall() ) {
824 mcall = NULL;
825 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
826 } else {
827 mcall = mach->as_MachCall();
828
829 // Is the call a MethodHandle call?
830 if (mcall->is_MachCallJava()) {
831 if (mcall->as_MachCallJava()->_method_handle_invoke) {
832 assert(has_method_handle_invokes(), "must have been set during call generation");
833 is_method_handle_invoke = true;
834 }
835 }
836
837 // Check if a call returns an object.
838 if (mcall->return_value_is_used() &&
839 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
840 return_oop = true;
841 }
842 safepoint_pc_offset += mcall->ret_addr_offset();
843 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
844 }
845
846 // Loop over the JVMState list to add scope information
847 // Do not skip safepoints with a NULL method, they need monitor info
848 JVMState* youngest_jvms = sfn->jvms();
849 int max_depth = youngest_jvms->depth();
850
851 // Allocate the object pool for scalar-replaced objects -- the map from
852 // small-integer keys (which can be recorded in the local and ostack
853 // arrays) to descriptions of the object state.
854 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
855
856 // Visit scopes from oldest to youngest.
857 for (int depth = 1; depth <= max_depth; depth++) {
858 JVMState* jvms = youngest_jvms->of_depth(depth);
859 int idx;
860 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
861 // Safepoints that do not have method() set only provide oop-map and monitor info
862 // to support GC; these do not support deoptimization.
863 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
864 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
865 int num_mon = jvms->nof_monitors();
866 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
867 "JVMS local count must match that of the method");
868
869 // Add Local and Expression Stack Information
870
871 // Insert locals into the locarray
872 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
873 for( idx = 0; idx < num_locs; idx++ ) {
874 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
875 }
876
877 // Insert expression stack entries into the exparray
878 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
879 for( idx = 0; idx < num_exps; idx++ ) {
880 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
881 }
882
883 // Add in mappings of the monitors
884 assert( !method ||
885 !method->is_synchronized() ||
886 method->is_native() ||
887 num_mon > 0 ||
888 !GenerateSynchronizationCode,
889 "monitors must always exist for synchronized methods");
890
891 // Build the growable array of ScopeValues for exp stack
892 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
893
894 // Loop over monitors and insert into array
895 for(idx = 0; idx < num_mon; idx++) {
896 // Grab the node that defines this monitor
897 Node* box_node = sfn->monitor_box(jvms, idx);
898 Node* obj_node = sfn->monitor_obj(jvms, idx);
899
900 // Create ScopeValue for object
901 ScopeValue *scval = NULL;
902
903 if( obj_node->is_SafePointScalarObject() ) {
904 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
905 scval = Compile::sv_for_node_id(objs, spobj->_idx);
906 if (scval == NULL) {
907 const Type *t = obj_node->bottom_type();
908 ciKlass* cik = t->is_oopptr()->klass();
909 assert(cik->is_instance_klass() ||
910 cik->is_array_klass(), "Not supported allocation.");
911 ObjectValue* sv = new ObjectValue(spobj->_idx,
912 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
913 Compile::set_sv_for_object_node(objs, sv);
914
915 uint first_ind = spobj->first_index();
916 for (uint i = 0; i < spobj->n_fields(); i++) {
917 Node* fld_node = sfn->in(first_ind+i);
918 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
919 }
920 scval = sv;
921 }
922 } else if( !obj_node->is_Con() ) {
923 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
924 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
925 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
926 } else {
927 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
928 }
929 } else {
930 const TypePtr *tp = obj_node->get_ptr_type();
931 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
932 }
933
934 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
935 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
936 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
937 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
938 }
939
940 // We dump the object pool first, since deoptimization reads it in first.
941 debug_info()->dump_object_pool(objs);
942
943 // Build first class objects to pass to scope
944 DebugToken *locvals = debug_info()->create_scope_values(locarray);
945 DebugToken *expvals = debug_info()->create_scope_values(exparray);
946 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
947
948 // Make method available for all Safepoints
949 ciMethod* scope_method = method ? method : _method;
950 // Describe the scope here
951 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
952 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
953 // Now we can describe the scope.
954 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
955 } // End jvms loop
956
957 // Mark the end of the scope set.
958 debug_info()->end_safepoint(safepoint_pc_offset);
959 }
960
961
962
963 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
964 class NonSafepointEmitter {
965 Compile* C;
966 JVMState* _pending_jvms;
967 int _pending_offset;
968
969 void emit_non_safepoint();
970
971 public:
972 NonSafepointEmitter(Compile* compile) {
973 this->C = compile;
974 _pending_jvms = NULL;
975 _pending_offset = 0;
976 }
977
978 void observe_instruction(Node* n, int pc_offset) {
979 if (!C->debug_info()->recording_non_safepoints()) return;
980
981 Node_Notes* nn = C->node_notes_at(n->_idx);
982 if (nn == NULL || nn->jvms() == NULL) return;
983 if (_pending_jvms != NULL &&
984 _pending_jvms->same_calls_as(nn->jvms())) {
985 // Repeated JVMS? Stretch it up here.
986 _pending_offset = pc_offset;
987 } else {
988 if (_pending_jvms != NULL &&
989 _pending_offset < pc_offset) {
990 emit_non_safepoint();
991 }
992 _pending_jvms = NULL;
993 if (pc_offset > C->debug_info()->last_pc_offset()) {
994 // This is the only way _pending_jvms can become non-NULL:
995 _pending_jvms = nn->jvms();
996 _pending_offset = pc_offset;
997 }
998 }
999 }
1000
1001 // Stay out of the way of real safepoints:
1002 void observe_safepoint(JVMState* jvms, int pc_offset) {
1003 if (_pending_jvms != NULL &&
1004 !_pending_jvms->same_calls_as(jvms) &&
1005 _pending_offset < pc_offset) {
1006 emit_non_safepoint();
1007 }
1008 _pending_jvms = NULL;
1009 }
1010
1011 void flush_at_end() {
1012 if (_pending_jvms != NULL) {
1013 emit_non_safepoint();
1014 }
1015 _pending_jvms = NULL;
1016 }
1017 };
1018
1019 void NonSafepointEmitter::emit_non_safepoint() {
1020 JVMState* youngest_jvms = _pending_jvms;
1021 int pc_offset = _pending_offset;
1022
1023 // Clear it now:
1024 _pending_jvms = NULL;
1025
1026 DebugInformationRecorder* debug_info = C->debug_info();
1027 assert(debug_info->recording_non_safepoints(), "sanity");
1028
1029 debug_info->add_non_safepoint(pc_offset);
1030 int max_depth = youngest_jvms->depth();
1031
1032 // Visit scopes from oldest to youngest.
1033 for (int depth = 1; depth <= max_depth; depth++) {
1034 JVMState* jvms = youngest_jvms->of_depth(depth);
1035 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1036 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1037 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1038 }
1039
1040 // Mark the end of the scope set.
1041 debug_info->end_non_safepoint(pc_offset);
1042 }
1043
1044 //------------------------------init_buffer------------------------------------
1045 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1046
1047 // Set the initially allocated size
1048 int code_req = initial_code_capacity;
1049 int locs_req = initial_locs_capacity;
1050 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1051 int const_req = initial_const_capacity;
1052
1053 int pad_req = NativeCall::instruction_size;
1054 // The extra spacing after the code is necessary on some platforms.
1055 // Sometimes we need to patch in a jump after the last instruction,
1056 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1057
1058 // Compute the byte offset where we can store the deopt pc.
1059 if (fixed_slots() != 0) {
1060 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1061 }
1062
1063 // Compute prolog code size
1064 _method_size = 0;
1065 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1066 #ifdef IA64
1067 if (save_argument_registers()) {
1068 // 4815101: this is a stub with implicit and unknown precision fp args.
1069 // The usual spill mechanism can only generate stfd's in this case, which
1070 // doesn't work if the fp reg to spill contains a single-precision denorm.
1071 // Instead, we hack around the normal spill mechanism using stfspill's and
1072 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1073 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1074 //
1075 // If we ever implement 16-byte 'registers' == stack slots, we can
1076 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1077 // instead of stfd/stfs/ldfd/ldfs.
1078 _frame_slots += 8*(16/BytesPerInt);
1079 }
1080 #endif
1081 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1082
1083 if (has_mach_constant_base_node()) {
1084 // Fill the constant table.
1085 // Note: This must happen before shorten_branches.
1086 for (uint i = 0; i < _cfg->_num_blocks; i++) {
1087 Block* b = _cfg->_blocks[i];
1088
1089 for (uint j = 0; j < b->_nodes.size(); j++) {
1090 Node* n = b->_nodes[j];
1091
1092 // If the node is a MachConstantNode evaluate the constant
1093 // value section.
1094 if (n->is_MachConstant()) {
1095 MachConstantNode* machcon = n->as_MachConstant();
1096 machcon->eval_constant(C);
1097 }
1098 }
1099 }
1100
1101 // Calculate the offsets of the constants and the size of the
1102 // constant table (including the padding to the next section).
1103 constant_table().calculate_offsets_and_size();
1104 const_req = constant_table().size();
1105 }
1106
1107 // Initialize the space for the BufferBlob used to find and verify
1108 // instruction size in MachNode::emit_size()
1109 init_scratch_buffer_blob(const_req);
1110 if (failing()) return NULL; // Out of memory
1111
1112 // Pre-compute the length of blocks and replace
1113 // long branches with short if machine supports it.
1114 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1115
1116 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1117 int exception_handler_req = size_exception_handler();
1118 int deopt_handler_req = size_deopt_handler();
1119 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1120 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1121 stub_req += MAX_stubs_size; // ensure per-stub margin
1122 code_req += MAX_inst_size; // ensure per-instruction margin
1123
1124 if (StressCodeBuffers)
1125 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1126
1127 int total_req =
1128 const_req +
1129 code_req +
1130 pad_req +
1131 stub_req +
1132 exception_handler_req +
1133 deopt_handler_req; // deopt handler
1134
1135 if (has_method_handle_invokes())
1136 total_req += deopt_handler_req; // deopt MH handler
1137
1138 CodeBuffer* cb = code_buffer();
1139 cb->initialize(total_req, locs_req);
1140
1141 // Have we run out of code space?
1142 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1143 C->record_failure("CodeCache is full");
1144 return NULL;
1145 }
1146 // Configure the code buffer.
1147 cb->initialize_consts_size(const_req);
1148 cb->initialize_stubs_size(stub_req);
1149 cb->initialize_oop_recorder(env()->oop_recorder());
1150
1151 // fill in the nop array for bundling computations
1152 MachNode *_nop_list[Bundle::_nop_count];
1153 Bundle::initialize_nops(_nop_list, this);
1154
1155 return cb;
1156 }
1157
1158 //------------------------------fill_buffer------------------------------------
1159 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1160 // blk_starts[] contains offsets calculated during short branches processing,
1161 // offsets should not be increased during following steps.
1162
1163 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1164 // of a loop. It is used to determine the padding for loop alignment.
1165 compute_loop_first_inst_sizes();
1166
1167 // Create oopmap set.
1168 _oop_map_set = new OopMapSet();
1169
1170 // !!!!! This preserves old handling of oopmaps for now
1171 debug_info()->set_oopmaps(_oop_map_set);
1172
1173 uint nblocks = _cfg->_num_blocks;
1174 // Count and start of implicit null check instructions
1175 uint inct_cnt = 0;
1176 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1177
1178 // Count and start of calls
1179 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1180
1181 uint return_offset = 0;
1182 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1183
1184 int previous_offset = 0;
1185 int current_offset = 0;
1186 int last_call_offset = -1;
1187 int last_avoid_back_to_back_offset = -1;
1188 #ifdef ASSERT
1189 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1190 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1191 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1192 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1193 #endif
1194
1195 // Create an array of unused labels, one for each basic block, if printing is enabled
1196 #ifndef PRODUCT
1197 int *node_offsets = NULL;
1198 uint node_offset_limit = unique();
1199
1200 if (print_assembly())
1201 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1202 #endif
1203
1204 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1205
1206 // Emit the constant table.
1207 if (has_mach_constant_base_node()) {
1208 constant_table().emit(*cb);
1209 }
1210
1211 // Create an array of labels, one for each basic block
1212 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1213 for (uint i=0; i <= nblocks; i++) {
1214 blk_labels[i].init();
1215 }
1216
1217 // ------------------
1218 // Now fill in the code buffer
1219 Node *delay_slot = NULL;
1220
1221 for (uint i=0; i < nblocks; i++) {
1222 Block *b = _cfg->_blocks[i];
1223
1224 Node *head = b->head();
1225
1226 // If this block needs to start aligned (i.e, can be reached other
1227 // than by falling-thru from the previous block), then force the
1228 // start of a new bundle.
1229 if (Pipeline::requires_bundling() && starts_bundle(head))
1230 cb->flush_bundle(true);
1231
1232 #ifdef ASSERT
1233 if (!b->is_connector()) {
1234 stringStream st;
1235 b->dump_head(_cfg, &st);
1236 MacroAssembler(cb).block_comment(st.as_string());
1237 }
1238 jmp_target[i] = 0;
1239 jmp_offset[i] = 0;
1240 jmp_size[i] = 0;
1241 jmp_rule[i] = 0;
1242 #endif
1243 int blk_offset = current_offset;
1244
1245 // Define the label at the beginning of the basic block
1246 MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
1247
1248 uint last_inst = b->_nodes.size();
1249
1250 // Emit block normally, except for last instruction.
1251 // Emit means "dump code bits into code buffer".
1252 for (uint j = 0; j<last_inst; j++) {
1253
1254 // Get the node
1255 Node* n = b->_nodes[j];
1256
1257 // See if delay slots are supported
1258 if (valid_bundle_info(n) &&
1259 node_bundling(n)->used_in_unconditional_delay()) {
1260 assert(delay_slot == NULL, "no use of delay slot node");
1261 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1262
1263 delay_slot = n;
1264 continue;
1265 }
1266
1267 // If this starts a new instruction group, then flush the current one
1268 // (but allow split bundles)
1269 if (Pipeline::requires_bundling() && starts_bundle(n))
1270 cb->flush_bundle(false);
1271
1272 // The following logic is duplicated in the code ifdeffed for
1273 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1274 // should be factored out. Or maybe dispersed to the nodes?
1275
1276 // Special handling for SafePoint/Call Nodes
1277 bool is_mcall = false;
1278 if (n->is_Mach()) {
1279 MachNode *mach = n->as_Mach();
1280 is_mcall = n->is_MachCall();
1281 bool is_sfn = n->is_MachSafePoint();
1282
1283 // If this requires all previous instructions be flushed, then do so
1284 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1285 cb->flush_bundle(true);
1286 current_offset = cb->insts_size();
1287 }
1288
1289 // A padding may be needed again since a previous instruction
1290 // could be moved to delay slot.
1291
1292 // align the instruction if necessary
1293 int padding = mach->compute_padding(current_offset);
1294 // Make sure safepoint node for polling is distinct from a call's
1295 // return by adding a nop if needed.
1296 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1297 padding = nop_size;
1298 }
1299 if (padding == 0 && mach->avoid_back_to_back() &&
1300 current_offset == last_avoid_back_to_back_offset) {
1301 // Avoid back to back some instructions.
1302 padding = nop_size;
1303 }
1304
1305 if(padding > 0) {
1306 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1307 int nops_cnt = padding / nop_size;
1308 MachNode *nop = new (this) MachNopNode(nops_cnt);
1309 b->_nodes.insert(j++, nop);
1310 last_inst++;
1311 _cfg->map_node_to_block(nop, b);
1312 nop->emit(*cb, _regalloc);
1313 cb->flush_bundle(true);
1314 current_offset = cb->insts_size();
1315 }
1316
1317 // Remember the start of the last call in a basic block
1318 if (is_mcall) {
1319 MachCallNode *mcall = mach->as_MachCall();
1320
1321 // This destination address is NOT PC-relative
1322 mcall->method_set((intptr_t)mcall->entry_point());
1323
1324 // Save the return address
1325 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset();
1326
1327 if (mcall->is_MachCallLeaf()) {
1328 is_mcall = false;
1329 is_sfn = false;
1330 }
1331 }
1332
1333 // sfn will be valid whenever mcall is valid now because of inheritance
1334 if (is_sfn || is_mcall) {
1335
1336 // Handle special safepoint nodes for synchronization
1337 if (!is_mcall) {
1338 MachSafePointNode *sfn = mach->as_MachSafePoint();
1339 // !!!!! Stubs only need an oopmap right now, so bail out
1340 if (sfn->jvms()->method() == NULL) {
1341 // Write the oopmap directly to the code blob??!!
1342 # ifdef ENABLE_ZAP_DEAD_LOCALS
1343 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1344 # endif
1345 continue;
1346 }
1347 } // End synchronization
1348
1349 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1350 current_offset);
1351 Process_OopMap_Node(mach, current_offset);
1352 } // End if safepoint
1353
1354 // If this is a null check, then add the start of the previous instruction to the list
1355 else if( mach->is_MachNullCheck() ) {
1356 inct_starts[inct_cnt++] = previous_offset;
1357 }
1358
1359 // If this is a branch, then fill in the label with the target BB's label
1360 else if (mach->is_MachBranch()) {
1361 // This requires the TRUE branch target be in succs[0]
1362 uint block_num = b->non_connector_successor(0)->_pre_order;
1363
1364 // Try to replace long branch if delay slot is not used,
1365 // it is mostly for back branches since forward branch's
1366 // distance is not updated yet.
1367 bool delay_slot_is_used = valid_bundle_info(n) &&
1368 node_bundling(n)->use_unconditional_delay();
1369 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1370 assert(delay_slot == NULL, "not expecting delay slot node");
1371 int br_size = n->size(_regalloc);
1372 int offset = blk_starts[block_num] - current_offset;
1373 if (block_num >= i) {
1374 // Current and following block's offset are not
1375 // finilized yet, adjust distance by the difference
1376 // between calculated and final offsets of current block.
1377 offset -= (blk_starts[i] - blk_offset);
1378 }
1379 // In the following code a nop could be inserted before
1380 // the branch which will increase the backward distance.
1381 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1382 if (needs_padding && offset <= 0)
1383 offset -= nop_size;
1384
1385 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1386 // We've got a winner. Replace this branch.
1387 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1388
1389 // Update the jmp_size.
1390 int new_size = replacement->size(_regalloc);
1391 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1392 // Insert padding between avoid_back_to_back branches.
1393 if (needs_padding && replacement->avoid_back_to_back()) {
1394 MachNode *nop = new (this) MachNopNode();
1395 b->_nodes.insert(j++, nop);
1396 _cfg->map_node_to_block(nop, b);
1397 last_inst++;
1398 nop->emit(*cb, _regalloc);
1399 cb->flush_bundle(true);
1400 current_offset = cb->insts_size();
1401 }
1402 #ifdef ASSERT
1403 jmp_target[i] = block_num;
1404 jmp_offset[i] = current_offset - blk_offset;
1405 jmp_size[i] = new_size;
1406 jmp_rule[i] = mach->rule();
1407 #endif
1408 b->_nodes.map(j, replacement);
1409 mach->subsume_by(replacement, C);
1410 n = replacement;
1411 mach = replacement;
1412 }
1413 }
1414 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1415 } else if (mach->ideal_Opcode() == Op_Jump) {
1416 for (uint h = 0; h < b->_num_succs; h++) {
1417 Block* succs_block = b->_succs[h];
1418 for (uint j = 1; j < succs_block->num_preds(); j++) {
1419 Node* jpn = succs_block->pred(j);
1420 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1421 uint block_num = succs_block->non_connector()->_pre_order;
1422 Label *blkLabel = &blk_labels[block_num];
1423 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1424 }
1425 }
1426 }
1427 }
1428
1429 #ifdef ASSERT
1430 // Check that oop-store precedes the card-mark
1431 else if (mach->ideal_Opcode() == Op_StoreCM) {
1432 uint storeCM_idx = j;
1433 int count = 0;
1434 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1435 Node *oop_store = mach->in(prec); // Precedence edge
1436 if (oop_store == NULL) continue;
1437 count++;
1438 uint i4;
1439 for( i4 = 0; i4 < last_inst; ++i4 ) {
1440 if( b->_nodes[i4] == oop_store ) break;
1441 }
1442 // Note: This test can provide a false failure if other precedence
1443 // edges have been added to the storeCMNode.
1444 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1445 }
1446 assert(count > 0, "storeCM expects at least one precedence edge");
1447 }
1448 #endif
1449
1450 else if (!n->is_Proj()) {
1451 // Remember the beginning of the previous instruction, in case
1452 // it's followed by a flag-kill and a null-check. Happens on
1453 // Intel all the time, with add-to-memory kind of opcodes.
1454 previous_offset = current_offset;
1455 }
1456 }
1457
1458 // Verify that there is sufficient space remaining
1459 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1460 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1461 C->record_failure("CodeCache is full");
1462 return;
1463 }
1464
1465 // Save the offset for the listing
1466 #ifndef PRODUCT
1467 if (node_offsets && n->_idx < node_offset_limit)
1468 node_offsets[n->_idx] = cb->insts_size();
1469 #endif
1470
1471 // "Normal" instruction case
1472 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1473 n->emit(*cb, _regalloc);
1474 current_offset = cb->insts_size();
1475
1476 #ifdef ASSERT
1477 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1478 n->dump();
1479 assert(false, "wrong size of mach node");
1480 }
1481 #endif
1482 non_safepoints.observe_instruction(n, current_offset);
1483
1484 // mcall is last "call" that can be a safepoint
1485 // record it so we can see if a poll will directly follow it
1486 // in which case we'll need a pad to make the PcDesc sites unique
1487 // see 5010568. This can be slightly inaccurate but conservative
1488 // in the case that return address is not actually at current_offset.
1489 // This is a small price to pay.
1490
1491 if (is_mcall) {
1492 last_call_offset = current_offset;
1493 }
1494
1495 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
1496 // Avoid back to back some instructions.
1497 last_avoid_back_to_back_offset = current_offset;
1498 }
1499
1500 // See if this instruction has a delay slot
1501 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1502 assert(delay_slot != NULL, "expecting delay slot node");
1503
1504 // Back up 1 instruction
1505 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1506
1507 // Save the offset for the listing
1508 #ifndef PRODUCT
1509 if (node_offsets && delay_slot->_idx < node_offset_limit)
1510 node_offsets[delay_slot->_idx] = cb->insts_size();
1511 #endif
1512
1513 // Support a SafePoint in the delay slot
1514 if (delay_slot->is_MachSafePoint()) {
1515 MachNode *mach = delay_slot->as_Mach();
1516 // !!!!! Stubs only need an oopmap right now, so bail out
1517 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1518 // Write the oopmap directly to the code blob??!!
1519 # ifdef ENABLE_ZAP_DEAD_LOCALS
1520 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1521 # endif
1522 delay_slot = NULL;
1523 continue;
1524 }
1525
1526 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1527 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1528 adjusted_offset);
1529 // Generate an OopMap entry
1530 Process_OopMap_Node(mach, adjusted_offset);
1531 }
1532
1533 // Insert the delay slot instruction
1534 delay_slot->emit(*cb, _regalloc);
1535
1536 // Don't reuse it
1537 delay_slot = NULL;
1538 }
1539
1540 } // End for all instructions in block
1541
1542 // If the next block is the top of a loop, pad this block out to align
1543 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1544 if (i < nblocks-1) {
1545 Block *nb = _cfg->_blocks[i+1];
1546 int padding = nb->alignment_padding(current_offset);
1547 if( padding > 0 ) {
1548 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1549 b->_nodes.insert( b->_nodes.size(), nop );
1550 _cfg->map_node_to_block(nop, b);
1551 nop->emit(*cb, _regalloc);
1552 current_offset = cb->insts_size();
1553 }
1554 }
1555 // Verify that the distance for generated before forward
1556 // short branches is still valid.
1557 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1558
1559 // Save new block start offset
1560 blk_starts[i] = blk_offset;
1561 } // End of for all blocks
1562 blk_starts[nblocks] = current_offset;
1563
1564 non_safepoints.flush_at_end();
1565
1566 // Offset too large?
1567 if (failing()) return;
1568
1569 // Define a pseudo-label at the end of the code
1570 MacroAssembler(cb).bind( blk_labels[nblocks] );
1571
1572 // Compute the size of the first block
1573 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1574
1575 assert(cb->insts_size() < 500000, "method is unreasonably large");
1576
1577 #ifdef ASSERT
1578 for (uint i = 0; i < nblocks; i++) { // For all blocks
1579 if (jmp_target[i] != 0) {
1580 int br_size = jmp_size[i];
1581 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1582 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1583 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]);
1584 assert(false, "Displacement too large for short jmp");
1585 }
1586 }
1587 }
1588 #endif
1589
1590 // ------------------
1591
1592 #ifndef PRODUCT
1593 // Information on the size of the method, without the extraneous code
1594 Scheduling::increment_method_size(cb->insts_size());
1595 #endif
1596
1597 // ------------------
1598 // Fill in exception table entries.
1599 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1600
1601 // Only java methods have exception handlers and deopt handlers
1602 if (_method) {
1603 // Emit the exception handler code.
1604 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1605 // Emit the deopt handler code.
1606 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1607
1608 // Emit the MethodHandle deopt handler code (if required).
1609 if (has_method_handle_invokes()) {
1610 // We can use the same code as for the normal deopt handler, we
1611 // just need a different entry point address.
1612 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1613 }
1614 }
1615
1616 // One last check for failed CodeBuffer::expand:
1617 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1618 C->record_failure("CodeCache is full");
1619 return;
1620 }
1621
1622 #ifndef PRODUCT
1623 // Dump the assembly code, including basic-block numbers
1624 if (print_assembly()) {
1625 ttyLocker ttyl; // keep the following output all in one block
1626 if (!VMThread::should_terminate()) { // test this under the tty lock
1627 // This output goes directly to the tty, not the compiler log.
1628 // To enable tools to match it up with the compilation activity,
1629 // be sure to tag this tty output with the compile ID.
1630 if (xtty != NULL) {
1631 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1632 is_osr_compilation() ? " compile_kind='osr'" :
1633 "");
1634 }
1635 if (method() != NULL) {
1636 method()->print_metadata();
1637 }
1638 dump_asm(node_offsets, node_offset_limit);
1639 if (xtty != NULL) {
1640 xtty->tail("opto_assembly");
1641 }
1642 }
1643 }
1644 #endif
1645
1646 }
1647
1648 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1649 _inc_table.set_size(cnt);
1650
1651 uint inct_cnt = 0;
1652 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1653 Block *b = _cfg->_blocks[i];
1654 Node *n = NULL;
1655 int j;
1656
1657 // Find the branch; ignore trailing NOPs.
1658 for( j = b->_nodes.size()-1; j>=0; j-- ) {
1659 n = b->_nodes[j];
1660 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con )
1661 break;
1662 }
1663
1664 // If we didn't find anything, continue
1665 if( j < 0 ) continue;
1666
1667 // Compute ExceptionHandlerTable subtable entry and add it
1668 // (skip empty blocks)
1669 if( n->is_Catch() ) {
1670
1671 // Get the offset of the return from the call
1672 uint call_return = call_returns[b->_pre_order];
1673 #ifdef ASSERT
1674 assert( call_return > 0, "no call seen for this basic block" );
1675 while( b->_nodes[--j]->is_MachProj() ) ;
1676 assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" );
1677 #endif
1678 // last instruction is a CatchNode, find it's CatchProjNodes
1679 int nof_succs = b->_num_succs;
1680 // allocate space
1681 GrowableArray<intptr_t> handler_bcis(nof_succs);
1682 GrowableArray<intptr_t> handler_pcos(nof_succs);
1683 // iterate through all successors
1684 for (int j = 0; j < nof_succs; j++) {
1685 Block* s = b->_succs[j];
1686 bool found_p = false;
1687 for( uint k = 1; k < s->num_preds(); k++ ) {
1688 Node *pk = s->pred(k);
1689 if( pk->is_CatchProj() && pk->in(0) == n ) {
1690 const CatchProjNode* p = pk->as_CatchProj();
1691 found_p = true;
1692 // add the corresponding handler bci & pco information
1693 if( p->_con != CatchProjNode::fall_through_index ) {
1694 // p leads to an exception handler (and is not fall through)
1695 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering");
1696 // no duplicates, please
1697 if( !handler_bcis.contains(p->handler_bci()) ) {
1698 uint block_num = s->non_connector()->_pre_order;
1699 handler_bcis.append(p->handler_bci());
1700 handler_pcos.append(blk_labels[block_num].loc_pos());
1701 }
1702 }
1703 }
1704 }
1705 assert(found_p, "no matching predecessor found");
1706 // Note: Due to empty block removal, one block may have
1707 // several CatchProj inputs, from the same Catch.
1708 }
1709
1710 // Set the offset of the return from the call
1711 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1712 continue;
1713 }
1714
1715 // Handle implicit null exception table updates
1716 if( n->is_MachNullCheck() ) {
1717 uint block_num = b->non_connector_successor(0)->_pre_order;
1718 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() );
1719 continue;
1720 }
1721 } // End of for all blocks fill in exception table entries
1722 }
1723
1724 // Static Variables
1725 #ifndef PRODUCT
1726 uint Scheduling::_total_nop_size = 0;
1727 uint Scheduling::_total_method_size = 0;
1728 uint Scheduling::_total_branches = 0;
1729 uint Scheduling::_total_unconditional_delays = 0;
1730 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1731 #endif
1732
1733 // Initializer for class Scheduling
1734
1735 Scheduling::Scheduling(Arena *arena, Compile &compile)
1736 : _arena(arena),
1737 _cfg(compile.cfg()),
1738 _regalloc(compile.regalloc()),
1739 _reg_node(arena),
1740 _bundle_instr_count(0),
1741 _bundle_cycle_number(0),
1742 _scheduled(arena),
1743 _available(arena),
1744 _next_node(NULL),
1745 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1746 _pinch_free_list(arena)
1747 #ifndef PRODUCT
1748 , _branches(0)
1749 , _unconditional_delays(0)
1750 #endif
1751 {
1752 // Create a MachNopNode
1753 _nop = new (&compile) MachNopNode();
1754
1755 // Now that the nops are in the array, save the count
1756 // (but allow entries for the nops)
1757 _node_bundling_limit = compile.unique();
1758 uint node_max = _regalloc->node_regs_max_index();
1759
1760 compile.set_node_bundling_limit(_node_bundling_limit);
1761
1762 // This one is persistent within the Compile class
1763 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1764
1765 // Allocate space for fixed-size arrays
1766 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1767 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1768 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1769
1770 // Clear the arrays
1771 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1772 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1773 memset(_uses, 0, node_max * sizeof(short));
1774 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1775
1776 // Clear the bundling information
1777 memcpy(_bundle_use_elements,
1778 Pipeline_Use::elaborated_elements,
1779 sizeof(Pipeline_Use::elaborated_elements));
1780
1781 // Get the last node
1782 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1];
1783
1784 _next_node = bb->_nodes[bb->_nodes.size()-1];
1785 }
1786
1787 #ifndef PRODUCT
1788 // Scheduling destructor
1789 Scheduling::~Scheduling() {
1790 _total_branches += _branches;
1791 _total_unconditional_delays += _unconditional_delays;
1792 }
1793 #endif
1794
1795 // Step ahead "i" cycles
1796 void Scheduling::step(uint i) {
1797
1798 Bundle *bundle = node_bundling(_next_node);
1799 bundle->set_starts_bundle();
1800
1801 // Update the bundle record, but leave the flags information alone
1802 if (_bundle_instr_count > 0) {
1803 bundle->set_instr_count(_bundle_instr_count);
1804 bundle->set_resources_used(_bundle_use.resourcesUsed());
1805 }
1806
1807 // Update the state information
1808 _bundle_instr_count = 0;
1809 _bundle_cycle_number += i;
1810 _bundle_use.step(i);
1811 }
1812
1813 void Scheduling::step_and_clear() {
1814 Bundle *bundle = node_bundling(_next_node);
1815 bundle->set_starts_bundle();
1816
1817 // Update the bundle record
1818 if (_bundle_instr_count > 0) {
1819 bundle->set_instr_count(_bundle_instr_count);
1820 bundle->set_resources_used(_bundle_use.resourcesUsed());
1821
1822 _bundle_cycle_number += 1;
1823 }
1824
1825 // Clear the bundling information
1826 _bundle_instr_count = 0;
1827 _bundle_use.reset();
1828
1829 memcpy(_bundle_use_elements,
1830 Pipeline_Use::elaborated_elements,
1831 sizeof(Pipeline_Use::elaborated_elements));
1832 }
1833
1834 //------------------------------ScheduleAndBundle------------------------------
1835 // Perform instruction scheduling and bundling over the sequence of
1836 // instructions in backwards order.
1837 void Compile::ScheduleAndBundle() {
1838
1839 // Don't optimize this if it isn't a method
1840 if (!_method)
1841 return;
1842
1843 // Don't optimize this if scheduling is disabled
1844 if (!do_scheduling())
1845 return;
1846
1847 // Scheduling code works only with pairs (8 bytes) maximum.
1848 if (max_vector_size() > 8)
1849 return;
1850
1851 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1852
1853 // Create a data structure for all the scheduling information
1854 Scheduling scheduling(Thread::current()->resource_area(), *this);
1855
1856 // Walk backwards over each basic block, computing the needed alignment
1857 // Walk over all the basic blocks
1858 scheduling.DoScheduling();
1859 }
1860
1861 //------------------------------ComputeLocalLatenciesForward-------------------
1862 // Compute the latency of all the instructions. This is fairly simple,
1863 // because we already have a legal ordering. Walk over the instructions
1864 // from first to last, and compute the latency of the instruction based
1865 // on the latency of the preceding instruction(s).
1866 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1867 #ifndef PRODUCT
1868 if (_cfg->C->trace_opto_output())
1869 tty->print("# -> ComputeLocalLatenciesForward\n");
1870 #endif
1871
1872 // Walk over all the schedulable instructions
1873 for( uint j=_bb_start; j < _bb_end; j++ ) {
1874
1875 // This is a kludge, forcing all latency calculations to start at 1.
1876 // Used to allow latency 0 to force an instruction to the beginning
1877 // of the bb
1878 uint latency = 1;
1879 Node *use = bb->_nodes[j];
1880 uint nlen = use->len();
1881
1882 // Walk over all the inputs
1883 for ( uint k=0; k < nlen; k++ ) {
1884 Node *def = use->in(k);
1885 if (!def)
1886 continue;
1887
1888 uint l = _node_latency[def->_idx] + use->latency(k);
1889 if (latency < l)
1890 latency = l;
1891 }
1892
1893 _node_latency[use->_idx] = latency;
1894
1895 #ifndef PRODUCT
1896 if (_cfg->C->trace_opto_output()) {
1897 tty->print("# latency %4d: ", latency);
1898 use->dump();
1899 }
1900 #endif
1901 }
1902
1903 #ifndef PRODUCT
1904 if (_cfg->C->trace_opto_output())
1905 tty->print("# <- ComputeLocalLatenciesForward\n");
1906 #endif
1907
1908 } // end ComputeLocalLatenciesForward
1909
1910 // See if this node fits into the present instruction bundle
1911 bool Scheduling::NodeFitsInBundle(Node *n) {
1912 uint n_idx = n->_idx;
1913
1914 // If this is the unconditional delay instruction, then it fits
1915 if (n == _unconditional_delay_slot) {
1916 #ifndef PRODUCT
1917 if (_cfg->C->trace_opto_output())
1918 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1919 #endif
1920 return (true);
1921 }
1922
1923 // If the node cannot be scheduled this cycle, skip it
1924 if (_current_latency[n_idx] > _bundle_cycle_number) {
1925 #ifndef PRODUCT
1926 if (_cfg->C->trace_opto_output())
1927 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1928 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1929 #endif
1930 return (false);
1931 }
1932
1933 const Pipeline *node_pipeline = n->pipeline();
1934
1935 uint instruction_count = node_pipeline->instructionCount();
1936 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1937 instruction_count = 0;
1938 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1939 instruction_count++;
1940
1941 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1942 #ifndef PRODUCT
1943 if (_cfg->C->trace_opto_output())
1944 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1945 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1946 #endif
1947 return (false);
1948 }
1949
1950 // Don't allow non-machine nodes to be handled this way
1951 if (!n->is_Mach() && instruction_count == 0)
1952 return (false);
1953
1954 // See if there is any overlap
1955 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1956
1957 if (delay > 0) {
1958 #ifndef PRODUCT
1959 if (_cfg->C->trace_opto_output())
1960 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1961 #endif
1962 return false;
1963 }
1964
1965 #ifndef PRODUCT
1966 if (_cfg->C->trace_opto_output())
1967 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1968 #endif
1969
1970 return true;
1971 }
1972
1973 Node * Scheduling::ChooseNodeToBundle() {
1974 uint siz = _available.size();
1975
1976 if (siz == 0) {
1977
1978 #ifndef PRODUCT
1979 if (_cfg->C->trace_opto_output())
1980 tty->print("# ChooseNodeToBundle: NULL\n");
1981 #endif
1982 return (NULL);
1983 }
1984
1985 // Fast path, if only 1 instruction in the bundle
1986 if (siz == 1) {
1987 #ifndef PRODUCT
1988 if (_cfg->C->trace_opto_output()) {
1989 tty->print("# ChooseNodeToBundle (only 1): ");
1990 _available[0]->dump();
1991 }
1992 #endif
1993 return (_available[0]);
1994 }
1995
1996 // Don't bother, if the bundle is already full
1997 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1998 for ( uint i = 0; i < siz; i++ ) {
1999 Node *n = _available[i];
2000
2001 // Skip projections, we'll handle them another way
2002 if (n->is_Proj())
2003 continue;
2004
2005 // This presupposed that instructions are inserted into the
2006 // available list in a legality order; i.e. instructions that
2007 // must be inserted first are at the head of the list
2008 if (NodeFitsInBundle(n)) {
2009 #ifndef PRODUCT
2010 if (_cfg->C->trace_opto_output()) {
2011 tty->print("# ChooseNodeToBundle: ");
2012 n->dump();
2013 }
2014 #endif
2015 return (n);
2016 }
2017 }
2018 }
2019
2020 // Nothing fits in this bundle, choose the highest priority
2021 #ifndef PRODUCT
2022 if (_cfg->C->trace_opto_output()) {
2023 tty->print("# ChooseNodeToBundle: ");
2024 _available[0]->dump();
2025 }
2026 #endif
2027
2028 return _available[0];
2029 }
2030
2031 //------------------------------AddNodeToAvailableList-------------------------
2032 void Scheduling::AddNodeToAvailableList(Node *n) {
2033 assert( !n->is_Proj(), "projections never directly made available" );
2034 #ifndef PRODUCT
2035 if (_cfg->C->trace_opto_output()) {
2036 tty->print("# AddNodeToAvailableList: ");
2037 n->dump();
2038 }
2039 #endif
2040
2041 int latency = _current_latency[n->_idx];
2042
2043 // Insert in latency order (insertion sort)
2044 uint i;
2045 for ( i=0; i < _available.size(); i++ )
2046 if (_current_latency[_available[i]->_idx] > latency)
2047 break;
2048
2049 // Special Check for compares following branches
2050 if( n->is_Mach() && _scheduled.size() > 0 ) {
2051 int op = n->as_Mach()->ideal_Opcode();
2052 Node *last = _scheduled[0];
2053 if( last->is_MachIf() && last->in(1) == n &&
2054 ( op == Op_CmpI ||
2055 op == Op_CmpU ||
2056 op == Op_CmpP ||
2057 op == Op_CmpF ||
2058 op == Op_CmpD ||
2059 op == Op_CmpL ) ) {
2060
2061 // Recalculate position, moving to front of same latency
2062 for ( i=0 ; i < _available.size(); i++ )
2063 if (_current_latency[_available[i]->_idx] >= latency)
2064 break;
2065 }
2066 }
2067
2068 // Insert the node in the available list
2069 _available.insert(i, n);
2070
2071 #ifndef PRODUCT
2072 if (_cfg->C->trace_opto_output())
2073 dump_available();
2074 #endif
2075 }
2076
2077 //------------------------------DecrementUseCounts-----------------------------
2078 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2079 for ( uint i=0; i < n->len(); i++ ) {
2080 Node *def = n->in(i);
2081 if (!def) continue;
2082 if( def->is_Proj() ) // If this is a machine projection, then
2083 def = def->in(0); // propagate usage thru to the base instruction
2084
2085 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2086 continue;
2087 }
2088
2089 // Compute the latency
2090 uint l = _bundle_cycle_number + n->latency(i);
2091 if (_current_latency[def->_idx] < l)
2092 _current_latency[def->_idx] = l;
2093
2094 // If this does not have uses then schedule it
2095 if ((--_uses[def->_idx]) == 0)
2096 AddNodeToAvailableList(def);
2097 }
2098 }
2099
2100 //------------------------------AddNodeToBundle--------------------------------
2101 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2102 #ifndef PRODUCT
2103 if (_cfg->C->trace_opto_output()) {
2104 tty->print("# AddNodeToBundle: ");
2105 n->dump();
2106 }
2107 #endif
2108
2109 // Remove this from the available list
2110 uint i;
2111 for (i = 0; i < _available.size(); i++)
2112 if (_available[i] == n)
2113 break;
2114 assert(i < _available.size(), "entry in _available list not found");
2115 _available.remove(i);
2116
2117 // See if this fits in the current bundle
2118 const Pipeline *node_pipeline = n->pipeline();
2119 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2120
2121 // Check for instructions to be placed in the delay slot. We
2122 // do this before we actually schedule the current instruction,
2123 // because the delay slot follows the current instruction.
2124 if (Pipeline::_branch_has_delay_slot &&
2125 node_pipeline->hasBranchDelay() &&
2126 !_unconditional_delay_slot) {
2127
2128 uint siz = _available.size();
2129
2130 // Conditional branches can support an instruction that
2131 // is unconditionally executed and not dependent by the
2132 // branch, OR a conditionally executed instruction if
2133 // the branch is taken. In practice, this means that
2134 // the first instruction at the branch target is
2135 // copied to the delay slot, and the branch goes to
2136 // the instruction after that at the branch target
2137 if ( n->is_MachBranch() ) {
2138
2139 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2140 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2141
2142 #ifndef PRODUCT
2143 _branches++;
2144 #endif
2145
2146 // At least 1 instruction is on the available list
2147 // that is not dependent on the branch
2148 for (uint i = 0; i < siz; i++) {
2149 Node *d = _available[i];
2150 const Pipeline *avail_pipeline = d->pipeline();
2151
2152 // Don't allow safepoints in the branch shadow, that will
2153 // cause a number of difficulties
2154 if ( avail_pipeline->instructionCount() == 1 &&
2155 !avail_pipeline->hasMultipleBundles() &&
2156 !avail_pipeline->hasBranchDelay() &&
2157 Pipeline::instr_has_unit_size() &&
2158 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2159 NodeFitsInBundle(d) &&
2160 !node_bundling(d)->used_in_delay()) {
2161
2162 if (d->is_Mach() && !d->is_MachSafePoint()) {
2163 // A node that fits in the delay slot was found, so we need to
2164 // set the appropriate bits in the bundle pipeline information so
2165 // that it correctly indicates resource usage. Later, when we
2166 // attempt to add this instruction to the bundle, we will skip
2167 // setting the resource usage.
2168 _unconditional_delay_slot = d;
2169 node_bundling(n)->set_use_unconditional_delay();
2170 node_bundling(d)->set_used_in_unconditional_delay();
2171 _bundle_use.add_usage(avail_pipeline->resourceUse());
2172 _current_latency[d->_idx] = _bundle_cycle_number;
2173 _next_node = d;
2174 ++_bundle_instr_count;
2175 #ifndef PRODUCT
2176 _unconditional_delays++;
2177 #endif
2178 break;
2179 }
2180 }
2181 }
2182 }
2183
2184 // No delay slot, add a nop to the usage
2185 if (!_unconditional_delay_slot) {
2186 // See if adding an instruction in the delay slot will overflow
2187 // the bundle.
2188 if (!NodeFitsInBundle(_nop)) {
2189 #ifndef PRODUCT
2190 if (_cfg->C->trace_opto_output())
2191 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2192 #endif
2193 step(1);
2194 }
2195
2196 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2197 _next_node = _nop;
2198 ++_bundle_instr_count;
2199 }
2200
2201 // See if the instruction in the delay slot requires a
2202 // step of the bundles
2203 if (!NodeFitsInBundle(n)) {
2204 #ifndef PRODUCT
2205 if (_cfg->C->trace_opto_output())
2206 tty->print("# *** STEP(branch won't fit) ***\n");
2207 #endif
2208 // Update the state information
2209 _bundle_instr_count = 0;
2210 _bundle_cycle_number += 1;
2211 _bundle_use.step(1);
2212 }
2213 }
2214
2215 // Get the number of instructions
2216 uint instruction_count = node_pipeline->instructionCount();
2217 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2218 instruction_count = 0;
2219
2220 // Compute the latency information
2221 uint delay = 0;
2222
2223 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2224 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2225 if (relative_latency < 0)
2226 relative_latency = 0;
2227
2228 delay = _bundle_use.full_latency(relative_latency, node_usage);
2229
2230 // Does not fit in this bundle, start a new one
2231 if (delay > 0) {
2232 step(delay);
2233
2234 #ifndef PRODUCT
2235 if (_cfg->C->trace_opto_output())
2236 tty->print("# *** STEP(%d) ***\n", delay);
2237 #endif
2238 }
2239 }
2240
2241 // If this was placed in the delay slot, ignore it
2242 if (n != _unconditional_delay_slot) {
2243
2244 if (delay == 0) {
2245 if (node_pipeline->hasMultipleBundles()) {
2246 #ifndef PRODUCT
2247 if (_cfg->C->trace_opto_output())
2248 tty->print("# *** STEP(multiple instructions) ***\n");
2249 #endif
2250 step(1);
2251 }
2252
2253 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2254 #ifndef PRODUCT
2255 if (_cfg->C->trace_opto_output())
2256 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2257 instruction_count + _bundle_instr_count,
2258 Pipeline::_max_instrs_per_cycle);
2259 #endif
2260 step(1);
2261 }
2262 }
2263
2264 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2265 _bundle_instr_count++;
2266
2267 // Set the node's latency
2268 _current_latency[n->_idx] = _bundle_cycle_number;
2269
2270 // Now merge the functional unit information
2271 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2272 _bundle_use.add_usage(node_usage);
2273
2274 // Increment the number of instructions in this bundle
2275 _bundle_instr_count += instruction_count;
2276
2277 // Remember this node for later
2278 if (n->is_Mach())
2279 _next_node = n;
2280 }
2281
2282 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2283 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2284 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2285 // into the block. All other scheduled nodes get put in the schedule here.
2286 int op = n->Opcode();
2287 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2288 (op != Op_Node && // Not an unused antidepedence node and
2289 // not an unallocated boxlock
2290 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2291
2292 // Push any trailing projections
2293 if( bb->_nodes[bb->_nodes.size()-1] != n ) {
2294 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2295 Node *foi = n->fast_out(i);
2296 if( foi->is_Proj() )
2297 _scheduled.push(foi);
2298 }
2299 }
2300
2301 // Put the instruction in the schedule list
2302 _scheduled.push(n);
2303 }
2304
2305 #ifndef PRODUCT
2306 if (_cfg->C->trace_opto_output())
2307 dump_available();
2308 #endif
2309
2310 // Walk all the definitions, decrementing use counts, and
2311 // if a definition has a 0 use count, place it in the available list.
2312 DecrementUseCounts(n,bb);
2313 }
2314
2315 //------------------------------ComputeUseCount--------------------------------
2316 // This method sets the use count within a basic block. We will ignore all
2317 // uses outside the current basic block. As we are doing a backwards walk,
2318 // any node we reach that has a use count of 0 may be scheduled. This also
2319 // avoids the problem of cyclic references from phi nodes, as long as phi
2320 // nodes are at the front of the basic block. This method also initializes
2321 // the available list to the set of instructions that have no uses within this
2322 // basic block.
2323 void Scheduling::ComputeUseCount(const Block *bb) {
2324 #ifndef PRODUCT
2325 if (_cfg->C->trace_opto_output())
2326 tty->print("# -> ComputeUseCount\n");
2327 #endif
2328
2329 // Clear the list of available and scheduled instructions, just in case
2330 _available.clear();
2331 _scheduled.clear();
2332
2333 // No delay slot specified
2334 _unconditional_delay_slot = NULL;
2335
2336 #ifdef ASSERT
2337 for( uint i=0; i < bb->_nodes.size(); i++ )
2338 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );
2339 #endif
2340
2341 // Force the _uses count to never go to zero for unscheduable pieces
2342 // of the block
2343 for( uint k = 0; k < _bb_start; k++ )
2344 _uses[bb->_nodes[k]->_idx] = 1;
2345 for( uint l = _bb_end; l < bb->_nodes.size(); l++ )
2346 _uses[bb->_nodes[l]->_idx] = 1;
2347
2348 // Iterate backwards over the instructions in the block. Don't count the
2349 // branch projections at end or the block header instructions.
2350 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2351 Node *n = bb->_nodes[j];
2352 if( n->is_Proj() ) continue; // Projections handled another way
2353
2354 // Account for all uses
2355 for ( uint k = 0; k < n->len(); k++ ) {
2356 Node *inp = n->in(k);
2357 if (!inp) continue;
2358 assert(inp != n, "no cycles allowed" );
2359 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2360 if (inp->is_Proj()) { // Skip through Proj's
2361 inp = inp->in(0);
2362 }
2363 ++_uses[inp->_idx]; // Count 1 block-local use
2364 }
2365 }
2366
2367 // If this instruction has a 0 use count, then it is available
2368 if (!_uses[n->_idx]) {
2369 _current_latency[n->_idx] = _bundle_cycle_number;
2370 AddNodeToAvailableList(n);
2371 }
2372
2373 #ifndef PRODUCT
2374 if (_cfg->C->trace_opto_output()) {
2375 tty->print("# uses: %3d: ", _uses[n->_idx]);
2376 n->dump();
2377 }
2378 #endif
2379 }
2380
2381 #ifndef PRODUCT
2382 if (_cfg->C->trace_opto_output())
2383 tty->print("# <- ComputeUseCount\n");
2384 #endif
2385 }
2386
2387 // This routine performs scheduling on each basic block in reverse order,
2388 // using instruction latencies and taking into account function unit
2389 // availability.
2390 void Scheduling::DoScheduling() {
2391 #ifndef PRODUCT
2392 if (_cfg->C->trace_opto_output())
2393 tty->print("# -> DoScheduling\n");
2394 #endif
2395
2396 Block *succ_bb = NULL;
2397 Block *bb;
2398
2399 // Walk over all the basic blocks in reverse order
2400 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) {
2401 bb = _cfg->_blocks[i];
2402
2403 #ifndef PRODUCT
2404 if (_cfg->C->trace_opto_output()) {
2405 tty->print("# Schedule BB#%03d (initial)\n", i);
2406 for (uint j = 0; j < bb->_nodes.size(); j++)
2407 bb->_nodes[j]->dump();
2408 }
2409 #endif
2410
2411 // On the head node, skip processing
2412 if( bb == _cfg->_broot )
2413 continue;
2414
2415 // Skip empty, connector blocks
2416 if (bb->is_connector())
2417 continue;
2418
2419 // If the following block is not the sole successor of
2420 // this one, then reset the pipeline information
2421 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2422 #ifndef PRODUCT
2423 if (_cfg->C->trace_opto_output()) {
2424 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2425 _next_node->_idx, _bundle_instr_count);
2426 }
2427 #endif
2428 step_and_clear();
2429 }
2430
2431 // Leave untouched the starting instruction, any Phis, a CreateEx node
2432 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction.
2433 _bb_end = bb->_nodes.size()-1;
2434 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2435 Node *n = bb->_nodes[_bb_start];
2436 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2437 // Also, MachIdealNodes do not get scheduled
2438 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2439 MachNode *mach = n->as_Mach();
2440 int iop = mach->ideal_Opcode();
2441 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2442 if( iop == Op_Con ) continue; // Do not schedule Top
2443 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2444 mach->pipeline() == MachNode::pipeline_class() &&
2445 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2446 continue;
2447 break; // Funny loop structure to be sure...
2448 }
2449 // Compute last "interesting" instruction in block - last instruction we
2450 // might schedule. _bb_end points just after last schedulable inst. We
2451 // normally schedule conditional branches (despite them being forced last
2452 // in the block), because they have delay slots we can fill. Calls all
2453 // have their delay slots filled in the template expansions, so we don't
2454 // bother scheduling them.
2455 Node *last = bb->_nodes[_bb_end];
2456 // Ignore trailing NOPs.
2457 while (_bb_end > 0 && last->is_Mach() &&
2458 last->as_Mach()->ideal_Opcode() == Op_Con) {
2459 last = bb->_nodes[--_bb_end];
2460 }
2461 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2462 if( last->is_Catch() ||
2463 // Exclude unreachable path case when Halt node is in a separate block.
2464 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2465 // There must be a prior call. Skip it.
2466 while( !bb->_nodes[--_bb_end]->is_MachCall() ) {
2467 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" );
2468 }
2469 } else if( last->is_MachNullCheck() ) {
2470 // Backup so the last null-checked memory instruction is
2471 // outside the schedulable range. Skip over the nullcheck,
2472 // projection, and the memory nodes.
2473 Node *mem = last->in(1);
2474 do {
2475 _bb_end--;
2476 } while (mem != bb->_nodes[_bb_end]);
2477 } else {
2478 // Set _bb_end to point after last schedulable inst.
2479 _bb_end++;
2480 }
2481
2482 assert( _bb_start <= _bb_end, "inverted block ends" );
2483
2484 // Compute the register antidependencies for the basic block
2485 ComputeRegisterAntidependencies(bb);
2486 if (_cfg->C->failing()) return; // too many D-U pinch points
2487
2488 // Compute intra-bb latencies for the nodes
2489 ComputeLocalLatenciesForward(bb);
2490
2491 // Compute the usage within the block, and set the list of all nodes
2492 // in the block that have no uses within the block.
2493 ComputeUseCount(bb);
2494
2495 // Schedule the remaining instructions in the block
2496 while ( _available.size() > 0 ) {
2497 Node *n = ChooseNodeToBundle();
2498 guarantee(n != NULL, "no nodes available");
2499 AddNodeToBundle(n,bb);
2500 }
2501
2502 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2503 #ifdef ASSERT
2504 for( uint l = _bb_start; l < _bb_end; l++ ) {
2505 Node *n = bb->_nodes[l];
2506 uint m;
2507 for( m = 0; m < _bb_end-_bb_start; m++ )
2508 if( _scheduled[m] == n )
2509 break;
2510 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2511 }
2512 #endif
2513
2514 // Now copy the instructions (in reverse order) back to the block
2515 for ( uint k = _bb_start; k < _bb_end; k++ )
2516 bb->_nodes.map(k, _scheduled[_bb_end-k-1]);
2517
2518 #ifndef PRODUCT
2519 if (_cfg->C->trace_opto_output()) {
2520 tty->print("# Schedule BB#%03d (final)\n", i);
2521 uint current = 0;
2522 for (uint j = 0; j < bb->_nodes.size(); j++) {
2523 Node *n = bb->_nodes[j];
2524 if( valid_bundle_info(n) ) {
2525 Bundle *bundle = node_bundling(n);
2526 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2527 tty->print("*** Bundle: ");
2528 bundle->dump();
2529 }
2530 n->dump();
2531 }
2532 }
2533 }
2534 #endif
2535 #ifdef ASSERT
2536 verify_good_schedule(bb,"after block local scheduling");
2537 #endif
2538 }
2539
2540 #ifndef PRODUCT
2541 if (_cfg->C->trace_opto_output())
2542 tty->print("# <- DoScheduling\n");
2543 #endif
2544
2545 // Record final node-bundling array location
2546 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2547
2548 } // end DoScheduling
2549
2550 //------------------------------verify_good_schedule---------------------------
2551 // Verify that no live-range used in the block is killed in the block by a
2552 // wrong DEF. This doesn't verify live-ranges that span blocks.
2553
2554 // Check for edge existence. Used to avoid adding redundant precedence edges.
2555 static bool edge_from_to( Node *from, Node *to ) {
2556 for( uint i=0; i<from->len(); i++ )
2557 if( from->in(i) == to )
2558 return true;
2559 return false;
2560 }
2561
2562 #ifdef ASSERT
2563 //------------------------------verify_do_def----------------------------------
2564 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2565 // Check for bad kills
2566 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2567 Node *prior_use = _reg_node[def];
2568 if( prior_use && !edge_from_to(prior_use,n) ) {
2569 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2570 n->dump();
2571 tty->print_cr("...");
2572 prior_use->dump();
2573 assert(edge_from_to(prior_use,n),msg);
2574 }
2575 _reg_node.map(def,NULL); // Kill live USEs
2576 }
2577 }
2578
2579 //------------------------------verify_good_schedule---------------------------
2580 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2581
2582 // Zap to something reasonable for the verify code
2583 _reg_node.clear();
2584
2585 // Walk over the block backwards. Check to make sure each DEF doesn't
2586 // kill a live value (other than the one it's supposed to). Add each
2587 // USE to the live set.
2588 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {
2589 Node *n = b->_nodes[i];
2590 int n_op = n->Opcode();
2591 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2592 // Fat-proj kills a slew of registers
2593 RegMask rm = n->out_RegMask();// Make local copy
2594 while( rm.is_NotEmpty() ) {
2595 OptoReg::Name kill = rm.find_first_elem();
2596 rm.Remove(kill);
2597 verify_do_def( n, kill, msg );
2598 }
2599 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2600 // Get DEF'd registers the normal way
2601 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2602 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2603 }
2604
2605 // Now make all USEs live
2606 for( uint i=1; i<n->req(); i++ ) {
2607 Node *def = n->in(i);
2608 assert(def != 0, "input edge required");
2609 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2610 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2611 if( OptoReg::is_valid(reg_lo) ) {
2612 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2613 _reg_node.map(reg_lo,n);
2614 }
2615 if( OptoReg::is_valid(reg_hi) ) {
2616 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2617 _reg_node.map(reg_hi,n);
2618 }
2619 }
2620
2621 }
2622
2623 // Zap to something reasonable for the Antidependence code
2624 _reg_node.clear();
2625 }
2626 #endif
2627
2628 // Conditionally add precedence edges. Avoid putting edges on Projs.
2629 static void add_prec_edge_from_to( Node *from, Node *to ) {
2630 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2631 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2632 from = from->in(0);
2633 }
2634 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2635 !edge_from_to( from, to ) ) // Avoid duplicate edge
2636 from->add_prec(to);
2637 }
2638
2639 //------------------------------anti_do_def------------------------------------
2640 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2641 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2642 return;
2643
2644 Node *pinch = _reg_node[def_reg]; // Get pinch point
2645 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2646 is_def ) { // Check for a true def (not a kill)
2647 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2648 return;
2649 }
2650
2651 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2652 debug_only( def = (Node*)0xdeadbeef; )
2653
2654 // After some number of kills there _may_ be a later def
2655 Node *later_def = NULL;
2656
2657 // Finding a kill requires a real pinch-point.
2658 // Check for not already having a pinch-point.
2659 // Pinch points are Op_Node's.
2660 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2661 later_def = pinch; // Must be def/kill as optimistic pinch-point
2662 if ( _pinch_free_list.size() > 0) {
2663 pinch = _pinch_free_list.pop();
2664 } else {
2665 pinch = new (_cfg->C) Node(1); // Pinch point to-be
2666 }
2667 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2668 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2669 return;
2670 }
2671 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2672 _reg_node.map(def_reg,pinch); // Record pinch-point
2673 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2674 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2675 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2676 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2677 later_def = NULL; // and no later def
2678 }
2679 pinch->set_req(0,later_def); // Hook later def so we can find it
2680 } else { // Else have valid pinch point
2681 if( pinch->in(0) ) // If there is a later-def
2682 later_def = pinch->in(0); // Get it
2683 }
2684
2685 // Add output-dependence edge from later def to kill
2686 if( later_def ) // If there is some original def
2687 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2688
2689 // See if current kill is also a use, and so is forced to be the pinch-point.
2690 if( pinch->Opcode() == Op_Node ) {
2691 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2692 for( uint i=1; i<uses->req(); i++ ) {
2693 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2694 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2695 // Yes, found a use/kill pinch-point
2696 pinch->set_req(0,NULL); //
2697 pinch->replace_by(kill); // Move anti-dep edges up
2698 pinch = kill;
2699 _reg_node.map(def_reg,pinch);
2700 return;
2701 }
2702 }
2703 }
2704
2705 // Add edge from kill to pinch-point
2706 add_prec_edge_from_to(kill,pinch);
2707 }
2708
2709 //------------------------------anti_do_use------------------------------------
2710 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2711 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2712 return;
2713 Node *pinch = _reg_node[use_reg]; // Get pinch point
2714 // Check for no later def_reg/kill in block
2715 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2716 // Use has to be block-local as well
2717 _cfg->get_block_for_node(use) == b) {
2718 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2719 pinch->req() == 1 ) { // pinch not yet in block?
2720 pinch->del_req(0); // yank pointer to later-def, also set flag
2721 // Insert the pinch-point in the block just after the last use
2722 b->_nodes.insert(b->find_node(use)+1,pinch);
2723 _bb_end++; // Increase size scheduled region in block
2724 }
2725
2726 add_prec_edge_from_to(pinch,use);
2727 }
2728 }
2729
2730 //------------------------------ComputeRegisterAntidependences-----------------
2731 // We insert antidependences between the reads and following write of
2732 // allocated registers to prevent illegal code motion. Hopefully, the
2733 // number of added references should be fairly small, especially as we
2734 // are only adding references within the current basic block.
2735 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2736
2737 #ifdef ASSERT
2738 verify_good_schedule(b,"before block local scheduling");
2739 #endif
2740
2741 // A valid schedule, for each register independently, is an endless cycle
2742 // of: a def, then some uses (connected to the def by true dependencies),
2743 // then some kills (defs with no uses), finally the cycle repeats with a new
2744 // def. The uses are allowed to float relative to each other, as are the
2745 // kills. No use is allowed to slide past a kill (or def). This requires
2746 // antidependencies between all uses of a single def and all kills that
2747 // follow, up to the next def. More edges are redundant, because later defs
2748 // & kills are already serialized with true or antidependencies. To keep
2749 // the edge count down, we add a 'pinch point' node if there's more than
2750 // one use or more than one kill/def.
2751
2752 // We add dependencies in one bottom-up pass.
2753
2754 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2755
2756 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2757 // register. If not, we record the DEF/KILL in _reg_node, the
2758 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2759 // "pinch point", a new Node that's in the graph but not in the block.
2760 // We put edges from the prior and current DEF/KILLs to the pinch point.
2761 // We put the pinch point in _reg_node. If there's already a pinch point
2762 // we merely add an edge from the current DEF/KILL to the pinch point.
2763
2764 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2765 // put an edge from the pinch point to the USE.
2766
2767 // To be expedient, the _reg_node array is pre-allocated for the whole
2768 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2769 // or a valid def/kill/pinch-point, or a leftover node from some prior
2770 // block. Leftover node from some prior block is treated like a NULL (no
2771 // prior def, so no anti-dependence needed). Valid def is distinguished by
2772 // it being in the current block.
2773 bool fat_proj_seen = false;
2774 uint last_safept = _bb_end-1;
2775 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL;
2776 Node* last_safept_node = end_node;
2777 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2778 Node *n = b->_nodes[i];
2779 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2780 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2781 // Fat-proj kills a slew of registers
2782 // This can add edges to 'n' and obscure whether or not it was a def,
2783 // hence the is_def flag.
2784 fat_proj_seen = true;
2785 RegMask rm = n->out_RegMask();// Make local copy
2786 while( rm.is_NotEmpty() ) {
2787 OptoReg::Name kill = rm.find_first_elem();
2788 rm.Remove(kill);
2789 anti_do_def( b, n, kill, is_def );
2790 }
2791 } else {
2792 // Get DEF'd registers the normal way
2793 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2794 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2795 }
2796
2797 // Kill projections on a branch should appear to occur on the
2798 // branch, not afterwards, so grab the masks from the projections
2799 // and process them.
2800 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2801 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2802 Node* use = n->fast_out(i);
2803 if (use->is_Proj()) {
2804 RegMask rm = use->out_RegMask();// Make local copy
2805 while( rm.is_NotEmpty() ) {
2806 OptoReg::Name kill = rm.find_first_elem();
2807 rm.Remove(kill);
2808 anti_do_def( b, n, kill, false );
2809 }
2810 }
2811 }
2812 }
2813
2814 // Check each register used by this instruction for a following DEF/KILL
2815 // that must occur afterward and requires an anti-dependence edge.
2816 for( uint j=0; j<n->req(); j++ ) {
2817 Node *def = n->in(j);
2818 if( def ) {
2819 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2820 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2821 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2822 }
2823 }
2824 // Do not allow defs of new derived values to float above GC
2825 // points unless the base is definitely available at the GC point.
2826
2827 Node *m = b->_nodes[i];
2828
2829 // Add precedence edge from following safepoint to use of derived pointer
2830 if( last_safept_node != end_node &&
2831 m != last_safept_node) {
2832 for (uint k = 1; k < m->req(); k++) {
2833 const Type *t = m->in(k)->bottom_type();
2834 if( t->isa_oop_ptr() &&
2835 t->is_ptr()->offset() != 0 ) {
2836 last_safept_node->add_prec( m );
2837 break;
2838 }
2839 }
2840 }
2841
2842 if( n->jvms() ) { // Precedence edge from derived to safept
2843 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2844 if( b->_nodes[last_safept] != last_safept_node ) {
2845 last_safept = b->find_node(last_safept_node);
2846 }
2847 for( uint j=last_safept; j > i; j-- ) {
2848 Node *mach = b->_nodes[j];
2849 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2850 mach->add_prec( n );
2851 }
2852 last_safept = i;
2853 last_safept_node = m;
2854 }
2855 }
2856
2857 if (fat_proj_seen) {
2858 // Garbage collect pinch nodes that were not consumed.
2859 // They are usually created by a fat kill MachProj for a call.
2860 garbage_collect_pinch_nodes();
2861 }
2862 }
2863
2864 //------------------------------garbage_collect_pinch_nodes-------------------------------
2865
2866 // Garbage collect pinch nodes for reuse by other blocks.
2867 //
2868 // The block scheduler's insertion of anti-dependence
2869 // edges creates many pinch nodes when the block contains
2870 // 2 or more Calls. A pinch node is used to prevent a
2871 // combinatorial explosion of edges. If a set of kills for a
2872 // register is anti-dependent on a set of uses (or defs), rather
2873 // than adding an edge in the graph between each pair of kill
2874 // and use (or def), a pinch is inserted between them:
2875 //
2876 // use1 use2 use3
2877 // \ | /
2878 // \ | /
2879 // pinch
2880 // / | \
2881 // / | \
2882 // kill1 kill2 kill3
2883 //
2884 // One pinch node is created per register killed when
2885 // the second call is encountered during a backwards pass
2886 // over the block. Most of these pinch nodes are never
2887 // wired into the graph because the register is never
2888 // used or def'ed in the block.
2889 //
2890 void Scheduling::garbage_collect_pinch_nodes() {
2891 #ifndef PRODUCT
2892 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2893 #endif
2894 int trace_cnt = 0;
2895 for (uint k = 0; k < _reg_node.Size(); k++) {
2896 Node* pinch = _reg_node[k];
2897 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2898 // no predecence input edges
2899 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2900 cleanup_pinch(pinch);
2901 _pinch_free_list.push(pinch);
2902 _reg_node.map(k, NULL);
2903 #ifndef PRODUCT
2904 if (_cfg->C->trace_opto_output()) {
2905 trace_cnt++;
2906 if (trace_cnt > 40) {
2907 tty->print("\n");
2908 trace_cnt = 0;
2909 }
2910 tty->print(" %d", pinch->_idx);
2911 }
2912 #endif
2913 }
2914 }
2915 #ifndef PRODUCT
2916 if (_cfg->C->trace_opto_output()) tty->print("\n");
2917 #endif
2918 }
2919
2920 // Clean up a pinch node for reuse.
2921 void Scheduling::cleanup_pinch( Node *pinch ) {
2922 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2923
2924 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2925 Node* use = pinch->last_out(i);
2926 uint uses_found = 0;
2927 for (uint j = use->req(); j < use->len(); j++) {
2928 if (use->in(j) == pinch) {
2929 use->rm_prec(j);
2930 uses_found++;
2931 }
2932 }
2933 assert(uses_found > 0, "must be a precedence edge");
2934 i -= uses_found; // we deleted 1 or more copies of this edge
2935 }
2936 // May have a later_def entry
2937 pinch->set_req(0, NULL);
2938 }
2939
2940 //------------------------------print_statistics-------------------------------
2941 #ifndef PRODUCT
2942
2943 void Scheduling::dump_available() const {
2944 tty->print("#Availist ");
2945 for (uint i = 0; i < _available.size(); i++)
2946 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2947 tty->cr();
2948 }
2949
2950 // Print Scheduling Statistics
2951 void Scheduling::print_statistics() {
2952 // Print the size added by nops for bundling
2953 tty->print("Nops added %d bytes to total of %d bytes",
2954 _total_nop_size, _total_method_size);
2955 if (_total_method_size > 0)
2956 tty->print(", for %.2f%%",
2957 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2958 tty->print("\n");
2959
2960 // Print the number of branch shadows filled
2961 if (Pipeline::_branch_has_delay_slot) {
2962 tty->print("Of %d branches, %d had unconditional delay slots filled",
2963 _total_branches, _total_unconditional_delays);
2964 if (_total_branches > 0)
2965 tty->print(", for %.2f%%",
2966 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2967 tty->print("\n");
2968 }
2969
2970 uint total_instructions = 0, total_bundles = 0;
2971
2972 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2973 uint bundle_count = _total_instructions_per_bundle[i];
2974 total_instructions += bundle_count * i;
2975 total_bundles += bundle_count;
2976 }
2977
2978 if (total_bundles > 0)
2979 tty->print("Average ILP (excluding nops) is %.2f\n",
2980 ((double)total_instructions) / ((double)total_bundles));
2981 }
2982 #endif