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