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