Thu, 11 Apr 2013 13:57:44 +0200
8006952: Slow VM due to excessive code cache freelist iteration
Summary: Remove continous free block requirement
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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"
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();
49 #ifndef PRODUCT
50 #define DEBUG_ARG(x) , x
51 #else
52 #define DEBUG_ARG(x)
53 #endif
55 extern int emit_exception_handler(CodeBuffer &cbuf);
56 extern int emit_deopt_handler(CodeBuffer &cbuf);
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, "" );
64 // The number of new nodes (mostly MachNop) is proportional to
65 // the number of java calls and inner loops which are aligned.
66 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
67 C->inner_loops()*(OptoLoopAlignment-1)),
68 "out of nodes before code generation" ) ) {
69 return;
70 }
71 // Make sure I can find the Start Node
72 Block_Array& bbs = _cfg->_bbs;
73 Block *entry = _cfg->_blocks[1];
74 Block *broot = _cfg->_broot;
76 const StartNode *start = entry->_nodes[0]->as_Start();
78 // Replace StartNode with prolog
79 MachPrologNode *prolog = new (this) MachPrologNode();
80 entry->_nodes.map( 0, prolog );
81 bbs.map( prolog->_idx, entry );
82 bbs.map( start->_idx, NULL ); // start is no longer in any block
84 // Virtual methods need an unverified entry point
86 if( is_osr_compilation() ) {
87 if( PoisonOSREntry ) {
88 // TODO: Should use a ShouldNotReachHereNode...
89 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
90 }
91 } else {
92 if( _method && !_method->flags().is_static() ) {
93 // Insert unvalidated entry point
94 _cfg->insert( broot, 0, new (this) MachUEPNode() );
95 }
97 }
100 // Break before main entry point
101 if( (_method && _method->break_at_execute())
102 #ifndef PRODUCT
103 ||(OptoBreakpoint && is_method_compilation())
104 ||(OptoBreakpointOSR && is_osr_compilation())
105 ||(OptoBreakpointC2R && !_method)
106 #endif
107 ) {
108 // checking for _method means that OptoBreakpoint does not apply to
109 // runtime stubs or frame converters
110 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
111 }
113 // Insert epilogs before every return
114 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
115 Block *b = _cfg->_blocks[i];
116 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point?
117 Node *m = b->end();
118 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) {
119 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
120 b->add_inst( epilog );
121 bbs.map(epilog->_idx, b);
122 //_regalloc->set_bad(epilog->_idx); // Already initialized this way.
123 }
124 }
125 }
127 # ifdef ENABLE_ZAP_DEAD_LOCALS
128 if ( ZapDeadCompiledLocals ) Insert_zap_nodes();
129 # endif
131 uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
132 blk_starts[0] = 0;
134 // Initialize code buffer and process short branches.
135 CodeBuffer* cb = init_buffer(blk_starts);
137 if (cb == NULL || failing()) return;
139 ScheduleAndBundle();
141 #ifndef PRODUCT
142 if (trace_opto_output()) {
143 tty->print("\n---- After ScheduleAndBundle ----\n");
144 for (uint i = 0; i < _cfg->_num_blocks; i++) {
145 tty->print("\nBB#%03d:\n", i);
146 Block *bb = _cfg->_blocks[i];
147 for (uint j = 0; j < bb->_nodes.size(); j++) {
148 Node *n = bb->_nodes[j];
149 OptoReg::Name reg = _regalloc->get_reg_first(n);
150 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
151 n->dump();
152 }
153 }
154 }
155 #endif
157 if (failing()) return;
159 BuildOopMaps();
161 if (failing()) return;
163 fill_buffer(cb, blk_starts);
164 }
166 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
167 // Determine if we need to generate a stack overflow check.
168 // Do it if the method is not a stub function and
169 // has java calls or has frame size > vm_page_size/8.
170 return (UseStackBanging && stub_function() == NULL &&
171 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
172 }
174 bool Compile::need_register_stack_bang() const {
175 // Determine if we need to generate a register stack overflow check.
176 // This is only used on architectures which have split register
177 // and memory stacks (ie. IA64).
178 // Bang if the method is not a stub function and has java calls
179 return (stub_function() == NULL && has_java_calls());
180 }
182 # ifdef ENABLE_ZAP_DEAD_LOCALS
185 // In order to catch compiler oop-map bugs, we have implemented
186 // a debugging mode called ZapDeadCompilerLocals.
187 // This mode causes the compiler to insert a call to a runtime routine,
188 // "zap_dead_locals", right before each place in compiled code
189 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
190 // The runtime routine checks that locations mapped as oops are really
191 // oops, that locations mapped as values do not look like oops,
192 // and that locations mapped as dead are not used later
193 // (by zapping them to an invalid address).
195 int Compile::_CompiledZap_count = 0;
197 void Compile::Insert_zap_nodes() {
198 bool skip = false;
201 // Dink with static counts because code code without the extra
202 // runtime calls is MUCH faster for debugging purposes
204 if ( CompileZapFirst == 0 ) ; // nothing special
205 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
206 else if ( CompileZapFirst == CompiledZap_count() )
207 warning("starting zap compilation after skipping");
209 if ( CompileZapLast == -1 ) ; // nothing special
210 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
211 else if ( CompileZapLast == CompiledZap_count() )
212 warning("about to compile last zap");
214 ++_CompiledZap_count; // counts skipped zaps, too
216 if ( skip ) return;
219 if ( _method == NULL )
220 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
222 // Insert call to zap runtime stub before every node with an oop map
223 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
224 Block *b = _cfg->_blocks[i];
225 for ( uint j = 0; j < b->_nodes.size(); ++j ) {
226 Node *n = b->_nodes[j];
228 // Determining if we should insert a zap-a-lot node in output.
229 // We do that for all nodes that has oopmap info, except for calls
230 // to allocation. Calls to allocation passes in the old top-of-eden pointer
231 // and expect the C code to reset it. Hence, there can be no safepoints between
232 // the inlined-allocation and the call to new_Java, etc.
233 // We also cannot zap monitor calls, as they must hold the microlock
234 // during the call to Zap, which also wants to grab the microlock.
235 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
236 if ( insert ) { // it is MachSafePoint
237 if ( !n->is_MachCall() ) {
238 insert = false;
239 } else if ( n->is_MachCall() ) {
240 MachCallNode* call = n->as_MachCall();
241 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
242 call->entry_point() == OptoRuntime::new_array_Java() ||
243 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
244 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
245 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
246 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
247 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
248 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
249 ) {
250 insert = false;
251 }
252 }
253 if (insert) {
254 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
255 b->_nodes.insert( j, zap );
256 _cfg->_bbs.map( zap->_idx, b );
257 ++j;
258 }
259 }
260 }
261 }
262 }
265 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
266 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
267 CallStaticJavaNode* ideal_node =
268 new (this) CallStaticJavaNode( tf,
269 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
270 "call zap dead locals stub", 0, TypePtr::BOTTOM);
271 // We need to copy the OopMap from the site we're zapping at.
272 // We have to make a copy, because the zap site might not be
273 // a call site, and zap_dead is a call site.
274 OopMap* clone = node_to_check->oop_map()->deep_copy();
276 // Add the cloned OopMap to the zap node
277 ideal_node->set_oop_map(clone);
278 return _matcher->match_sfpt(ideal_node);
279 }
281 //------------------------------is_node_getting_a_safepoint--------------------
282 bool Compile::is_node_getting_a_safepoint( Node* n) {
283 // This code duplicates the logic prior to the call of add_safepoint
284 // below in this file.
285 if( n->is_MachSafePoint() ) return true;
286 return false;
287 }
289 # endif // ENABLE_ZAP_DEAD_LOCALS
291 //------------------------------compute_loop_first_inst_sizes------------------
292 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
293 // of a loop. When aligning a loop we need to provide enough instructions
294 // in cpu's fetch buffer to feed decoders. The loop alignment could be
295 // avoided if we have enough instructions in fetch buffer at the head of a loop.
296 // By default, the size is set to 999999 by Block's constructor so that
297 // a loop will be aligned if the size is not reset here.
298 //
299 // Note: Mach instructions could contain several HW instructions
300 // so the size is estimated only.
301 //
302 void Compile::compute_loop_first_inst_sizes() {
303 // The next condition is used to gate the loop alignment optimization.
304 // Don't aligned a loop if there are enough instructions at the head of a loop
305 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
306 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
307 // equal to 11 bytes which is the largest address NOP instruction.
308 if( MaxLoopPad < OptoLoopAlignment-1 ) {
309 uint last_block = _cfg->_num_blocks-1;
310 for( uint i=1; i <= last_block; i++ ) {
311 Block *b = _cfg->_blocks[i];
312 // Check the first loop's block which requires an alignment.
313 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {
314 uint sum_size = 0;
315 uint inst_cnt = NumberOfLoopInstrToAlign;
316 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
318 // Check subsequent fallthrough blocks if the loop's first
319 // block(s) does not have enough instructions.
320 Block *nb = b;
321 while( inst_cnt > 0 &&
322 i < last_block &&
323 !_cfg->_blocks[i+1]->has_loop_alignment() &&
324 !nb->has_successor(b) ) {
325 i++;
326 nb = _cfg->_blocks[i];
327 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
328 } // while( inst_cnt > 0 && i < last_block )
330 b->set_first_inst_size(sum_size);
331 } // f( b->head()->is_Loop() )
332 } // for( i <= last_block )
333 } // if( MaxLoopPad < OptoLoopAlignment-1 )
334 }
336 //----------------------shorten_branches---------------------------------------
337 // The architecture description provides short branch variants for some long
338 // branch instructions. Replace eligible long branches with short branches.
339 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
341 // ------------------
342 // Compute size of each block, method size, and relocation information size
343 uint nblocks = _cfg->_num_blocks;
345 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
346 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
347 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
348 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
349 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
351 bool has_short_branch_candidate = false;
353 // Initialize the sizes to 0
354 code_size = 0; // Size in bytes of generated code
355 stub_size = 0; // Size in bytes of all stub entries
356 // Size in bytes of all relocation entries, including those in local stubs.
357 // Start with 2-bytes of reloc info for the unvalidated entry point
358 reloc_size = 1; // Number of relocation entries
360 // Make three passes. The first computes pessimistic blk_starts,
361 // relative jmp_offset and reloc_size information. The second performs
362 // short branch substitution using the pessimistic sizing. The
363 // third inserts nops where needed.
365 // Step one, perform a pessimistic sizing pass.
366 uint last_call_adr = max_uint;
367 uint last_avoid_back_to_back_adr = max_uint;
368 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
369 for (uint i = 0; i < nblocks; i++) { // For all blocks
370 Block *b = _cfg->_blocks[i];
372 // During short branch replacement, we store the relative (to blk_starts)
373 // offset of jump in jmp_offset, rather than the absolute offset of jump.
374 // This is so that we do not need to recompute sizes of all nodes when
375 // we compute correct blk_starts in our next sizing pass.
376 jmp_offset[i] = 0;
377 jmp_size[i] = 0;
378 jmp_nidx[i] = -1;
379 DEBUG_ONLY( jmp_target[i] = 0; )
380 DEBUG_ONLY( jmp_rule[i] = 0; )
382 // Sum all instruction sizes to compute block size
383 uint last_inst = b->_nodes.size();
384 uint blk_size = 0;
385 for (uint j = 0; j < last_inst; j++) {
386 Node* nj = b->_nodes[j];
387 // Handle machine instruction nodes
388 if (nj->is_Mach()) {
389 MachNode *mach = nj->as_Mach();
390 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
391 reloc_size += mach->reloc();
392 if( mach->is_MachCall() ) {
393 MachCallNode *mcall = mach->as_MachCall();
394 // This destination address is NOT PC-relative
396 mcall->method_set((intptr_t)mcall->entry_point());
398 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) {
399 stub_size += size_java_to_interp();
400 reloc_size += reloc_java_to_interp();
401 }
402 } else if (mach->is_MachSafePoint()) {
403 // If call/safepoint are adjacent, account for possible
404 // nop to disambiguate the two safepoints.
405 // ScheduleAndBundle() can rearrange nodes in a block,
406 // check for all offsets inside this block.
407 if (last_call_adr >= blk_starts[i]) {
408 blk_size += nop_size;
409 }
410 }
411 if (mach->avoid_back_to_back()) {
412 // Nop is inserted between "avoid back to back" instructions.
413 // ScheduleAndBundle() can rearrange nodes in a block,
414 // check for all offsets inside this block.
415 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
416 blk_size += nop_size;
417 }
418 }
419 if (mach->may_be_short_branch()) {
420 if (!nj->is_MachBranch()) {
421 #ifndef PRODUCT
422 nj->dump(3);
423 #endif
424 Unimplemented();
425 }
426 assert(jmp_nidx[i] == -1, "block should have only one branch");
427 jmp_offset[i] = blk_size;
428 jmp_size[i] = nj->size(_regalloc);
429 jmp_nidx[i] = j;
430 has_short_branch_candidate = true;
431 }
432 }
433 blk_size += nj->size(_regalloc);
434 // Remember end of call offset
435 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
436 last_call_adr = blk_starts[i]+blk_size;
437 }
438 // Remember end of avoid_back_to_back offset
439 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
440 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
441 }
442 }
444 // When the next block starts a loop, we may insert pad NOP
445 // instructions. Since we cannot know our future alignment,
446 // assume the worst.
447 if (i< nblocks-1) {
448 Block *nb = _cfg->_blocks[i+1];
449 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
450 if (max_loop_pad > 0) {
451 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
452 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
453 // If either is the last instruction in this block, bump by
454 // max_loop_pad in lock-step with blk_size, so sizing
455 // calculations in subsequent blocks still can conservatively
456 // detect that it may the last instruction in this block.
457 if (last_call_adr == blk_starts[i]+blk_size) {
458 last_call_adr += max_loop_pad;
459 }
460 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
461 last_avoid_back_to_back_adr += max_loop_pad;
462 }
463 blk_size += max_loop_pad;
464 }
465 }
467 // Save block size; update total method size
468 blk_starts[i+1] = blk_starts[i]+blk_size;
469 }
471 // Step two, replace eligible long jumps.
472 bool progress = true;
473 uint last_may_be_short_branch_adr = max_uint;
474 while (has_short_branch_candidate && progress) {
475 progress = false;
476 has_short_branch_candidate = false;
477 int adjust_block_start = 0;
478 for (uint i = 0; i < nblocks; i++) {
479 Block *b = _cfg->_blocks[i];
480 int idx = jmp_nidx[i];
481 MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach();
482 if (mach != NULL && mach->may_be_short_branch()) {
483 #ifdef ASSERT
484 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
485 int j;
486 // Find the branch; ignore trailing NOPs.
487 for (j = b->_nodes.size()-1; j>=0; j--) {
488 Node* n = b->_nodes[j];
489 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
490 break;
491 }
492 assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity");
493 #endif
494 int br_size = jmp_size[i];
495 int br_offs = blk_starts[i] + jmp_offset[i];
497 // This requires the TRUE branch target be in succs[0]
498 uint bnum = b->non_connector_successor(0)->_pre_order;
499 int offset = blk_starts[bnum] - br_offs;
500 if (bnum > i) { // adjust following block's offset
501 offset -= adjust_block_start;
502 }
503 // In the following code a nop could be inserted before
504 // the branch which will increase the backward distance.
505 bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);
506 if (needs_padding && offset <= 0)
507 offset -= nop_size;
509 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
510 // We've got a winner. Replace this branch.
511 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
513 // Update the jmp_size.
514 int new_size = replacement->size(_regalloc);
515 int diff = br_size - new_size;
516 assert(diff >= (int)nop_size, "short_branch size should be smaller");
517 // Conservatively take into accound padding between
518 // avoid_back_to_back branches. Previous branch could be
519 // converted into avoid_back_to_back branch during next
520 // rounds.
521 if (needs_padding && replacement->avoid_back_to_back()) {
522 jmp_offset[i] += nop_size;
523 diff -= nop_size;
524 }
525 adjust_block_start += diff;
526 b->_nodes.map(idx, replacement);
527 mach->subsume_by(replacement, C);
528 mach = replacement;
529 progress = true;
531 jmp_size[i] = new_size;
532 DEBUG_ONLY( jmp_target[i] = bnum; );
533 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
534 } else {
535 // The jump distance is not short, try again during next iteration.
536 has_short_branch_candidate = true;
537 }
538 } // (mach->may_be_short_branch())
539 if (mach != NULL && (mach->may_be_short_branch() ||
540 mach->avoid_back_to_back())) {
541 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
542 }
543 blk_starts[i+1] -= adjust_block_start;
544 }
545 }
547 #ifdef ASSERT
548 for (uint i = 0; i < nblocks; i++) { // For all blocks
549 if (jmp_target[i] != 0) {
550 int br_size = jmp_size[i];
551 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
552 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
553 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
554 }
555 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
556 }
557 }
558 #endif
560 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
561 // after ScheduleAndBundle().
563 // ------------------
564 // Compute size for code buffer
565 code_size = blk_starts[nblocks];
567 // Relocation records
568 reloc_size += 1; // Relo entry for exception handler
570 // Adjust reloc_size to number of record of relocation info
571 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
572 // a relocation index.
573 // The CodeBuffer will expand the locs array if this estimate is too low.
574 reloc_size *= 10 / sizeof(relocInfo);
575 }
577 //------------------------------FillLocArray-----------------------------------
578 // Create a bit of debug info and append it to the array. The mapping is from
579 // Java local or expression stack to constant, register or stack-slot. For
580 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
581 // entry has been taken care of and caller should skip it).
582 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
583 // This should never have accepted Bad before
584 assert(OptoReg::is_valid(regnum), "location must be valid");
585 return (OptoReg::is_reg(regnum))
586 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
587 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
588 }
591 ObjectValue*
592 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
593 for (int i = 0; i < objs->length(); i++) {
594 assert(objs->at(i)->is_object(), "corrupt object cache");
595 ObjectValue* sv = (ObjectValue*) objs->at(i);
596 if (sv->id() == id) {
597 return sv;
598 }
599 }
600 // Otherwise..
601 return NULL;
602 }
604 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
605 ObjectValue* sv ) {
606 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
607 objs->append(sv);
608 }
611 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
612 GrowableArray<ScopeValue*> *array,
613 GrowableArray<ScopeValue*> *objs ) {
614 assert( local, "use _top instead of null" );
615 if (array->length() != idx) {
616 assert(array->length() == idx + 1, "Unexpected array count");
617 // Old functionality:
618 // return
619 // New functionality:
620 // Assert if the local is not top. In product mode let the new node
621 // override the old entry.
622 assert(local == top(), "LocArray collision");
623 if (local == top()) {
624 return;
625 }
626 array->pop();
627 }
628 const Type *t = local->bottom_type();
630 // Is it a safepoint scalar object node?
631 if (local->is_SafePointScalarObject()) {
632 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
634 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
635 if (sv == NULL) {
636 ciKlass* cik = t->is_oopptr()->klass();
637 assert(cik->is_instance_klass() ||
638 cik->is_array_klass(), "Not supported allocation.");
639 sv = new ObjectValue(spobj->_idx,
640 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
641 Compile::set_sv_for_object_node(objs, sv);
643 uint first_ind = spobj->first_index();
644 for (uint i = 0; i < spobj->n_fields(); i++) {
645 Node* fld_node = sfpt->in(first_ind+i);
646 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
647 }
648 }
649 array->append(sv);
650 return;
651 }
653 // Grab the register number for the local
654 OptoReg::Name regnum = _regalloc->get_reg_first(local);
655 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
656 // Record the double as two float registers.
657 // The register mask for such a value always specifies two adjacent
658 // float registers, with the lower register number even.
659 // Normally, the allocation of high and low words to these registers
660 // is irrelevant, because nearly all operations on register pairs
661 // (e.g., StoreD) treat them as a single unit.
662 // Here, we assume in addition that the words in these two registers
663 // stored "naturally" (by operations like StoreD and double stores
664 // within the interpreter) such that the lower-numbered register
665 // is written to the lower memory address. This may seem like
666 // a machine dependency, but it is not--it is a requirement on
667 // the author of the <arch>.ad file to ensure that, for every
668 // even/odd double-register pair to which a double may be allocated,
669 // the word in the even single-register is stored to the first
670 // memory word. (Note that register numbers are completely
671 // arbitrary, and are not tied to any machine-level encodings.)
672 #ifdef _LP64
673 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
674 array->append(new ConstantIntValue(0));
675 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
676 } else if ( t->base() == Type::Long ) {
677 array->append(new ConstantIntValue(0));
678 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
679 } else if ( t->base() == Type::RawPtr ) {
680 // jsr/ret return address which must be restored into a the full
681 // width 64-bit stack slot.
682 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
683 }
684 #else //_LP64
685 #ifdef SPARC
686 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
687 // For SPARC we have to swap high and low words for
688 // long values stored in a single-register (g0-g7).
689 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
690 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
691 } else
692 #endif //SPARC
693 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
694 // Repack the double/long as two jints.
695 // The convention the interpreter uses is that the second local
696 // holds the first raw word of the native double representation.
697 // This is actually reasonable, since locals and stack arrays
698 // grow downwards in all implementations.
699 // (If, on some machine, the interpreter's Java locals or stack
700 // were to grow upwards, the embedded doubles would be word-swapped.)
701 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
702 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
703 }
704 #endif //_LP64
705 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
706 OptoReg::is_reg(regnum) ) {
707 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
708 ? Location::float_in_dbl : Location::normal ));
709 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
710 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
711 ? Location::int_in_long : Location::normal ));
712 } else if( t->base() == Type::NarrowOop ) {
713 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
714 } else {
715 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
716 }
717 return;
718 }
720 // No register. It must be constant data.
721 switch (t->base()) {
722 case Type::Half: // Second half of a double
723 ShouldNotReachHere(); // Caller should skip 2nd halves
724 break;
725 case Type::AnyPtr:
726 array->append(new ConstantOopWriteValue(NULL));
727 break;
728 case Type::AryPtr:
729 case Type::InstPtr: // fall through
730 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
731 break;
732 case Type::NarrowOop:
733 if (t == TypeNarrowOop::NULL_PTR) {
734 array->append(new ConstantOopWriteValue(NULL));
735 } else {
736 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
737 }
738 break;
739 case Type::Int:
740 array->append(new ConstantIntValue(t->is_int()->get_con()));
741 break;
742 case Type::RawPtr:
743 // A return address (T_ADDRESS).
744 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
745 #ifdef _LP64
746 // Must be restored to the full-width 64-bit stack slot.
747 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
748 #else
749 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
750 #endif
751 break;
752 case Type::FloatCon: {
753 float f = t->is_float_constant()->getf();
754 array->append(new ConstantIntValue(jint_cast(f)));
755 break;
756 }
757 case Type::DoubleCon: {
758 jdouble d = t->is_double_constant()->getd();
759 #ifdef _LP64
760 array->append(new ConstantIntValue(0));
761 array->append(new ConstantDoubleValue(d));
762 #else
763 // Repack the double as two jints.
764 // The convention the interpreter uses is that the second local
765 // holds the first raw word of the native double representation.
766 // This is actually reasonable, since locals and stack arrays
767 // grow downwards in all implementations.
768 // (If, on some machine, the interpreter's Java locals or stack
769 // were to grow upwards, the embedded doubles would be word-swapped.)
770 jint *dp = (jint*)&d;
771 array->append(new ConstantIntValue(dp[1]));
772 array->append(new ConstantIntValue(dp[0]));
773 #endif
774 break;
775 }
776 case Type::Long: {
777 jlong d = t->is_long()->get_con();
778 #ifdef _LP64
779 array->append(new ConstantIntValue(0));
780 array->append(new ConstantLongValue(d));
781 #else
782 // Repack the long as two jints.
783 // The convention the interpreter uses is that the second local
784 // holds the first raw word of the native double representation.
785 // This is actually reasonable, since locals and stack arrays
786 // grow downwards in all implementations.
787 // (If, on some machine, the interpreter's Java locals or stack
788 // were to grow upwards, the embedded doubles would be word-swapped.)
789 jint *dp = (jint*)&d;
790 array->append(new ConstantIntValue(dp[1]));
791 array->append(new ConstantIntValue(dp[0]));
792 #endif
793 break;
794 }
795 case Type::Top: // Add an illegal value here
796 array->append(new LocationValue(Location()));
797 break;
798 default:
799 ShouldNotReachHere();
800 break;
801 }
802 }
804 // Determine if this node starts a bundle
805 bool Compile::starts_bundle(const Node *n) const {
806 return (_node_bundling_limit > n->_idx &&
807 _node_bundling_base[n->_idx].starts_bundle());
808 }
810 //--------------------------Process_OopMap_Node--------------------------------
811 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
813 // Handle special safepoint nodes for synchronization
814 MachSafePointNode *sfn = mach->as_MachSafePoint();
815 MachCallNode *mcall;
817 #ifdef ENABLE_ZAP_DEAD_LOCALS
818 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
819 #endif
821 int safepoint_pc_offset = current_offset;
822 bool is_method_handle_invoke = false;
823 bool return_oop = false;
825 // Add the safepoint in the DebugInfoRecorder
826 if( !mach->is_MachCall() ) {
827 mcall = NULL;
828 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
829 } else {
830 mcall = mach->as_MachCall();
832 // Is the call a MethodHandle call?
833 if (mcall->is_MachCallJava()) {
834 if (mcall->as_MachCallJava()->_method_handle_invoke) {
835 assert(has_method_handle_invokes(), "must have been set during call generation");
836 is_method_handle_invoke = true;
837 }
838 }
840 // Check if a call returns an object.
841 if (mcall->return_value_is_used() &&
842 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
843 return_oop = true;
844 }
845 safepoint_pc_offset += mcall->ret_addr_offset();
846 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
847 }
849 // Loop over the JVMState list to add scope information
850 // Do not skip safepoints with a NULL method, they need monitor info
851 JVMState* youngest_jvms = sfn->jvms();
852 int max_depth = youngest_jvms->depth();
854 // Allocate the object pool for scalar-replaced objects -- the map from
855 // small-integer keys (which can be recorded in the local and ostack
856 // arrays) to descriptions of the object state.
857 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
859 // Visit scopes from oldest to youngest.
860 for (int depth = 1; depth <= max_depth; depth++) {
861 JVMState* jvms = youngest_jvms->of_depth(depth);
862 int idx;
863 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
864 // Safepoints that do not have method() set only provide oop-map and monitor info
865 // to support GC; these do not support deoptimization.
866 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
867 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
868 int num_mon = jvms->nof_monitors();
869 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
870 "JVMS local count must match that of the method");
872 // Add Local and Expression Stack Information
874 // Insert locals into the locarray
875 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
876 for( idx = 0; idx < num_locs; idx++ ) {
877 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
878 }
880 // Insert expression stack entries into the exparray
881 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
882 for( idx = 0; idx < num_exps; idx++ ) {
883 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
884 }
886 // Add in mappings of the monitors
887 assert( !method ||
888 !method->is_synchronized() ||
889 method->is_native() ||
890 num_mon > 0 ||
891 !GenerateSynchronizationCode,
892 "monitors must always exist for synchronized methods");
894 // Build the growable array of ScopeValues for exp stack
895 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
897 // Loop over monitors and insert into array
898 for(idx = 0; idx < num_mon; idx++) {
899 // Grab the node that defines this monitor
900 Node* box_node = sfn->monitor_box(jvms, idx);
901 Node* obj_node = sfn->monitor_obj(jvms, idx);
903 // Create ScopeValue for object
904 ScopeValue *scval = NULL;
906 if( obj_node->is_SafePointScalarObject() ) {
907 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
908 scval = Compile::sv_for_node_id(objs, spobj->_idx);
909 if (scval == NULL) {
910 const Type *t = obj_node->bottom_type();
911 ciKlass* cik = t->is_oopptr()->klass();
912 assert(cik->is_instance_klass() ||
913 cik->is_array_klass(), "Not supported allocation.");
914 ObjectValue* sv = new ObjectValue(spobj->_idx,
915 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
916 Compile::set_sv_for_object_node(objs, sv);
918 uint first_ind = spobj->first_index();
919 for (uint i = 0; i < spobj->n_fields(); i++) {
920 Node* fld_node = sfn->in(first_ind+i);
921 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
922 }
923 scval = sv;
924 }
925 } else if( !obj_node->is_Con() ) {
926 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
927 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
928 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
929 } else {
930 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
931 }
932 } else {
933 const TypePtr *tp = obj_node->bottom_type()->make_ptr();
934 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
935 }
937 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
938 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
939 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
940 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
941 }
943 // We dump the object pool first, since deoptimization reads it in first.
944 debug_info()->dump_object_pool(objs);
946 // Build first class objects to pass to scope
947 DebugToken *locvals = debug_info()->create_scope_values(locarray);
948 DebugToken *expvals = debug_info()->create_scope_values(exparray);
949 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
951 // Make method available for all Safepoints
952 ciMethod* scope_method = method ? method : _method;
953 // Describe the scope here
954 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
955 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
956 // Now we can describe the scope.
957 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
958 } // End jvms loop
960 // Mark the end of the scope set.
961 debug_info()->end_safepoint(safepoint_pc_offset);
962 }
966 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
967 class NonSafepointEmitter {
968 Compile* C;
969 JVMState* _pending_jvms;
970 int _pending_offset;
972 void emit_non_safepoint();
974 public:
975 NonSafepointEmitter(Compile* compile) {
976 this->C = compile;
977 _pending_jvms = NULL;
978 _pending_offset = 0;
979 }
981 void observe_instruction(Node* n, int pc_offset) {
982 if (!C->debug_info()->recording_non_safepoints()) return;
984 Node_Notes* nn = C->node_notes_at(n->_idx);
985 if (nn == NULL || nn->jvms() == NULL) return;
986 if (_pending_jvms != NULL &&
987 _pending_jvms->same_calls_as(nn->jvms())) {
988 // Repeated JVMS? Stretch it up here.
989 _pending_offset = pc_offset;
990 } else {
991 if (_pending_jvms != NULL &&
992 _pending_offset < pc_offset) {
993 emit_non_safepoint();
994 }
995 _pending_jvms = NULL;
996 if (pc_offset > C->debug_info()->last_pc_offset()) {
997 // This is the only way _pending_jvms can become non-NULL:
998 _pending_jvms = nn->jvms();
999 _pending_offset = pc_offset;
1000 }
1001 }
1002 }
1004 // Stay out of the way of real safepoints:
1005 void observe_safepoint(JVMState* jvms, int pc_offset) {
1006 if (_pending_jvms != NULL &&
1007 !_pending_jvms->same_calls_as(jvms) &&
1008 _pending_offset < pc_offset) {
1009 emit_non_safepoint();
1010 }
1011 _pending_jvms = NULL;
1012 }
1014 void flush_at_end() {
1015 if (_pending_jvms != NULL) {
1016 emit_non_safepoint();
1017 }
1018 _pending_jvms = NULL;
1019 }
1020 };
1022 void NonSafepointEmitter::emit_non_safepoint() {
1023 JVMState* youngest_jvms = _pending_jvms;
1024 int pc_offset = _pending_offset;
1026 // Clear it now:
1027 _pending_jvms = NULL;
1029 DebugInformationRecorder* debug_info = C->debug_info();
1030 assert(debug_info->recording_non_safepoints(), "sanity");
1032 debug_info->add_non_safepoint(pc_offset);
1033 int max_depth = youngest_jvms->depth();
1035 // Visit scopes from oldest to youngest.
1036 for (int depth = 1; depth <= max_depth; depth++) {
1037 JVMState* jvms = youngest_jvms->of_depth(depth);
1038 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1039 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1040 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1041 }
1043 // Mark the end of the scope set.
1044 debug_info->end_non_safepoint(pc_offset);
1045 }
1047 //------------------------------init_buffer------------------------------------
1048 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1050 // Set the initially allocated size
1051 int code_req = initial_code_capacity;
1052 int locs_req = initial_locs_capacity;
1053 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1054 int const_req = initial_const_capacity;
1056 int pad_req = NativeCall::instruction_size;
1057 // The extra spacing after the code is necessary on some platforms.
1058 // Sometimes we need to patch in a jump after the last instruction,
1059 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1061 // Compute the byte offset where we can store the deopt pc.
1062 if (fixed_slots() != 0) {
1063 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1064 }
1066 // Compute prolog code size
1067 _method_size = 0;
1068 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1069 #ifdef IA64
1070 if (save_argument_registers()) {
1071 // 4815101: this is a stub with implicit and unknown precision fp args.
1072 // The usual spill mechanism can only generate stfd's in this case, which
1073 // doesn't work if the fp reg to spill contains a single-precision denorm.
1074 // Instead, we hack around the normal spill mechanism using stfspill's and
1075 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1076 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1077 //
1078 // If we ever implement 16-byte 'registers' == stack slots, we can
1079 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1080 // instead of stfd/stfs/ldfd/ldfs.
1081 _frame_slots += 8*(16/BytesPerInt);
1082 }
1083 #endif
1084 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1086 if (has_mach_constant_base_node()) {
1087 // Fill the constant table.
1088 // Note: This must happen before shorten_branches.
1089 for (uint i = 0; i < _cfg->_num_blocks; i++) {
1090 Block* b = _cfg->_blocks[i];
1092 for (uint j = 0; j < b->_nodes.size(); j++) {
1093 Node* n = b->_nodes[j];
1095 // If the node is a MachConstantNode evaluate the constant
1096 // value section.
1097 if (n->is_MachConstant()) {
1098 MachConstantNode* machcon = n->as_MachConstant();
1099 machcon->eval_constant(C);
1100 }
1101 }
1102 }
1104 // Calculate the offsets of the constants and the size of the
1105 // constant table (including the padding to the next section).
1106 constant_table().calculate_offsets_and_size();
1107 const_req = constant_table().size();
1108 }
1110 // Initialize the space for the BufferBlob used to find and verify
1111 // instruction size in MachNode::emit_size()
1112 init_scratch_buffer_blob(const_req);
1113 if (failing()) return NULL; // Out of memory
1115 // Pre-compute the length of blocks and replace
1116 // long branches with short if machine supports it.
1117 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1119 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1120 int exception_handler_req = size_exception_handler();
1121 int deopt_handler_req = size_deopt_handler();
1122 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1123 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1124 stub_req += MAX_stubs_size; // ensure per-stub margin
1125 code_req += MAX_inst_size; // ensure per-instruction margin
1127 if (StressCodeBuffers)
1128 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1130 int total_req =
1131 const_req +
1132 code_req +
1133 pad_req +
1134 stub_req +
1135 exception_handler_req +
1136 deopt_handler_req; // deopt handler
1138 if (has_method_handle_invokes())
1139 total_req += deopt_handler_req; // deopt MH handler
1141 CodeBuffer* cb = code_buffer();
1142 cb->initialize(total_req, locs_req);
1144 // Have we run out of code space?
1145 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1146 C->record_failure("CodeCache is full");
1147 return NULL;
1148 }
1149 // Configure the code buffer.
1150 cb->initialize_consts_size(const_req);
1151 cb->initialize_stubs_size(stub_req);
1152 cb->initialize_oop_recorder(env()->oop_recorder());
1154 // fill in the nop array for bundling computations
1155 MachNode *_nop_list[Bundle::_nop_count];
1156 Bundle::initialize_nops(_nop_list, this);
1158 return cb;
1159 }
1161 //------------------------------fill_buffer------------------------------------
1162 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1163 // blk_starts[] contains offsets calculated during short branches processing,
1164 // offsets should not be increased during following steps.
1166 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1167 // of a loop. It is used to determine the padding for loop alignment.
1168 compute_loop_first_inst_sizes();
1170 // Create oopmap set.
1171 _oop_map_set = new OopMapSet();
1173 // !!!!! This preserves old handling of oopmaps for now
1174 debug_info()->set_oopmaps(_oop_map_set);
1176 uint nblocks = _cfg->_num_blocks;
1177 // Count and start of implicit null check instructions
1178 uint inct_cnt = 0;
1179 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1181 // Count and start of calls
1182 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1184 uint return_offset = 0;
1185 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1187 int previous_offset = 0;
1188 int current_offset = 0;
1189 int last_call_offset = -1;
1190 int last_avoid_back_to_back_offset = -1;
1191 #ifdef ASSERT
1192 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1193 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1194 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1195 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1196 #endif
1198 // Create an array of unused labels, one for each basic block, if printing is enabled
1199 #ifndef PRODUCT
1200 int *node_offsets = NULL;
1201 uint node_offset_limit = unique();
1203 if (print_assembly())
1204 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1205 #endif
1207 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1209 // Emit the constant table.
1210 if (has_mach_constant_base_node()) {
1211 constant_table().emit(*cb);
1212 }
1214 // Create an array of labels, one for each basic block
1215 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1216 for (uint i=0; i <= nblocks; i++) {
1217 blk_labels[i].init();
1218 }
1220 // ------------------
1221 // Now fill in the code buffer
1222 Node *delay_slot = NULL;
1224 for (uint i=0; i < nblocks; i++) {
1225 Block *b = _cfg->_blocks[i];
1227 Node *head = b->head();
1229 // If this block needs to start aligned (i.e, can be reached other
1230 // than by falling-thru from the previous block), then force the
1231 // start of a new bundle.
1232 if (Pipeline::requires_bundling() && starts_bundle(head))
1233 cb->flush_bundle(true);
1235 #ifdef ASSERT
1236 if (!b->is_connector()) {
1237 stringStream st;
1238 b->dump_head(&_cfg->_bbs, &st);
1239 MacroAssembler(cb).block_comment(st.as_string());
1240 }
1241 jmp_target[i] = 0;
1242 jmp_offset[i] = 0;
1243 jmp_size[i] = 0;
1244 jmp_rule[i] = 0;
1245 #endif
1246 int blk_offset = current_offset;
1248 // Define the label at the beginning of the basic block
1249 MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
1251 uint last_inst = b->_nodes.size();
1253 // Emit block normally, except for last instruction.
1254 // Emit means "dump code bits into code buffer".
1255 for (uint j = 0; j<last_inst; j++) {
1257 // Get the node
1258 Node* n = b->_nodes[j];
1260 // See if delay slots are supported
1261 if (valid_bundle_info(n) &&
1262 node_bundling(n)->used_in_unconditional_delay()) {
1263 assert(delay_slot == NULL, "no use of delay slot node");
1264 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1266 delay_slot = n;
1267 continue;
1268 }
1270 // If this starts a new instruction group, then flush the current one
1271 // (but allow split bundles)
1272 if (Pipeline::requires_bundling() && starts_bundle(n))
1273 cb->flush_bundle(false);
1275 // The following logic is duplicated in the code ifdeffed for
1276 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1277 // should be factored out. Or maybe dispersed to the nodes?
1279 // Special handling for SafePoint/Call Nodes
1280 bool is_mcall = false;
1281 if (n->is_Mach()) {
1282 MachNode *mach = n->as_Mach();
1283 is_mcall = n->is_MachCall();
1284 bool is_sfn = n->is_MachSafePoint();
1286 // If this requires all previous instructions be flushed, then do so
1287 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1288 cb->flush_bundle(true);
1289 current_offset = cb->insts_size();
1290 }
1292 // A padding may be needed again since a previous instruction
1293 // could be moved to delay slot.
1295 // align the instruction if necessary
1296 int padding = mach->compute_padding(current_offset);
1297 // Make sure safepoint node for polling is distinct from a call's
1298 // return by adding a nop if needed.
1299 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1300 padding = nop_size;
1301 }
1302 if (padding == 0 && mach->avoid_back_to_back() &&
1303 current_offset == last_avoid_back_to_back_offset) {
1304 // Avoid back to back some instructions.
1305 padding = nop_size;
1306 }
1308 if(padding > 0) {
1309 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1310 int nops_cnt = padding / nop_size;
1311 MachNode *nop = new (this) MachNopNode(nops_cnt);
1312 b->_nodes.insert(j++, nop);
1313 last_inst++;
1314 _cfg->_bbs.map( nop->_idx, b );
1315 nop->emit(*cb, _regalloc);
1316 cb->flush_bundle(true);
1317 current_offset = cb->insts_size();
1318 }
1320 // Remember the start of the last call in a basic block
1321 if (is_mcall) {
1322 MachCallNode *mcall = mach->as_MachCall();
1324 // This destination address is NOT PC-relative
1325 mcall->method_set((intptr_t)mcall->entry_point());
1327 // Save the return address
1328 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset();
1330 if (mcall->is_MachCallLeaf()) {
1331 is_mcall = false;
1332 is_sfn = false;
1333 }
1334 }
1336 // sfn will be valid whenever mcall is valid now because of inheritance
1337 if (is_sfn || is_mcall) {
1339 // Handle special safepoint nodes for synchronization
1340 if (!is_mcall) {
1341 MachSafePointNode *sfn = mach->as_MachSafePoint();
1342 // !!!!! Stubs only need an oopmap right now, so bail out
1343 if (sfn->jvms()->method() == NULL) {
1344 // Write the oopmap directly to the code blob??!!
1345 # ifdef ENABLE_ZAP_DEAD_LOCALS
1346 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1347 # endif
1348 continue;
1349 }
1350 } // End synchronization
1352 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1353 current_offset);
1354 Process_OopMap_Node(mach, current_offset);
1355 } // End if safepoint
1357 // If this is a null check, then add the start of the previous instruction to the list
1358 else if( mach->is_MachNullCheck() ) {
1359 inct_starts[inct_cnt++] = previous_offset;
1360 }
1362 // If this is a branch, then fill in the label with the target BB's label
1363 else if (mach->is_MachBranch()) {
1364 // This requires the TRUE branch target be in succs[0]
1365 uint block_num = b->non_connector_successor(0)->_pre_order;
1367 // Try to replace long branch if delay slot is not used,
1368 // it is mostly for back branches since forward branch's
1369 // distance is not updated yet.
1370 bool delay_slot_is_used = valid_bundle_info(n) &&
1371 node_bundling(n)->use_unconditional_delay();
1372 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1373 assert(delay_slot == NULL, "not expecting delay slot node");
1374 int br_size = n->size(_regalloc);
1375 int offset = blk_starts[block_num] - current_offset;
1376 if (block_num >= i) {
1377 // Current and following block's offset are not
1378 // finilized yet, adjust distance by the difference
1379 // between calculated and final offsets of current block.
1380 offset -= (blk_starts[i] - blk_offset);
1381 }
1382 // In the following code a nop could be inserted before
1383 // the branch which will increase the backward distance.
1384 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1385 if (needs_padding && offset <= 0)
1386 offset -= nop_size;
1388 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1389 // We've got a winner. Replace this branch.
1390 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1392 // Update the jmp_size.
1393 int new_size = replacement->size(_regalloc);
1394 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1395 // Insert padding between avoid_back_to_back branches.
1396 if (needs_padding && replacement->avoid_back_to_back()) {
1397 MachNode *nop = new (this) MachNopNode();
1398 b->_nodes.insert(j++, nop);
1399 _cfg->_bbs.map(nop->_idx, b);
1400 last_inst++;
1401 nop->emit(*cb, _regalloc);
1402 cb->flush_bundle(true);
1403 current_offset = cb->insts_size();
1404 }
1405 #ifdef ASSERT
1406 jmp_target[i] = block_num;
1407 jmp_offset[i] = current_offset - blk_offset;
1408 jmp_size[i] = new_size;
1409 jmp_rule[i] = mach->rule();
1410 #endif
1411 b->_nodes.map(j, replacement);
1412 mach->subsume_by(replacement, C);
1413 n = replacement;
1414 mach = replacement;
1415 }
1416 }
1417 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1418 } else if (mach->ideal_Opcode() == Op_Jump) {
1419 for (uint h = 0; h < b->_num_succs; h++) {
1420 Block* succs_block = b->_succs[h];
1421 for (uint j = 1; j < succs_block->num_preds(); j++) {
1422 Node* jpn = succs_block->pred(j);
1423 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1424 uint block_num = succs_block->non_connector()->_pre_order;
1425 Label *blkLabel = &blk_labels[block_num];
1426 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1427 }
1428 }
1429 }
1430 }
1432 #ifdef ASSERT
1433 // Check that oop-store precedes the card-mark
1434 else if (mach->ideal_Opcode() == Op_StoreCM) {
1435 uint storeCM_idx = j;
1436 int count = 0;
1437 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1438 Node *oop_store = mach->in(prec); // Precedence edge
1439 if (oop_store == NULL) continue;
1440 count++;
1441 uint i4;
1442 for( i4 = 0; i4 < last_inst; ++i4 ) {
1443 if( b->_nodes[i4] == oop_store ) break;
1444 }
1445 // Note: This test can provide a false failure if other precedence
1446 // edges have been added to the storeCMNode.
1447 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1448 }
1449 assert(count > 0, "storeCM expects at least one precedence edge");
1450 }
1451 #endif
1453 else if (!n->is_Proj()) {
1454 // Remember the beginning of the previous instruction, in case
1455 // it's followed by a flag-kill and a null-check. Happens on
1456 // Intel all the time, with add-to-memory kind of opcodes.
1457 previous_offset = current_offset;
1458 }
1459 }
1461 // Verify that there is sufficient space remaining
1462 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1463 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1464 C->record_failure("CodeCache is full");
1465 return;
1466 }
1468 // Save the offset for the listing
1469 #ifndef PRODUCT
1470 if (node_offsets && n->_idx < node_offset_limit)
1471 node_offsets[n->_idx] = cb->insts_size();
1472 #endif
1474 // "Normal" instruction case
1475 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1476 n->emit(*cb, _regalloc);
1477 current_offset = cb->insts_size();
1479 #ifdef ASSERT
1480 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1481 n->dump();
1482 assert(false, "wrong size of mach node");
1483 }
1484 #endif
1485 non_safepoints.observe_instruction(n, current_offset);
1487 // mcall is last "call" that can be a safepoint
1488 // record it so we can see if a poll will directly follow it
1489 // in which case we'll need a pad to make the PcDesc sites unique
1490 // see 5010568. This can be slightly inaccurate but conservative
1491 // in the case that return address is not actually at current_offset.
1492 // This is a small price to pay.
1494 if (is_mcall) {
1495 last_call_offset = current_offset;
1496 }
1498 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
1499 // Avoid back to back some instructions.
1500 last_avoid_back_to_back_offset = current_offset;
1501 }
1503 // See if this instruction has a delay slot
1504 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1505 assert(delay_slot != NULL, "expecting delay slot node");
1507 // Back up 1 instruction
1508 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1510 // Save the offset for the listing
1511 #ifndef PRODUCT
1512 if (node_offsets && delay_slot->_idx < node_offset_limit)
1513 node_offsets[delay_slot->_idx] = cb->insts_size();
1514 #endif
1516 // Support a SafePoint in the delay slot
1517 if (delay_slot->is_MachSafePoint()) {
1518 MachNode *mach = delay_slot->as_Mach();
1519 // !!!!! Stubs only need an oopmap right now, so bail out
1520 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1521 // Write the oopmap directly to the code blob??!!
1522 # ifdef ENABLE_ZAP_DEAD_LOCALS
1523 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1524 # endif
1525 delay_slot = NULL;
1526 continue;
1527 }
1529 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1530 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1531 adjusted_offset);
1532 // Generate an OopMap entry
1533 Process_OopMap_Node(mach, adjusted_offset);
1534 }
1536 // Insert the delay slot instruction
1537 delay_slot->emit(*cb, _regalloc);
1539 // Don't reuse it
1540 delay_slot = NULL;
1541 }
1543 } // End for all instructions in block
1545 // If the next block is the top of a loop, pad this block out to align
1546 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1547 if (i < nblocks-1) {
1548 Block *nb = _cfg->_blocks[i+1];
1549 int padding = nb->alignment_padding(current_offset);
1550 if( padding > 0 ) {
1551 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1552 b->_nodes.insert( b->_nodes.size(), nop );
1553 _cfg->_bbs.map( nop->_idx, b );
1554 nop->emit(*cb, _regalloc);
1555 current_offset = cb->insts_size();
1556 }
1557 }
1558 // Verify that the distance for generated before forward
1559 // short branches is still valid.
1560 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1562 // Save new block start offset
1563 blk_starts[i] = blk_offset;
1564 } // End of for all blocks
1565 blk_starts[nblocks] = current_offset;
1567 non_safepoints.flush_at_end();
1569 // Offset too large?
1570 if (failing()) return;
1572 // Define a pseudo-label at the end of the code
1573 MacroAssembler(cb).bind( blk_labels[nblocks] );
1575 // Compute the size of the first block
1576 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1578 assert(cb->insts_size() < 500000, "method is unreasonably large");
1580 #ifdef ASSERT
1581 for (uint i = 0; i < nblocks; i++) { // For all blocks
1582 if (jmp_target[i] != 0) {
1583 int br_size = jmp_size[i];
1584 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1585 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1586 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
1587 assert(false, "Displacement too large for short jmp");
1588 }
1589 }
1590 }
1591 #endif
1593 // ------------------
1595 #ifndef PRODUCT
1596 // Information on the size of the method, without the extraneous code
1597 Scheduling::increment_method_size(cb->insts_size());
1598 #endif
1600 // ------------------
1601 // Fill in exception table entries.
1602 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1604 // Only java methods have exception handlers and deopt handlers
1605 if (_method) {
1606 // Emit the exception handler code.
1607 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1608 // Emit the deopt handler code.
1609 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1611 // Emit the MethodHandle deopt handler code (if required).
1612 if (has_method_handle_invokes()) {
1613 // We can use the same code as for the normal deopt handler, we
1614 // just need a different entry point address.
1615 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1616 }
1617 }
1619 // One last check for failed CodeBuffer::expand:
1620 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1621 C->record_failure("CodeCache is full");
1622 return;
1623 }
1625 #ifndef PRODUCT
1626 // Dump the assembly code, including basic-block numbers
1627 if (print_assembly()) {
1628 ttyLocker ttyl; // keep the following output all in one block
1629 if (!VMThread::should_terminate()) { // test this under the tty lock
1630 // This output goes directly to the tty, not the compiler log.
1631 // To enable tools to match it up with the compilation activity,
1632 // be sure to tag this tty output with the compile ID.
1633 if (xtty != NULL) {
1634 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1635 is_osr_compilation() ? " compile_kind='osr'" :
1636 "");
1637 }
1638 if (method() != NULL) {
1639 method()->print_metadata();
1640 }
1641 dump_asm(node_offsets, node_offset_limit);
1642 if (xtty != NULL) {
1643 xtty->tail("opto_assembly");
1644 }
1645 }
1646 }
1647 #endif
1649 }
1651 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1652 _inc_table.set_size(cnt);
1654 uint inct_cnt = 0;
1655 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1656 Block *b = _cfg->_blocks[i];
1657 Node *n = NULL;
1658 int j;
1660 // Find the branch; ignore trailing NOPs.
1661 for( j = b->_nodes.size()-1; j>=0; j-- ) {
1662 n = b->_nodes[j];
1663 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con )
1664 break;
1665 }
1667 // If we didn't find anything, continue
1668 if( j < 0 ) continue;
1670 // Compute ExceptionHandlerTable subtable entry and add it
1671 // (skip empty blocks)
1672 if( n->is_Catch() ) {
1674 // Get the offset of the return from the call
1675 uint call_return = call_returns[b->_pre_order];
1676 #ifdef ASSERT
1677 assert( call_return > 0, "no call seen for this basic block" );
1678 while( b->_nodes[--j]->is_MachProj() ) ;
1679 assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" );
1680 #endif
1681 // last instruction is a CatchNode, find it's CatchProjNodes
1682 int nof_succs = b->_num_succs;
1683 // allocate space
1684 GrowableArray<intptr_t> handler_bcis(nof_succs);
1685 GrowableArray<intptr_t> handler_pcos(nof_succs);
1686 // iterate through all successors
1687 for (int j = 0; j < nof_succs; j++) {
1688 Block* s = b->_succs[j];
1689 bool found_p = false;
1690 for( uint k = 1; k < s->num_preds(); k++ ) {
1691 Node *pk = s->pred(k);
1692 if( pk->is_CatchProj() && pk->in(0) == n ) {
1693 const CatchProjNode* p = pk->as_CatchProj();
1694 found_p = true;
1695 // add the corresponding handler bci & pco information
1696 if( p->_con != CatchProjNode::fall_through_index ) {
1697 // p leads to an exception handler (and is not fall through)
1698 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering");
1699 // no duplicates, please
1700 if( !handler_bcis.contains(p->handler_bci()) ) {
1701 uint block_num = s->non_connector()->_pre_order;
1702 handler_bcis.append(p->handler_bci());
1703 handler_pcos.append(blk_labels[block_num].loc_pos());
1704 }
1705 }
1706 }
1707 }
1708 assert(found_p, "no matching predecessor found");
1709 // Note: Due to empty block removal, one block may have
1710 // several CatchProj inputs, from the same Catch.
1711 }
1713 // Set the offset of the return from the call
1714 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1715 continue;
1716 }
1718 // Handle implicit null exception table updates
1719 if( n->is_MachNullCheck() ) {
1720 uint block_num = b->non_connector_successor(0)->_pre_order;
1721 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() );
1722 continue;
1723 }
1724 } // End of for all blocks fill in exception table entries
1725 }
1727 // Static Variables
1728 #ifndef PRODUCT
1729 uint Scheduling::_total_nop_size = 0;
1730 uint Scheduling::_total_method_size = 0;
1731 uint Scheduling::_total_branches = 0;
1732 uint Scheduling::_total_unconditional_delays = 0;
1733 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1734 #endif
1736 // Initializer for class Scheduling
1738 Scheduling::Scheduling(Arena *arena, Compile &compile)
1739 : _arena(arena),
1740 _cfg(compile.cfg()),
1741 _bbs(compile.cfg()->_bbs),
1742 _regalloc(compile.regalloc()),
1743 _reg_node(arena),
1744 _bundle_instr_count(0),
1745 _bundle_cycle_number(0),
1746 _scheduled(arena),
1747 _available(arena),
1748 _next_node(NULL),
1749 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1750 _pinch_free_list(arena)
1751 #ifndef PRODUCT
1752 , _branches(0)
1753 , _unconditional_delays(0)
1754 #endif
1755 {
1756 // Create a MachNopNode
1757 _nop = new (&compile) MachNopNode();
1759 // Now that the nops are in the array, save the count
1760 // (but allow entries for the nops)
1761 _node_bundling_limit = compile.unique();
1762 uint node_max = _regalloc->node_regs_max_index();
1764 compile.set_node_bundling_limit(_node_bundling_limit);
1766 // This one is persistent within the Compile class
1767 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1769 // Allocate space for fixed-size arrays
1770 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1771 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1772 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1774 // Clear the arrays
1775 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1776 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1777 memset(_uses, 0, node_max * sizeof(short));
1778 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1780 // Clear the bundling information
1781 memcpy(_bundle_use_elements,
1782 Pipeline_Use::elaborated_elements,
1783 sizeof(Pipeline_Use::elaborated_elements));
1785 // Get the last node
1786 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1];
1788 _next_node = bb->_nodes[bb->_nodes.size()-1];
1789 }
1791 #ifndef PRODUCT
1792 // Scheduling destructor
1793 Scheduling::~Scheduling() {
1794 _total_branches += _branches;
1795 _total_unconditional_delays += _unconditional_delays;
1796 }
1797 #endif
1799 // Step ahead "i" cycles
1800 void Scheduling::step(uint i) {
1802 Bundle *bundle = node_bundling(_next_node);
1803 bundle->set_starts_bundle();
1805 // Update the bundle record, but leave the flags information alone
1806 if (_bundle_instr_count > 0) {
1807 bundle->set_instr_count(_bundle_instr_count);
1808 bundle->set_resources_used(_bundle_use.resourcesUsed());
1809 }
1811 // Update the state information
1812 _bundle_instr_count = 0;
1813 _bundle_cycle_number += i;
1814 _bundle_use.step(i);
1815 }
1817 void Scheduling::step_and_clear() {
1818 Bundle *bundle = node_bundling(_next_node);
1819 bundle->set_starts_bundle();
1821 // Update the bundle record
1822 if (_bundle_instr_count > 0) {
1823 bundle->set_instr_count(_bundle_instr_count);
1824 bundle->set_resources_used(_bundle_use.resourcesUsed());
1826 _bundle_cycle_number += 1;
1827 }
1829 // Clear the bundling information
1830 _bundle_instr_count = 0;
1831 _bundle_use.reset();
1833 memcpy(_bundle_use_elements,
1834 Pipeline_Use::elaborated_elements,
1835 sizeof(Pipeline_Use::elaborated_elements));
1836 }
1838 //------------------------------ScheduleAndBundle------------------------------
1839 // Perform instruction scheduling and bundling over the sequence of
1840 // instructions in backwards order.
1841 void Compile::ScheduleAndBundle() {
1843 // Don't optimize this if it isn't a method
1844 if (!_method)
1845 return;
1847 // Don't optimize this if scheduling is disabled
1848 if (!do_scheduling())
1849 return;
1851 // Scheduling code works only with pairs (8 bytes) maximum.
1852 if (max_vector_size() > 8)
1853 return;
1855 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1857 // Create a data structure for all the scheduling information
1858 Scheduling scheduling(Thread::current()->resource_area(), *this);
1860 // Walk backwards over each basic block, computing the needed alignment
1861 // Walk over all the basic blocks
1862 scheduling.DoScheduling();
1863 }
1865 //------------------------------ComputeLocalLatenciesForward-------------------
1866 // Compute the latency of all the instructions. This is fairly simple,
1867 // because we already have a legal ordering. Walk over the instructions
1868 // from first to last, and compute the latency of the instruction based
1869 // on the latency of the preceding instruction(s).
1870 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1871 #ifndef PRODUCT
1872 if (_cfg->C->trace_opto_output())
1873 tty->print("# -> ComputeLocalLatenciesForward\n");
1874 #endif
1876 // Walk over all the schedulable instructions
1877 for( uint j=_bb_start; j < _bb_end; j++ ) {
1879 // This is a kludge, forcing all latency calculations to start at 1.
1880 // Used to allow latency 0 to force an instruction to the beginning
1881 // of the bb
1882 uint latency = 1;
1883 Node *use = bb->_nodes[j];
1884 uint nlen = use->len();
1886 // Walk over all the inputs
1887 for ( uint k=0; k < nlen; k++ ) {
1888 Node *def = use->in(k);
1889 if (!def)
1890 continue;
1892 uint l = _node_latency[def->_idx] + use->latency(k);
1893 if (latency < l)
1894 latency = l;
1895 }
1897 _node_latency[use->_idx] = latency;
1899 #ifndef PRODUCT
1900 if (_cfg->C->trace_opto_output()) {
1901 tty->print("# latency %4d: ", latency);
1902 use->dump();
1903 }
1904 #endif
1905 }
1907 #ifndef PRODUCT
1908 if (_cfg->C->trace_opto_output())
1909 tty->print("# <- ComputeLocalLatenciesForward\n");
1910 #endif
1912 } // end ComputeLocalLatenciesForward
1914 // See if this node fits into the present instruction bundle
1915 bool Scheduling::NodeFitsInBundle(Node *n) {
1916 uint n_idx = n->_idx;
1918 // If this is the unconditional delay instruction, then it fits
1919 if (n == _unconditional_delay_slot) {
1920 #ifndef PRODUCT
1921 if (_cfg->C->trace_opto_output())
1922 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1923 #endif
1924 return (true);
1925 }
1927 // If the node cannot be scheduled this cycle, skip it
1928 if (_current_latency[n_idx] > _bundle_cycle_number) {
1929 #ifndef PRODUCT
1930 if (_cfg->C->trace_opto_output())
1931 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1932 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1933 #endif
1934 return (false);
1935 }
1937 const Pipeline *node_pipeline = n->pipeline();
1939 uint instruction_count = node_pipeline->instructionCount();
1940 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1941 instruction_count = 0;
1942 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1943 instruction_count++;
1945 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1946 #ifndef PRODUCT
1947 if (_cfg->C->trace_opto_output())
1948 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1949 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1950 #endif
1951 return (false);
1952 }
1954 // Don't allow non-machine nodes to be handled this way
1955 if (!n->is_Mach() && instruction_count == 0)
1956 return (false);
1958 // See if there is any overlap
1959 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1961 if (delay > 0) {
1962 #ifndef PRODUCT
1963 if (_cfg->C->trace_opto_output())
1964 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1965 #endif
1966 return false;
1967 }
1969 #ifndef PRODUCT
1970 if (_cfg->C->trace_opto_output())
1971 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1972 #endif
1974 return true;
1975 }
1977 Node * Scheduling::ChooseNodeToBundle() {
1978 uint siz = _available.size();
1980 if (siz == 0) {
1982 #ifndef PRODUCT
1983 if (_cfg->C->trace_opto_output())
1984 tty->print("# ChooseNodeToBundle: NULL\n");
1985 #endif
1986 return (NULL);
1987 }
1989 // Fast path, if only 1 instruction in the bundle
1990 if (siz == 1) {
1991 #ifndef PRODUCT
1992 if (_cfg->C->trace_opto_output()) {
1993 tty->print("# ChooseNodeToBundle (only 1): ");
1994 _available[0]->dump();
1995 }
1996 #endif
1997 return (_available[0]);
1998 }
2000 // Don't bother, if the bundle is already full
2001 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2002 for ( uint i = 0; i < siz; i++ ) {
2003 Node *n = _available[i];
2005 // Skip projections, we'll handle them another way
2006 if (n->is_Proj())
2007 continue;
2009 // This presupposed that instructions are inserted into the
2010 // available list in a legality order; i.e. instructions that
2011 // must be inserted first are at the head of the list
2012 if (NodeFitsInBundle(n)) {
2013 #ifndef PRODUCT
2014 if (_cfg->C->trace_opto_output()) {
2015 tty->print("# ChooseNodeToBundle: ");
2016 n->dump();
2017 }
2018 #endif
2019 return (n);
2020 }
2021 }
2022 }
2024 // Nothing fits in this bundle, choose the highest priority
2025 #ifndef PRODUCT
2026 if (_cfg->C->trace_opto_output()) {
2027 tty->print("# ChooseNodeToBundle: ");
2028 _available[0]->dump();
2029 }
2030 #endif
2032 return _available[0];
2033 }
2035 //------------------------------AddNodeToAvailableList-------------------------
2036 void Scheduling::AddNodeToAvailableList(Node *n) {
2037 assert( !n->is_Proj(), "projections never directly made available" );
2038 #ifndef PRODUCT
2039 if (_cfg->C->trace_opto_output()) {
2040 tty->print("# AddNodeToAvailableList: ");
2041 n->dump();
2042 }
2043 #endif
2045 int latency = _current_latency[n->_idx];
2047 // Insert in latency order (insertion sort)
2048 uint i;
2049 for ( i=0; i < _available.size(); i++ )
2050 if (_current_latency[_available[i]->_idx] > latency)
2051 break;
2053 // Special Check for compares following branches
2054 if( n->is_Mach() && _scheduled.size() > 0 ) {
2055 int op = n->as_Mach()->ideal_Opcode();
2056 Node *last = _scheduled[0];
2057 if( last->is_MachIf() && last->in(1) == n &&
2058 ( op == Op_CmpI ||
2059 op == Op_CmpU ||
2060 op == Op_CmpP ||
2061 op == Op_CmpF ||
2062 op == Op_CmpD ||
2063 op == Op_CmpL ) ) {
2065 // Recalculate position, moving to front of same latency
2066 for ( i=0 ; i < _available.size(); i++ )
2067 if (_current_latency[_available[i]->_idx] >= latency)
2068 break;
2069 }
2070 }
2072 // Insert the node in the available list
2073 _available.insert(i, n);
2075 #ifndef PRODUCT
2076 if (_cfg->C->trace_opto_output())
2077 dump_available();
2078 #endif
2079 }
2081 //------------------------------DecrementUseCounts-----------------------------
2082 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2083 for ( uint i=0; i < n->len(); i++ ) {
2084 Node *def = n->in(i);
2085 if (!def) continue;
2086 if( def->is_Proj() ) // If this is a machine projection, then
2087 def = def->in(0); // propagate usage thru to the base instruction
2089 if( _bbs[def->_idx] != bb ) // Ignore if not block-local
2090 continue;
2092 // Compute the latency
2093 uint l = _bundle_cycle_number + n->latency(i);
2094 if (_current_latency[def->_idx] < l)
2095 _current_latency[def->_idx] = l;
2097 // If this does not have uses then schedule it
2098 if ((--_uses[def->_idx]) == 0)
2099 AddNodeToAvailableList(def);
2100 }
2101 }
2103 //------------------------------AddNodeToBundle--------------------------------
2104 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2105 #ifndef PRODUCT
2106 if (_cfg->C->trace_opto_output()) {
2107 tty->print("# AddNodeToBundle: ");
2108 n->dump();
2109 }
2110 #endif
2112 // Remove this from the available list
2113 uint i;
2114 for (i = 0; i < _available.size(); i++)
2115 if (_available[i] == n)
2116 break;
2117 assert(i < _available.size(), "entry in _available list not found");
2118 _available.remove(i);
2120 // See if this fits in the current bundle
2121 const Pipeline *node_pipeline = n->pipeline();
2122 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2124 // Check for instructions to be placed in the delay slot. We
2125 // do this before we actually schedule the current instruction,
2126 // because the delay slot follows the current instruction.
2127 if (Pipeline::_branch_has_delay_slot &&
2128 node_pipeline->hasBranchDelay() &&
2129 !_unconditional_delay_slot) {
2131 uint siz = _available.size();
2133 // Conditional branches can support an instruction that
2134 // is unconditionally executed and not dependent by the
2135 // branch, OR a conditionally executed instruction if
2136 // the branch is taken. In practice, this means that
2137 // the first instruction at the branch target is
2138 // copied to the delay slot, and the branch goes to
2139 // the instruction after that at the branch target
2140 if ( n->is_MachBranch() ) {
2142 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2143 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2145 #ifndef PRODUCT
2146 _branches++;
2147 #endif
2149 // At least 1 instruction is on the available list
2150 // that is not dependent on the branch
2151 for (uint i = 0; i < siz; i++) {
2152 Node *d = _available[i];
2153 const Pipeline *avail_pipeline = d->pipeline();
2155 // Don't allow safepoints in the branch shadow, that will
2156 // cause a number of difficulties
2157 if ( avail_pipeline->instructionCount() == 1 &&
2158 !avail_pipeline->hasMultipleBundles() &&
2159 !avail_pipeline->hasBranchDelay() &&
2160 Pipeline::instr_has_unit_size() &&
2161 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2162 NodeFitsInBundle(d) &&
2163 !node_bundling(d)->used_in_delay()) {
2165 if (d->is_Mach() && !d->is_MachSafePoint()) {
2166 // A node that fits in the delay slot was found, so we need to
2167 // set the appropriate bits in the bundle pipeline information so
2168 // that it correctly indicates resource usage. Later, when we
2169 // attempt to add this instruction to the bundle, we will skip
2170 // setting the resource usage.
2171 _unconditional_delay_slot = d;
2172 node_bundling(n)->set_use_unconditional_delay();
2173 node_bundling(d)->set_used_in_unconditional_delay();
2174 _bundle_use.add_usage(avail_pipeline->resourceUse());
2175 _current_latency[d->_idx] = _bundle_cycle_number;
2176 _next_node = d;
2177 ++_bundle_instr_count;
2178 #ifndef PRODUCT
2179 _unconditional_delays++;
2180 #endif
2181 break;
2182 }
2183 }
2184 }
2185 }
2187 // No delay slot, add a nop to the usage
2188 if (!_unconditional_delay_slot) {
2189 // See if adding an instruction in the delay slot will overflow
2190 // the bundle.
2191 if (!NodeFitsInBundle(_nop)) {
2192 #ifndef PRODUCT
2193 if (_cfg->C->trace_opto_output())
2194 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2195 #endif
2196 step(1);
2197 }
2199 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2200 _next_node = _nop;
2201 ++_bundle_instr_count;
2202 }
2204 // See if the instruction in the delay slot requires a
2205 // step of the bundles
2206 if (!NodeFitsInBundle(n)) {
2207 #ifndef PRODUCT
2208 if (_cfg->C->trace_opto_output())
2209 tty->print("# *** STEP(branch won't fit) ***\n");
2210 #endif
2211 // Update the state information
2212 _bundle_instr_count = 0;
2213 _bundle_cycle_number += 1;
2214 _bundle_use.step(1);
2215 }
2216 }
2218 // Get the number of instructions
2219 uint instruction_count = node_pipeline->instructionCount();
2220 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2221 instruction_count = 0;
2223 // Compute the latency information
2224 uint delay = 0;
2226 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2227 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2228 if (relative_latency < 0)
2229 relative_latency = 0;
2231 delay = _bundle_use.full_latency(relative_latency, node_usage);
2233 // Does not fit in this bundle, start a new one
2234 if (delay > 0) {
2235 step(delay);
2237 #ifndef PRODUCT
2238 if (_cfg->C->trace_opto_output())
2239 tty->print("# *** STEP(%d) ***\n", delay);
2240 #endif
2241 }
2242 }
2244 // If this was placed in the delay slot, ignore it
2245 if (n != _unconditional_delay_slot) {
2247 if (delay == 0) {
2248 if (node_pipeline->hasMultipleBundles()) {
2249 #ifndef PRODUCT
2250 if (_cfg->C->trace_opto_output())
2251 tty->print("# *** STEP(multiple instructions) ***\n");
2252 #endif
2253 step(1);
2254 }
2256 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2257 #ifndef PRODUCT
2258 if (_cfg->C->trace_opto_output())
2259 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2260 instruction_count + _bundle_instr_count,
2261 Pipeline::_max_instrs_per_cycle);
2262 #endif
2263 step(1);
2264 }
2265 }
2267 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2268 _bundle_instr_count++;
2270 // Set the node's latency
2271 _current_latency[n->_idx] = _bundle_cycle_number;
2273 // Now merge the functional unit information
2274 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2275 _bundle_use.add_usage(node_usage);
2277 // Increment the number of instructions in this bundle
2278 _bundle_instr_count += instruction_count;
2280 // Remember this node for later
2281 if (n->is_Mach())
2282 _next_node = n;
2283 }
2285 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2286 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2287 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2288 // into the block. All other scheduled nodes get put in the schedule here.
2289 int op = n->Opcode();
2290 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2291 (op != Op_Node && // Not an unused antidepedence node and
2292 // not an unallocated boxlock
2293 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2295 // Push any trailing projections
2296 if( bb->_nodes[bb->_nodes.size()-1] != n ) {
2297 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2298 Node *foi = n->fast_out(i);
2299 if( foi->is_Proj() )
2300 _scheduled.push(foi);
2301 }
2302 }
2304 // Put the instruction in the schedule list
2305 _scheduled.push(n);
2306 }
2308 #ifndef PRODUCT
2309 if (_cfg->C->trace_opto_output())
2310 dump_available();
2311 #endif
2313 // Walk all the definitions, decrementing use counts, and
2314 // if a definition has a 0 use count, place it in the available list.
2315 DecrementUseCounts(n,bb);
2316 }
2318 //------------------------------ComputeUseCount--------------------------------
2319 // This method sets the use count within a basic block. We will ignore all
2320 // uses outside the current basic block. As we are doing a backwards walk,
2321 // any node we reach that has a use count of 0 may be scheduled. This also
2322 // avoids the problem of cyclic references from phi nodes, as long as phi
2323 // nodes are at the front of the basic block. This method also initializes
2324 // the available list to the set of instructions that have no uses within this
2325 // basic block.
2326 void Scheduling::ComputeUseCount(const Block *bb) {
2327 #ifndef PRODUCT
2328 if (_cfg->C->trace_opto_output())
2329 tty->print("# -> ComputeUseCount\n");
2330 #endif
2332 // Clear the list of available and scheduled instructions, just in case
2333 _available.clear();
2334 _scheduled.clear();
2336 // No delay slot specified
2337 _unconditional_delay_slot = NULL;
2339 #ifdef ASSERT
2340 for( uint i=0; i < bb->_nodes.size(); i++ )
2341 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );
2342 #endif
2344 // Force the _uses count to never go to zero for unscheduable pieces
2345 // of the block
2346 for( uint k = 0; k < _bb_start; k++ )
2347 _uses[bb->_nodes[k]->_idx] = 1;
2348 for( uint l = _bb_end; l < bb->_nodes.size(); l++ )
2349 _uses[bb->_nodes[l]->_idx] = 1;
2351 // Iterate backwards over the instructions in the block. Don't count the
2352 // branch projections at end or the block header instructions.
2353 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2354 Node *n = bb->_nodes[j];
2355 if( n->is_Proj() ) continue; // Projections handled another way
2357 // Account for all uses
2358 for ( uint k = 0; k < n->len(); k++ ) {
2359 Node *inp = n->in(k);
2360 if (!inp) continue;
2361 assert(inp != n, "no cycles allowed" );
2362 if( _bbs[inp->_idx] == bb ) { // Block-local use?
2363 if( inp->is_Proj() ) // Skip through Proj's
2364 inp = inp->in(0);
2365 ++_uses[inp->_idx]; // Count 1 block-local use
2366 }
2367 }
2369 // If this instruction has a 0 use count, then it is available
2370 if (!_uses[n->_idx]) {
2371 _current_latency[n->_idx] = _bundle_cycle_number;
2372 AddNodeToAvailableList(n);
2373 }
2375 #ifndef PRODUCT
2376 if (_cfg->C->trace_opto_output()) {
2377 tty->print("# uses: %3d: ", _uses[n->_idx]);
2378 n->dump();
2379 }
2380 #endif
2381 }
2383 #ifndef PRODUCT
2384 if (_cfg->C->trace_opto_output())
2385 tty->print("# <- ComputeUseCount\n");
2386 #endif
2387 }
2389 // This routine performs scheduling on each basic block in reverse order,
2390 // using instruction latencies and taking into account function unit
2391 // availability.
2392 void Scheduling::DoScheduling() {
2393 #ifndef PRODUCT
2394 if (_cfg->C->trace_opto_output())
2395 tty->print("# -> DoScheduling\n");
2396 #endif
2398 Block *succ_bb = NULL;
2399 Block *bb;
2401 // Walk over all the basic blocks in reverse order
2402 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) {
2403 bb = _cfg->_blocks[i];
2405 #ifndef PRODUCT
2406 if (_cfg->C->trace_opto_output()) {
2407 tty->print("# Schedule BB#%03d (initial)\n", i);
2408 for (uint j = 0; j < bb->_nodes.size(); j++)
2409 bb->_nodes[j]->dump();
2410 }
2411 #endif
2413 // On the head node, skip processing
2414 if( bb == _cfg->_broot )
2415 continue;
2417 // Skip empty, connector blocks
2418 if (bb->is_connector())
2419 continue;
2421 // If the following block is not the sole successor of
2422 // this one, then reset the pipeline information
2423 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2424 #ifndef PRODUCT
2425 if (_cfg->C->trace_opto_output()) {
2426 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2427 _next_node->_idx, _bundle_instr_count);
2428 }
2429 #endif
2430 step_and_clear();
2431 }
2433 // Leave untouched the starting instruction, any Phis, a CreateEx node
2434 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction.
2435 _bb_end = bb->_nodes.size()-1;
2436 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2437 Node *n = bb->_nodes[_bb_start];
2438 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2439 // Also, MachIdealNodes do not get scheduled
2440 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2441 MachNode *mach = n->as_Mach();
2442 int iop = mach->ideal_Opcode();
2443 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2444 if( iop == Op_Con ) continue; // Do not schedule Top
2445 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2446 mach->pipeline() == MachNode::pipeline_class() &&
2447 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2448 continue;
2449 break; // Funny loop structure to be sure...
2450 }
2451 // Compute last "interesting" instruction in block - last instruction we
2452 // might schedule. _bb_end points just after last schedulable inst. We
2453 // normally schedule conditional branches (despite them being forced last
2454 // in the block), because they have delay slots we can fill. Calls all
2455 // have their delay slots filled in the template expansions, so we don't
2456 // bother scheduling them.
2457 Node *last = bb->_nodes[_bb_end];
2458 // Ignore trailing NOPs.
2459 while (_bb_end > 0 && last->is_Mach() &&
2460 last->as_Mach()->ideal_Opcode() == Op_Con) {
2461 last = bb->_nodes[--_bb_end];
2462 }
2463 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2464 if( last->is_Catch() ||
2465 // Exclude unreachable path case when Halt node is in a separate block.
2466 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2467 // There must be a prior call. Skip it.
2468 while( !bb->_nodes[--_bb_end]->is_MachCall() ) {
2469 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" );
2470 }
2471 } else if( last->is_MachNullCheck() ) {
2472 // Backup so the last null-checked memory instruction is
2473 // outside the schedulable range. Skip over the nullcheck,
2474 // projection, and the memory nodes.
2475 Node *mem = last->in(1);
2476 do {
2477 _bb_end--;
2478 } while (mem != bb->_nodes[_bb_end]);
2479 } else {
2480 // Set _bb_end to point after last schedulable inst.
2481 _bb_end++;
2482 }
2484 assert( _bb_start <= _bb_end, "inverted block ends" );
2486 // Compute the register antidependencies for the basic block
2487 ComputeRegisterAntidependencies(bb);
2488 if (_cfg->C->failing()) return; // too many D-U pinch points
2490 // Compute intra-bb latencies for the nodes
2491 ComputeLocalLatenciesForward(bb);
2493 // Compute the usage within the block, and set the list of all nodes
2494 // in the block that have no uses within the block.
2495 ComputeUseCount(bb);
2497 // Schedule the remaining instructions in the block
2498 while ( _available.size() > 0 ) {
2499 Node *n = ChooseNodeToBundle();
2500 guarantee(n != NULL, "no nodes available");
2501 AddNodeToBundle(n,bb);
2502 }
2504 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2505 #ifdef ASSERT
2506 for( uint l = _bb_start; l < _bb_end; l++ ) {
2507 Node *n = bb->_nodes[l];
2508 uint m;
2509 for( m = 0; m < _bb_end-_bb_start; m++ )
2510 if( _scheduled[m] == n )
2511 break;
2512 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2513 }
2514 #endif
2516 // Now copy the instructions (in reverse order) back to the block
2517 for ( uint k = _bb_start; k < _bb_end; k++ )
2518 bb->_nodes.map(k, _scheduled[_bb_end-k-1]);
2520 #ifndef PRODUCT
2521 if (_cfg->C->trace_opto_output()) {
2522 tty->print("# Schedule BB#%03d (final)\n", i);
2523 uint current = 0;
2524 for (uint j = 0; j < bb->_nodes.size(); j++) {
2525 Node *n = bb->_nodes[j];
2526 if( valid_bundle_info(n) ) {
2527 Bundle *bundle = node_bundling(n);
2528 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2529 tty->print("*** Bundle: ");
2530 bundle->dump();
2531 }
2532 n->dump();
2533 }
2534 }
2535 }
2536 #endif
2537 #ifdef ASSERT
2538 verify_good_schedule(bb,"after block local scheduling");
2539 #endif
2540 }
2542 #ifndef PRODUCT
2543 if (_cfg->C->trace_opto_output())
2544 tty->print("# <- DoScheduling\n");
2545 #endif
2547 // Record final node-bundling array location
2548 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2550 } // end DoScheduling
2552 //------------------------------verify_good_schedule---------------------------
2553 // Verify that no live-range used in the block is killed in the block by a
2554 // wrong DEF. This doesn't verify live-ranges that span blocks.
2556 // Check for edge existence. Used to avoid adding redundant precedence edges.
2557 static bool edge_from_to( Node *from, Node *to ) {
2558 for( uint i=0; i<from->len(); i++ )
2559 if( from->in(i) == to )
2560 return true;
2561 return false;
2562 }
2564 #ifdef ASSERT
2565 //------------------------------verify_do_def----------------------------------
2566 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2567 // Check for bad kills
2568 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2569 Node *prior_use = _reg_node[def];
2570 if( prior_use && !edge_from_to(prior_use,n) ) {
2571 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2572 n->dump();
2573 tty->print_cr("...");
2574 prior_use->dump();
2575 assert(edge_from_to(prior_use,n),msg);
2576 }
2577 _reg_node.map(def,NULL); // Kill live USEs
2578 }
2579 }
2581 //------------------------------verify_good_schedule---------------------------
2582 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2584 // Zap to something reasonable for the verify code
2585 _reg_node.clear();
2587 // Walk over the block backwards. Check to make sure each DEF doesn't
2588 // kill a live value (other than the one it's supposed to). Add each
2589 // USE to the live set.
2590 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {
2591 Node *n = b->_nodes[i];
2592 int n_op = n->Opcode();
2593 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2594 // Fat-proj kills a slew of registers
2595 RegMask rm = n->out_RegMask();// Make local copy
2596 while( rm.is_NotEmpty() ) {
2597 OptoReg::Name kill = rm.find_first_elem();
2598 rm.Remove(kill);
2599 verify_do_def( n, kill, msg );
2600 }
2601 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2602 // Get DEF'd registers the normal way
2603 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2604 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2605 }
2607 // Now make all USEs live
2608 for( uint i=1; i<n->req(); i++ ) {
2609 Node *def = n->in(i);
2610 assert(def != 0, "input edge required");
2611 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2612 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2613 if( OptoReg::is_valid(reg_lo) ) {
2614 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2615 _reg_node.map(reg_lo,n);
2616 }
2617 if( OptoReg::is_valid(reg_hi) ) {
2618 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2619 _reg_node.map(reg_hi,n);
2620 }
2621 }
2623 }
2625 // Zap to something reasonable for the Antidependence code
2626 _reg_node.clear();
2627 }
2628 #endif
2630 // Conditionally add precedence edges. Avoid putting edges on Projs.
2631 static void add_prec_edge_from_to( Node *from, Node *to ) {
2632 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2633 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2634 from = from->in(0);
2635 }
2636 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2637 !edge_from_to( from, to ) ) // Avoid duplicate edge
2638 from->add_prec(to);
2639 }
2641 //------------------------------anti_do_def------------------------------------
2642 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2643 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2644 return;
2646 Node *pinch = _reg_node[def_reg]; // Get pinch point
2647 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet?
2648 is_def ) { // Check for a true def (not a kill)
2649 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2650 return;
2651 }
2653 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2654 debug_only( def = (Node*)0xdeadbeef; )
2656 // After some number of kills there _may_ be a later def
2657 Node *later_def = NULL;
2659 // Finding a kill requires a real pinch-point.
2660 // Check for not already having a pinch-point.
2661 // Pinch points are Op_Node's.
2662 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2663 later_def = pinch; // Must be def/kill as optimistic pinch-point
2664 if ( _pinch_free_list.size() > 0) {
2665 pinch = _pinch_free_list.pop();
2666 } else {
2667 pinch = new (_cfg->C) Node(1); // Pinch point to-be
2668 }
2669 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2670 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2671 return;
2672 }
2673 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init)
2674 _reg_node.map(def_reg,pinch); // Record pinch-point
2675 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2676 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2677 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2678 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2679 later_def = NULL; // and no later def
2680 }
2681 pinch->set_req(0,later_def); // Hook later def so we can find it
2682 } else { // Else have valid pinch point
2683 if( pinch->in(0) ) // If there is a later-def
2684 later_def = pinch->in(0); // Get it
2685 }
2687 // Add output-dependence edge from later def to kill
2688 if( later_def ) // If there is some original def
2689 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2691 // See if current kill is also a use, and so is forced to be the pinch-point.
2692 if( pinch->Opcode() == Op_Node ) {
2693 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2694 for( uint i=1; i<uses->req(); i++ ) {
2695 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2696 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2697 // Yes, found a use/kill pinch-point
2698 pinch->set_req(0,NULL); //
2699 pinch->replace_by(kill); // Move anti-dep edges up
2700 pinch = kill;
2701 _reg_node.map(def_reg,pinch);
2702 return;
2703 }
2704 }
2705 }
2707 // Add edge from kill to pinch-point
2708 add_prec_edge_from_to(kill,pinch);
2709 }
2711 //------------------------------anti_do_use------------------------------------
2712 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2713 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2714 return;
2715 Node *pinch = _reg_node[use_reg]; // Get pinch point
2716 // Check for no later def_reg/kill in block
2717 if( pinch && _bbs[pinch->_idx] == b &&
2718 // Use has to be block-local as well
2719 _bbs[use->_idx] == b ) {
2720 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2721 pinch->req() == 1 ) { // pinch not yet in block?
2722 pinch->del_req(0); // yank pointer to later-def, also set flag
2723 // Insert the pinch-point in the block just after the last use
2724 b->_nodes.insert(b->find_node(use)+1,pinch);
2725 _bb_end++; // Increase size scheduled region in block
2726 }
2728 add_prec_edge_from_to(pinch,use);
2729 }
2730 }
2732 //------------------------------ComputeRegisterAntidependences-----------------
2733 // We insert antidependences between the reads and following write of
2734 // allocated registers to prevent illegal code motion. Hopefully, the
2735 // number of added references should be fairly small, especially as we
2736 // are only adding references within the current basic block.
2737 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2739 #ifdef ASSERT
2740 verify_good_schedule(b,"before block local scheduling");
2741 #endif
2743 // A valid schedule, for each register independently, is an endless cycle
2744 // of: a def, then some uses (connected to the def by true dependencies),
2745 // then some kills (defs with no uses), finally the cycle repeats with a new
2746 // def. The uses are allowed to float relative to each other, as are the
2747 // kills. No use is allowed to slide past a kill (or def). This requires
2748 // antidependencies between all uses of a single def and all kills that
2749 // follow, up to the next def. More edges are redundant, because later defs
2750 // & kills are already serialized with true or antidependencies. To keep
2751 // the edge count down, we add a 'pinch point' node if there's more than
2752 // one use or more than one kill/def.
2754 // We add dependencies in one bottom-up pass.
2756 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2758 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2759 // register. If not, we record the DEF/KILL in _reg_node, the
2760 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2761 // "pinch point", a new Node that's in the graph but not in the block.
2762 // We put edges from the prior and current DEF/KILLs to the pinch point.
2763 // We put the pinch point in _reg_node. If there's already a pinch point
2764 // we merely add an edge from the current DEF/KILL to the pinch point.
2766 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2767 // put an edge from the pinch point to the USE.
2769 // To be expedient, the _reg_node array is pre-allocated for the whole
2770 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2771 // or a valid def/kill/pinch-point, or a leftover node from some prior
2772 // block. Leftover node from some prior block is treated like a NULL (no
2773 // prior def, so no anti-dependence needed). Valid def is distinguished by
2774 // it being in the current block.
2775 bool fat_proj_seen = false;
2776 uint last_safept = _bb_end-1;
2777 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL;
2778 Node* last_safept_node = end_node;
2779 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2780 Node *n = b->_nodes[i];
2781 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2782 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2783 // Fat-proj kills a slew of registers
2784 // This can add edges to 'n' and obscure whether or not it was a def,
2785 // hence the is_def flag.
2786 fat_proj_seen = true;
2787 RegMask rm = n->out_RegMask();// Make local copy
2788 while( rm.is_NotEmpty() ) {
2789 OptoReg::Name kill = rm.find_first_elem();
2790 rm.Remove(kill);
2791 anti_do_def( b, n, kill, is_def );
2792 }
2793 } else {
2794 // Get DEF'd registers the normal way
2795 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2796 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2797 }
2799 // Kill projections on a branch should appear to occur on the
2800 // branch, not afterwards, so grab the masks from the projections
2801 // and process them.
2802 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2803 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2804 Node* use = n->fast_out(i);
2805 if (use->is_Proj()) {
2806 RegMask rm = use->out_RegMask();// Make local copy
2807 while( rm.is_NotEmpty() ) {
2808 OptoReg::Name kill = rm.find_first_elem();
2809 rm.Remove(kill);
2810 anti_do_def( b, n, kill, false );
2811 }
2812 }
2813 }
2814 }
2816 // Check each register used by this instruction for a following DEF/KILL
2817 // that must occur afterward and requires an anti-dependence edge.
2818 for( uint j=0; j<n->req(); j++ ) {
2819 Node *def = n->in(j);
2820 if( def ) {
2821 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2822 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2823 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2824 }
2825 }
2826 // Do not allow defs of new derived values to float above GC
2827 // points unless the base is definitely available at the GC point.
2829 Node *m = b->_nodes[i];
2831 // Add precedence edge from following safepoint to use of derived pointer
2832 if( last_safept_node != end_node &&
2833 m != last_safept_node) {
2834 for (uint k = 1; k < m->req(); k++) {
2835 const Type *t = m->in(k)->bottom_type();
2836 if( t->isa_oop_ptr() &&
2837 t->is_ptr()->offset() != 0 ) {
2838 last_safept_node->add_prec( m );
2839 break;
2840 }
2841 }
2842 }
2844 if( n->jvms() ) { // Precedence edge from derived to safept
2845 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2846 if( b->_nodes[last_safept] != last_safept_node ) {
2847 last_safept = b->find_node(last_safept_node);
2848 }
2849 for( uint j=last_safept; j > i; j-- ) {
2850 Node *mach = b->_nodes[j];
2851 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2852 mach->add_prec( n );
2853 }
2854 last_safept = i;
2855 last_safept_node = m;
2856 }
2857 }
2859 if (fat_proj_seen) {
2860 // Garbage collect pinch nodes that were not consumed.
2861 // They are usually created by a fat kill MachProj for a call.
2862 garbage_collect_pinch_nodes();
2863 }
2864 }
2866 //------------------------------garbage_collect_pinch_nodes-------------------------------
2868 // Garbage collect pinch nodes for reuse by other blocks.
2869 //
2870 // The block scheduler's insertion of anti-dependence
2871 // edges creates many pinch nodes when the block contains
2872 // 2 or more Calls. A pinch node is used to prevent a
2873 // combinatorial explosion of edges. If a set of kills for a
2874 // register is anti-dependent on a set of uses (or defs), rather
2875 // than adding an edge in the graph between each pair of kill
2876 // and use (or def), a pinch is inserted between them:
2877 //
2878 // use1 use2 use3
2879 // \ | /
2880 // \ | /
2881 // pinch
2882 // / | \
2883 // / | \
2884 // kill1 kill2 kill3
2885 //
2886 // One pinch node is created per register killed when
2887 // the second call is encountered during a backwards pass
2888 // over the block. Most of these pinch nodes are never
2889 // wired into the graph because the register is never
2890 // used or def'ed in the block.
2891 //
2892 void Scheduling::garbage_collect_pinch_nodes() {
2893 #ifndef PRODUCT
2894 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2895 #endif
2896 int trace_cnt = 0;
2897 for (uint k = 0; k < _reg_node.Size(); k++) {
2898 Node* pinch = _reg_node[k];
2899 if (pinch != NULL && pinch->Opcode() == Op_Node &&
2900 // no predecence input edges
2901 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2902 cleanup_pinch(pinch);
2903 _pinch_free_list.push(pinch);
2904 _reg_node.map(k, NULL);
2905 #ifndef PRODUCT
2906 if (_cfg->C->trace_opto_output()) {
2907 trace_cnt++;
2908 if (trace_cnt > 40) {
2909 tty->print("\n");
2910 trace_cnt = 0;
2911 }
2912 tty->print(" %d", pinch->_idx);
2913 }
2914 #endif
2915 }
2916 }
2917 #ifndef PRODUCT
2918 if (_cfg->C->trace_opto_output()) tty->print("\n");
2919 #endif
2920 }
2922 // Clean up a pinch node for reuse.
2923 void Scheduling::cleanup_pinch( Node *pinch ) {
2924 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2926 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2927 Node* use = pinch->last_out(i);
2928 uint uses_found = 0;
2929 for (uint j = use->req(); j < use->len(); j++) {
2930 if (use->in(j) == pinch) {
2931 use->rm_prec(j);
2932 uses_found++;
2933 }
2934 }
2935 assert(uses_found > 0, "must be a precedence edge");
2936 i -= uses_found; // we deleted 1 or more copies of this edge
2937 }
2938 // May have a later_def entry
2939 pinch->set_req(0, NULL);
2940 }
2942 //------------------------------print_statistics-------------------------------
2943 #ifndef PRODUCT
2945 void Scheduling::dump_available() const {
2946 tty->print("#Availist ");
2947 for (uint i = 0; i < _available.size(); i++)
2948 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2949 tty->cr();
2950 }
2952 // Print Scheduling Statistics
2953 void Scheduling::print_statistics() {
2954 // Print the size added by nops for bundling
2955 tty->print("Nops added %d bytes to total of %d bytes",
2956 _total_nop_size, _total_method_size);
2957 if (_total_method_size > 0)
2958 tty->print(", for %.2f%%",
2959 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2960 tty->print("\n");
2962 // Print the number of branch shadows filled
2963 if (Pipeline::_branch_has_delay_slot) {
2964 tty->print("Of %d branches, %d had unconditional delay slots filled",
2965 _total_branches, _total_unconditional_delays);
2966 if (_total_branches > 0)
2967 tty->print(", for %.2f%%",
2968 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2969 tty->print("\n");
2970 }
2972 uint total_instructions = 0, total_bundles = 0;
2974 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2975 uint bundle_count = _total_instructions_per_bundle[i];
2976 total_instructions += bundle_count * i;
2977 total_bundles += bundle_count;
2978 }
2980 if (total_bundles > 0)
2981 tty->print("Average ILP (excluding nops) is %.2f\n",
2982 ((double)total_instructions) / ((double)total_bundles));
2983 }
2984 #endif