Fri, 28 Feb 2014 02:43:16 -0800
8035283: Second phase of branch shortening doesn't account for loop alignment
Summary: added missing check for loop padding case.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "code/compiledIC.hpp"
28 #include "code/debugInfo.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/output.hpp"
38 #include "opto/regalloc.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/type.hpp"
42 #include "runtime/handles.inline.hpp"
43 #include "utilities/xmlstream.hpp"
45 extern uint size_exception_handler();
46 extern uint size_deopt_handler();
48 #ifndef PRODUCT
49 #define DEBUG_ARG(x) , x
50 #else
51 #define DEBUG_ARG(x)
52 #endif
54 extern int emit_exception_handler(CodeBuffer &cbuf);
55 extern int emit_deopt_handler(CodeBuffer &cbuf);
57 // Convert Nodes to instruction bits and pass off to the VM
58 void Compile::Output() {
59 // RootNode goes
60 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
62 // The number of new nodes (mostly MachNop) is proportional to
63 // the number of java calls and inner loops which are aligned.
64 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
65 C->inner_loops()*(OptoLoopAlignment-1)),
66 "out of nodes before code generation" ) ) {
67 return;
68 }
69 // Make sure I can find the Start Node
70 Block *entry = _cfg->get_block(1);
71 Block *broot = _cfg->get_root_block();
73 const StartNode *start = entry->head()->as_Start();
75 // Replace StartNode with prolog
76 MachPrologNode *prolog = new (this) MachPrologNode();
77 entry->map_node(prolog, 0);
78 _cfg->map_node_to_block(prolog, entry);
79 _cfg->unmap_node_from_block(start); // start is no longer in any block
81 // Virtual methods need an unverified entry point
83 if( is_osr_compilation() ) {
84 if( PoisonOSREntry ) {
85 // TODO: Should use a ShouldNotReachHereNode...
86 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
87 }
88 } else {
89 if( _method && !_method->flags().is_static() ) {
90 // Insert unvalidated entry point
91 _cfg->insert( broot, 0, new (this) MachUEPNode() );
92 }
94 }
97 // Break before main entry point
98 if( (_method && _method->break_at_execute())
99 #ifndef PRODUCT
100 ||(OptoBreakpoint && is_method_compilation())
101 ||(OptoBreakpointOSR && is_osr_compilation())
102 ||(OptoBreakpointC2R && !_method)
103 #endif
104 ) {
105 // checking for _method means that OptoBreakpoint does not apply to
106 // runtime stubs or frame converters
107 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
108 }
110 // Insert epilogs before every return
111 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
112 Block* block = _cfg->get_block(i);
113 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
114 Node* m = block->end();
115 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
116 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
117 block->add_inst(epilog);
118 _cfg->map_node_to_block(epilog, block);
119 }
120 }
121 }
123 # ifdef ENABLE_ZAP_DEAD_LOCALS
124 if (ZapDeadCompiledLocals) {
125 Insert_zap_nodes();
126 }
127 # endif
129 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
130 blk_starts[0] = 0;
132 // Initialize code buffer and process short branches.
133 CodeBuffer* cb = init_buffer(blk_starts);
135 if (cb == NULL || failing()) {
136 return;
137 }
139 ScheduleAndBundle();
141 #ifndef PRODUCT
142 if (trace_opto_output()) {
143 tty->print("\n---- After ScheduleAndBundle ----\n");
144 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
145 tty->print("\nBB#%03d:\n", i);
146 Block* block = _cfg->get_block(i);
147 for (uint j = 0; j < block->number_of_nodes(); j++) {
148 Node* n = block->get_node(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()) {
158 return;
159 }
161 BuildOopMaps();
163 if (failing()) {
164 return;
165 }
167 fill_buffer(cb, blk_starts);
168 }
170 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
171 // Determine if we need to generate a stack overflow check.
172 // Do it if the method is not a stub function and
173 // has java calls or has frame size > vm_page_size/8.
174 return (UseStackBanging && stub_function() == NULL &&
175 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
176 }
178 bool Compile::need_register_stack_bang() const {
179 // Determine if we need to generate a register stack overflow check.
180 // This is only used on architectures which have split register
181 // and memory stacks (ie. IA64).
182 // Bang if the method is not a stub function and has java calls
183 return (stub_function() == NULL && has_java_calls());
184 }
186 # ifdef ENABLE_ZAP_DEAD_LOCALS
189 // In order to catch compiler oop-map bugs, we have implemented
190 // a debugging mode called ZapDeadCompilerLocals.
191 // This mode causes the compiler to insert a call to a runtime routine,
192 // "zap_dead_locals", right before each place in compiled code
193 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
194 // The runtime routine checks that locations mapped as oops are really
195 // oops, that locations mapped as values do not look like oops,
196 // and that locations mapped as dead are not used later
197 // (by zapping them to an invalid address).
199 int Compile::_CompiledZap_count = 0;
201 void Compile::Insert_zap_nodes() {
202 bool skip = false;
205 // Dink with static counts because code code without the extra
206 // runtime calls is MUCH faster for debugging purposes
208 if ( CompileZapFirst == 0 ) ; // nothing special
209 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
210 else if ( CompileZapFirst == CompiledZap_count() )
211 warning("starting zap compilation after skipping");
213 if ( CompileZapLast == -1 ) ; // nothing special
214 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
215 else if ( CompileZapLast == CompiledZap_count() )
216 warning("about to compile last zap");
218 ++_CompiledZap_count; // counts skipped zaps, too
220 if ( skip ) return;
223 if ( _method == NULL )
224 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
226 // Insert call to zap runtime stub before every node with an oop map
227 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
228 Block *b = _cfg->get_block(i);
229 for ( uint j = 0; j < b->number_of_nodes(); ++j ) {
230 Node *n = b->get_node(j);
232 // Determining if we should insert a zap-a-lot node in output.
233 // We do that for all nodes that has oopmap info, except for calls
234 // to allocation. Calls to allocation passes in the old top-of-eden pointer
235 // and expect the C code to reset it. Hence, there can be no safepoints between
236 // the inlined-allocation and the call to new_Java, etc.
237 // We also cannot zap monitor calls, as they must hold the microlock
238 // during the call to Zap, which also wants to grab the microlock.
239 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
240 if ( insert ) { // it is MachSafePoint
241 if ( !n->is_MachCall() ) {
242 insert = false;
243 } else if ( n->is_MachCall() ) {
244 MachCallNode* call = n->as_MachCall();
245 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
246 call->entry_point() == OptoRuntime::new_array_Java() ||
247 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
248 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
249 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
250 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
251 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
252 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
253 ) {
254 insert = false;
255 }
256 }
257 if (insert) {
258 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
259 b->insert_node(zap, j);
260 _cfg->map_node_to_block(zap, b);
261 ++j;
262 }
263 }
264 }
265 }
266 }
269 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
270 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
271 CallStaticJavaNode* ideal_node =
272 new (this) CallStaticJavaNode( tf,
273 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
274 "call zap dead locals stub", 0, TypePtr::BOTTOM);
275 // We need to copy the OopMap from the site we're zapping at.
276 // We have to make a copy, because the zap site might not be
277 // a call site, and zap_dead is a call site.
278 OopMap* clone = node_to_check->oop_map()->deep_copy();
280 // Add the cloned OopMap to the zap node
281 ideal_node->set_oop_map(clone);
282 return _matcher->match_sfpt(ideal_node);
283 }
285 bool Compile::is_node_getting_a_safepoint( Node* n) {
286 // This code duplicates the logic prior to the call of add_safepoint
287 // below in this file.
288 if( n->is_MachSafePoint() ) return true;
289 return false;
290 }
292 # endif // ENABLE_ZAP_DEAD_LOCALS
294 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
295 // of a loop. When aligning a loop we need to provide enough instructions
296 // in cpu's fetch buffer to feed decoders. The loop alignment could be
297 // avoided if we have enough instructions in fetch buffer at the head of a loop.
298 // By default, the size is set to 999999 by Block's constructor so that
299 // a loop will be aligned if the size is not reset here.
300 //
301 // Note: Mach instructions could contain several HW instructions
302 // so the size is estimated only.
303 //
304 void Compile::compute_loop_first_inst_sizes() {
305 // The next condition is used to gate the loop alignment optimization.
306 // Don't aligned a loop if there are enough instructions at the head of a loop
307 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
308 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
309 // equal to 11 bytes which is the largest address NOP instruction.
310 if (MaxLoopPad < OptoLoopAlignment - 1) {
311 uint last_block = _cfg->number_of_blocks() - 1;
312 for (uint i = 1; i <= last_block; i++) {
313 Block* block = _cfg->get_block(i);
314 // Check the first loop's block which requires an alignment.
315 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
316 uint sum_size = 0;
317 uint inst_cnt = NumberOfLoopInstrToAlign;
318 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
320 // Check subsequent fallthrough blocks if the loop's first
321 // block(s) does not have enough instructions.
322 Block *nb = block;
323 while(inst_cnt > 0 &&
324 i < last_block &&
325 !_cfg->get_block(i + 1)->has_loop_alignment() &&
326 !nb->has_successor(block)) {
327 i++;
328 nb = _cfg->get_block(i);
329 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
330 } // while( inst_cnt > 0 && i < last_block )
332 block->set_first_inst_size(sum_size);
333 } // f( b->head()->is_Loop() )
334 } // for( i <= last_block )
335 } // if( MaxLoopPad < OptoLoopAlignment-1 )
336 }
338 // The architecture description provides short branch variants for some long
339 // branch instructions. Replace eligible long branches with short branches.
340 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
341 // Compute size of each block, method size, and relocation information size
342 uint nblocks = _cfg->number_of_blocks();
344 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
345 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
346 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
348 // Collect worst case block paddings
349 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
350 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
352 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
353 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
355 bool has_short_branch_candidate = false;
357 // Initialize the sizes to 0
358 code_size = 0; // Size in bytes of generated code
359 stub_size = 0; // Size in bytes of all stub entries
360 // Size in bytes of all relocation entries, including those in local stubs.
361 // Start with 2-bytes of reloc info for the unvalidated entry point
362 reloc_size = 1; // Number of relocation entries
364 // Make three passes. The first computes pessimistic blk_starts,
365 // relative jmp_offset and reloc_size information. The second performs
366 // short branch substitution using the pessimistic sizing. The
367 // third inserts nops where needed.
369 // Step one, perform a pessimistic sizing pass.
370 uint last_call_adr = max_uint;
371 uint last_avoid_back_to_back_adr = max_uint;
372 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
373 for (uint i = 0; i < nblocks; i++) { // For all blocks
374 Block* block = _cfg->get_block(i);
376 // During short branch replacement, we store the relative (to blk_starts)
377 // offset of jump in jmp_offset, rather than the absolute offset of jump.
378 // This is so that we do not need to recompute sizes of all nodes when
379 // we compute correct blk_starts in our next sizing pass.
380 jmp_offset[i] = 0;
381 jmp_size[i] = 0;
382 jmp_nidx[i] = -1;
383 DEBUG_ONLY( jmp_target[i] = 0; )
384 DEBUG_ONLY( jmp_rule[i] = 0; )
386 // Sum all instruction sizes to compute block size
387 uint last_inst = block->number_of_nodes();
388 uint blk_size = 0;
389 for (uint j = 0; j < last_inst; j++) {
390 Node* nj = block->get_node(j);
391 // Handle machine instruction nodes
392 if (nj->is_Mach()) {
393 MachNode *mach = nj->as_Mach();
394 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
395 reloc_size += mach->reloc();
396 if (mach->is_MachCall()) {
397 MachCallNode *mcall = mach->as_MachCall();
398 // This destination address is NOT PC-relative
400 mcall->method_set((intptr_t)mcall->entry_point());
402 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
403 stub_size += CompiledStaticCall::to_interp_stub_size();
404 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
405 }
406 } else if (mach->is_MachSafePoint()) {
407 // If call/safepoint are adjacent, account for possible
408 // nop to disambiguate the two safepoints.
409 // ScheduleAndBundle() can rearrange nodes in a block,
410 // check for all offsets inside this block.
411 if (last_call_adr >= blk_starts[i]) {
412 blk_size += nop_size;
413 }
414 }
415 if (mach->avoid_back_to_back()) {
416 // Nop is inserted between "avoid back to back" instructions.
417 // ScheduleAndBundle() can rearrange nodes in a block,
418 // check for all offsets inside this block.
419 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
420 blk_size += nop_size;
421 }
422 }
423 if (mach->may_be_short_branch()) {
424 if (!nj->is_MachBranch()) {
425 #ifndef PRODUCT
426 nj->dump(3);
427 #endif
428 Unimplemented();
429 }
430 assert(jmp_nidx[i] == -1, "block should have only one branch");
431 jmp_offset[i] = blk_size;
432 jmp_size[i] = nj->size(_regalloc);
433 jmp_nidx[i] = j;
434 has_short_branch_candidate = true;
435 }
436 }
437 blk_size += nj->size(_regalloc);
438 // Remember end of call offset
439 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
440 last_call_adr = blk_starts[i]+blk_size;
441 }
442 // Remember end of avoid_back_to_back offset
443 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
444 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
445 }
446 }
448 // When the next block starts a loop, we may insert pad NOP
449 // instructions. Since we cannot know our future alignment,
450 // assume the worst.
451 if (i < nblocks - 1) {
452 Block* nb = _cfg->get_block(i + 1);
453 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
454 if (max_loop_pad > 0) {
455 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
456 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
457 // If either is the last instruction in this block, bump by
458 // max_loop_pad in lock-step with blk_size, so sizing
459 // calculations in subsequent blocks still can conservatively
460 // detect that it may the last instruction in this block.
461 if (last_call_adr == blk_starts[i]+blk_size) {
462 last_call_adr += max_loop_pad;
463 }
464 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
465 last_avoid_back_to_back_adr += max_loop_pad;
466 }
467 blk_size += max_loop_pad;
468 block_worst_case_pad[i + 1] = max_loop_pad;
469 }
470 }
472 // Save block size; update total method size
473 blk_starts[i+1] = blk_starts[i]+blk_size;
474 }
476 // Step two, replace eligible long jumps.
477 bool progress = true;
478 uint last_may_be_short_branch_adr = max_uint;
479 while (has_short_branch_candidate && progress) {
480 progress = false;
481 has_short_branch_candidate = false;
482 int adjust_block_start = 0;
483 for (uint i = 0; i < nblocks; i++) {
484 Block* block = _cfg->get_block(i);
485 int idx = jmp_nidx[i];
486 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
487 if (mach != NULL && mach->may_be_short_branch()) {
488 #ifdef ASSERT
489 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
490 int j;
491 // Find the branch; ignore trailing NOPs.
492 for (j = block->number_of_nodes()-1; j>=0; j--) {
493 Node* n = block->get_node(j);
494 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
495 break;
496 }
497 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
498 #endif
499 int br_size = jmp_size[i];
500 int br_offs = blk_starts[i] + jmp_offset[i];
502 // This requires the TRUE branch target be in succs[0]
503 uint bnum = block->non_connector_successor(0)->_pre_order;
504 int offset = blk_starts[bnum] - br_offs;
505 if (bnum > i) { // adjust following block's offset
506 offset -= adjust_block_start;
507 }
509 // This block can be a loop header, account for the padding
510 // in the previous block.
511 int block_padding = block_worst_case_pad[i];
512 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
513 // In the following code a nop could be inserted before
514 // the branch which will increase the backward distance.
515 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
516 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
518 if (needs_padding && offset <= 0)
519 offset -= nop_size;
521 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
522 // We've got a winner. Replace this branch.
523 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
525 // Update the jmp_size.
526 int new_size = replacement->size(_regalloc);
527 int diff = br_size - new_size;
528 assert(diff >= (int)nop_size, "short_branch size should be smaller");
529 // Conservatively take into accound padding between
530 // avoid_back_to_back branches. Previous branch could be
531 // converted into avoid_back_to_back branch during next
532 // rounds.
533 if (needs_padding && replacement->avoid_back_to_back()) {
534 jmp_offset[i] += nop_size;
535 diff -= nop_size;
536 }
537 adjust_block_start += diff;
538 block->map_node(replacement, idx);
539 mach->subsume_by(replacement, C);
540 mach = replacement;
541 progress = true;
543 jmp_size[i] = new_size;
544 DEBUG_ONLY( jmp_target[i] = bnum; );
545 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
546 } else {
547 // The jump distance is not short, try again during next iteration.
548 has_short_branch_candidate = true;
549 }
550 } // (mach->may_be_short_branch())
551 if (mach != NULL && (mach->may_be_short_branch() ||
552 mach->avoid_back_to_back())) {
553 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
554 }
555 blk_starts[i+1] -= adjust_block_start;
556 }
557 }
559 #ifdef ASSERT
560 for (uint i = 0; i < nblocks; i++) { // For all blocks
561 if (jmp_target[i] != 0) {
562 int br_size = jmp_size[i];
563 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
564 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
565 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]);
566 }
567 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
568 }
569 }
570 #endif
572 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
573 // after ScheduleAndBundle().
575 // ------------------
576 // Compute size for code buffer
577 code_size = blk_starts[nblocks];
579 // Relocation records
580 reloc_size += 1; // Relo entry for exception handler
582 // Adjust reloc_size to number of record of relocation info
583 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
584 // a relocation index.
585 // The CodeBuffer will expand the locs array if this estimate is too low.
586 reloc_size *= 10 / sizeof(relocInfo);
587 }
589 //------------------------------FillLocArray-----------------------------------
590 // Create a bit of debug info and append it to the array. The mapping is from
591 // Java local or expression stack to constant, register or stack-slot. For
592 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
593 // entry has been taken care of and caller should skip it).
594 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
595 // This should never have accepted Bad before
596 assert(OptoReg::is_valid(regnum), "location must be valid");
597 return (OptoReg::is_reg(regnum))
598 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
599 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
600 }
603 ObjectValue*
604 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
605 for (int i = 0; i < objs->length(); i++) {
606 assert(objs->at(i)->is_object(), "corrupt object cache");
607 ObjectValue* sv = (ObjectValue*) objs->at(i);
608 if (sv->id() == id) {
609 return sv;
610 }
611 }
612 // Otherwise..
613 return NULL;
614 }
616 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
617 ObjectValue* sv ) {
618 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
619 objs->append(sv);
620 }
623 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
624 GrowableArray<ScopeValue*> *array,
625 GrowableArray<ScopeValue*> *objs ) {
626 assert( local, "use _top instead of null" );
627 if (array->length() != idx) {
628 assert(array->length() == idx + 1, "Unexpected array count");
629 // Old functionality:
630 // return
631 // New functionality:
632 // Assert if the local is not top. In product mode let the new node
633 // override the old entry.
634 assert(local == top(), "LocArray collision");
635 if (local == top()) {
636 return;
637 }
638 array->pop();
639 }
640 const Type *t = local->bottom_type();
642 // Is it a safepoint scalar object node?
643 if (local->is_SafePointScalarObject()) {
644 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
646 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
647 if (sv == NULL) {
648 ciKlass* cik = t->is_oopptr()->klass();
649 assert(cik->is_instance_klass() ||
650 cik->is_array_klass(), "Not supported allocation.");
651 sv = new ObjectValue(spobj->_idx,
652 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
653 Compile::set_sv_for_object_node(objs, sv);
655 uint first_ind = spobj->first_index(sfpt->jvms());
656 for (uint i = 0; i < spobj->n_fields(); i++) {
657 Node* fld_node = sfpt->in(first_ind+i);
658 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
659 }
660 }
661 array->append(sv);
662 return;
663 }
665 // Grab the register number for the local
666 OptoReg::Name regnum = _regalloc->get_reg_first(local);
667 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
668 // Record the double as two float registers.
669 // The register mask for such a value always specifies two adjacent
670 // float registers, with the lower register number even.
671 // Normally, the allocation of high and low words to these registers
672 // is irrelevant, because nearly all operations on register pairs
673 // (e.g., StoreD) treat them as a single unit.
674 // Here, we assume in addition that the words in these two registers
675 // stored "naturally" (by operations like StoreD and double stores
676 // within the interpreter) such that the lower-numbered register
677 // is written to the lower memory address. This may seem like
678 // a machine dependency, but it is not--it is a requirement on
679 // the author of the <arch>.ad file to ensure that, for every
680 // even/odd double-register pair to which a double may be allocated,
681 // the word in the even single-register is stored to the first
682 // memory word. (Note that register numbers are completely
683 // arbitrary, and are not tied to any machine-level encodings.)
684 #ifdef _LP64
685 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
686 array->append(new ConstantIntValue(0));
687 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
688 } else if ( t->base() == Type::Long ) {
689 array->append(new ConstantIntValue(0));
690 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
691 } else if ( t->base() == Type::RawPtr ) {
692 // jsr/ret return address which must be restored into a the full
693 // width 64-bit stack slot.
694 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
695 }
696 #else //_LP64
697 #ifdef SPARC
698 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
699 // For SPARC we have to swap high and low words for
700 // long values stored in a single-register (g0-g7).
701 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
702 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
703 } else
704 #endif //SPARC
705 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
706 // Repack the double/long as two jints.
707 // The convention the interpreter uses is that the second local
708 // holds the first raw word of the native double representation.
709 // This is actually reasonable, since locals and stack arrays
710 // grow downwards in all implementations.
711 // (If, on some machine, the interpreter's Java locals or stack
712 // were to grow upwards, the embedded doubles would be word-swapped.)
713 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
714 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
715 }
716 #endif //_LP64
717 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
718 OptoReg::is_reg(regnum) ) {
719 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
720 ? Location::float_in_dbl : Location::normal ));
721 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
722 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
723 ? Location::int_in_long : Location::normal ));
724 } else if( t->base() == Type::NarrowOop ) {
725 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
726 } else {
727 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
728 }
729 return;
730 }
732 // No register. It must be constant data.
733 switch (t->base()) {
734 case Type::Half: // Second half of a double
735 ShouldNotReachHere(); // Caller should skip 2nd halves
736 break;
737 case Type::AnyPtr:
738 array->append(new ConstantOopWriteValue(NULL));
739 break;
740 case Type::AryPtr:
741 case Type::InstPtr: // fall through
742 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
743 break;
744 case Type::NarrowOop:
745 if (t == TypeNarrowOop::NULL_PTR) {
746 array->append(new ConstantOopWriteValue(NULL));
747 } else {
748 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
749 }
750 break;
751 case Type::Int:
752 array->append(new ConstantIntValue(t->is_int()->get_con()));
753 break;
754 case Type::RawPtr:
755 // A return address (T_ADDRESS).
756 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
757 #ifdef _LP64
758 // Must be restored to the full-width 64-bit stack slot.
759 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
760 #else
761 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
762 #endif
763 break;
764 case Type::FloatCon: {
765 float f = t->is_float_constant()->getf();
766 array->append(new ConstantIntValue(jint_cast(f)));
767 break;
768 }
769 case Type::DoubleCon: {
770 jdouble d = t->is_double_constant()->getd();
771 #ifdef _LP64
772 array->append(new ConstantIntValue(0));
773 array->append(new ConstantDoubleValue(d));
774 #else
775 // Repack the double as two jints.
776 // The convention the interpreter uses is that the second local
777 // holds the first raw word of the native double representation.
778 // This is actually reasonable, since locals and stack arrays
779 // grow downwards in all implementations.
780 // (If, on some machine, the interpreter's Java locals or stack
781 // were to grow upwards, the embedded doubles would be word-swapped.)
782 jint *dp = (jint*)&d;
783 array->append(new ConstantIntValue(dp[1]));
784 array->append(new ConstantIntValue(dp[0]));
785 #endif
786 break;
787 }
788 case Type::Long: {
789 jlong d = t->is_long()->get_con();
790 #ifdef _LP64
791 array->append(new ConstantIntValue(0));
792 array->append(new ConstantLongValue(d));
793 #else
794 // Repack the long as two jints.
795 // The convention the interpreter uses is that the second local
796 // holds the first raw word of the native double representation.
797 // This is actually reasonable, since locals and stack arrays
798 // grow downwards in all implementations.
799 // (If, on some machine, the interpreter's Java locals or stack
800 // were to grow upwards, the embedded doubles would be word-swapped.)
801 jint *dp = (jint*)&d;
802 array->append(new ConstantIntValue(dp[1]));
803 array->append(new ConstantIntValue(dp[0]));
804 #endif
805 break;
806 }
807 case Type::Top: // Add an illegal value here
808 array->append(new LocationValue(Location()));
809 break;
810 default:
811 ShouldNotReachHere();
812 break;
813 }
814 }
816 // Determine if this node starts a bundle
817 bool Compile::starts_bundle(const Node *n) const {
818 return (_node_bundling_limit > n->_idx &&
819 _node_bundling_base[n->_idx].starts_bundle());
820 }
822 //--------------------------Process_OopMap_Node--------------------------------
823 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
825 // Handle special safepoint nodes for synchronization
826 MachSafePointNode *sfn = mach->as_MachSafePoint();
827 MachCallNode *mcall;
829 #ifdef ENABLE_ZAP_DEAD_LOCALS
830 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
831 #endif
833 int safepoint_pc_offset = current_offset;
834 bool is_method_handle_invoke = false;
835 bool return_oop = false;
837 // Add the safepoint in the DebugInfoRecorder
838 if( !mach->is_MachCall() ) {
839 mcall = NULL;
840 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
841 } else {
842 mcall = mach->as_MachCall();
844 // Is the call a MethodHandle call?
845 if (mcall->is_MachCallJava()) {
846 if (mcall->as_MachCallJava()->_method_handle_invoke) {
847 assert(has_method_handle_invokes(), "must have been set during call generation");
848 is_method_handle_invoke = true;
849 }
850 }
852 // Check if a call returns an object.
853 if (mcall->return_value_is_used() &&
854 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
855 return_oop = true;
856 }
857 safepoint_pc_offset += mcall->ret_addr_offset();
858 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
859 }
861 // Loop over the JVMState list to add scope information
862 // Do not skip safepoints with a NULL method, they need monitor info
863 JVMState* youngest_jvms = sfn->jvms();
864 int max_depth = youngest_jvms->depth();
866 // Allocate the object pool for scalar-replaced objects -- the map from
867 // small-integer keys (which can be recorded in the local and ostack
868 // arrays) to descriptions of the object state.
869 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
871 // Visit scopes from oldest to youngest.
872 for (int depth = 1; depth <= max_depth; depth++) {
873 JVMState* jvms = youngest_jvms->of_depth(depth);
874 int idx;
875 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
876 // Safepoints that do not have method() set only provide oop-map and monitor info
877 // to support GC; these do not support deoptimization.
878 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
879 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
880 int num_mon = jvms->nof_monitors();
881 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
882 "JVMS local count must match that of the method");
884 // Add Local and Expression Stack Information
886 // Insert locals into the locarray
887 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
888 for( idx = 0; idx < num_locs; idx++ ) {
889 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
890 }
892 // Insert expression stack entries into the exparray
893 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
894 for( idx = 0; idx < num_exps; idx++ ) {
895 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
896 }
898 // Add in mappings of the monitors
899 assert( !method ||
900 !method->is_synchronized() ||
901 method->is_native() ||
902 num_mon > 0 ||
903 !GenerateSynchronizationCode,
904 "monitors must always exist for synchronized methods");
906 // Build the growable array of ScopeValues for exp stack
907 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
909 // Loop over monitors and insert into array
910 for (idx = 0; idx < num_mon; idx++) {
911 // Grab the node that defines this monitor
912 Node* box_node = sfn->monitor_box(jvms, idx);
913 Node* obj_node = sfn->monitor_obj(jvms, idx);
915 // Create ScopeValue for object
916 ScopeValue *scval = NULL;
918 if (obj_node->is_SafePointScalarObject()) {
919 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
920 scval = Compile::sv_for_node_id(objs, spobj->_idx);
921 if (scval == NULL) {
922 const Type *t = spobj->bottom_type();
923 ciKlass* cik = t->is_oopptr()->klass();
924 assert(cik->is_instance_klass() ||
925 cik->is_array_klass(), "Not supported allocation.");
926 ObjectValue* sv = new ObjectValue(spobj->_idx,
927 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
928 Compile::set_sv_for_object_node(objs, sv);
930 uint first_ind = spobj->first_index(youngest_jvms);
931 for (uint i = 0; i < spobj->n_fields(); i++) {
932 Node* fld_node = sfn->in(first_ind+i);
933 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
934 }
935 scval = sv;
936 }
937 } else if (!obj_node->is_Con()) {
938 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
939 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
940 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
941 } else {
942 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
943 }
944 } else {
945 const TypePtr *tp = obj_node->get_ptr_type();
946 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
947 }
949 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
950 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
951 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
952 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
953 }
955 // We dump the object pool first, since deoptimization reads it in first.
956 debug_info()->dump_object_pool(objs);
958 // Build first class objects to pass to scope
959 DebugToken *locvals = debug_info()->create_scope_values(locarray);
960 DebugToken *expvals = debug_info()->create_scope_values(exparray);
961 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
963 // Make method available for all Safepoints
964 ciMethod* scope_method = method ? method : _method;
965 // Describe the scope here
966 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
967 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
968 // Now we can describe the scope.
969 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
970 } // End jvms loop
972 // Mark the end of the scope set.
973 debug_info()->end_safepoint(safepoint_pc_offset);
974 }
978 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
979 class NonSafepointEmitter {
980 Compile* C;
981 JVMState* _pending_jvms;
982 int _pending_offset;
984 void emit_non_safepoint();
986 public:
987 NonSafepointEmitter(Compile* compile) {
988 this->C = compile;
989 _pending_jvms = NULL;
990 _pending_offset = 0;
991 }
993 void observe_instruction(Node* n, int pc_offset) {
994 if (!C->debug_info()->recording_non_safepoints()) return;
996 Node_Notes* nn = C->node_notes_at(n->_idx);
997 if (nn == NULL || nn->jvms() == NULL) return;
998 if (_pending_jvms != NULL &&
999 _pending_jvms->same_calls_as(nn->jvms())) {
1000 // Repeated JVMS? Stretch it up here.
1001 _pending_offset = pc_offset;
1002 } else {
1003 if (_pending_jvms != NULL &&
1004 _pending_offset < pc_offset) {
1005 emit_non_safepoint();
1006 }
1007 _pending_jvms = NULL;
1008 if (pc_offset > C->debug_info()->last_pc_offset()) {
1009 // This is the only way _pending_jvms can become non-NULL:
1010 _pending_jvms = nn->jvms();
1011 _pending_offset = pc_offset;
1012 }
1013 }
1014 }
1016 // Stay out of the way of real safepoints:
1017 void observe_safepoint(JVMState* jvms, int pc_offset) {
1018 if (_pending_jvms != NULL &&
1019 !_pending_jvms->same_calls_as(jvms) &&
1020 _pending_offset < pc_offset) {
1021 emit_non_safepoint();
1022 }
1023 _pending_jvms = NULL;
1024 }
1026 void flush_at_end() {
1027 if (_pending_jvms != NULL) {
1028 emit_non_safepoint();
1029 }
1030 _pending_jvms = NULL;
1031 }
1032 };
1034 void NonSafepointEmitter::emit_non_safepoint() {
1035 JVMState* youngest_jvms = _pending_jvms;
1036 int pc_offset = _pending_offset;
1038 // Clear it now:
1039 _pending_jvms = NULL;
1041 DebugInformationRecorder* debug_info = C->debug_info();
1042 assert(debug_info->recording_non_safepoints(), "sanity");
1044 debug_info->add_non_safepoint(pc_offset);
1045 int max_depth = youngest_jvms->depth();
1047 // Visit scopes from oldest to youngest.
1048 for (int depth = 1; depth <= max_depth; depth++) {
1049 JVMState* jvms = youngest_jvms->of_depth(depth);
1050 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1051 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1052 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1053 }
1055 // Mark the end of the scope set.
1056 debug_info->end_non_safepoint(pc_offset);
1057 }
1059 //------------------------------init_buffer------------------------------------
1060 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1062 // Set the initially allocated size
1063 int code_req = initial_code_capacity;
1064 int locs_req = initial_locs_capacity;
1065 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1066 int const_req = initial_const_capacity;
1068 int pad_req = NativeCall::instruction_size;
1069 // The extra spacing after the code is necessary on some platforms.
1070 // Sometimes we need to patch in a jump after the last instruction,
1071 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1073 // Compute the byte offset where we can store the deopt pc.
1074 if (fixed_slots() != 0) {
1075 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1076 }
1078 // Compute prolog code size
1079 _method_size = 0;
1080 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1081 #ifdef IA64
1082 if (save_argument_registers()) {
1083 // 4815101: this is a stub with implicit and unknown precision fp args.
1084 // The usual spill mechanism can only generate stfd's in this case, which
1085 // doesn't work if the fp reg to spill contains a single-precision denorm.
1086 // Instead, we hack around the normal spill mechanism using stfspill's and
1087 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1088 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1089 //
1090 // If we ever implement 16-byte 'registers' == stack slots, we can
1091 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1092 // instead of stfd/stfs/ldfd/ldfs.
1093 _frame_slots += 8*(16/BytesPerInt);
1094 }
1095 #endif
1096 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1098 if (has_mach_constant_base_node()) {
1099 // Fill the constant table.
1100 // Note: This must happen before shorten_branches.
1101 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1102 Block* b = _cfg->get_block(i);
1104 for (uint j = 0; j < b->number_of_nodes(); j++) {
1105 Node* n = b->get_node(j);
1107 // If the node is a MachConstantNode evaluate the constant
1108 // value section.
1109 if (n->is_MachConstant()) {
1110 MachConstantNode* machcon = n->as_MachConstant();
1111 machcon->eval_constant(C);
1112 }
1113 }
1114 }
1116 // Calculate the offsets of the constants and the size of the
1117 // constant table (including the padding to the next section).
1118 constant_table().calculate_offsets_and_size();
1119 const_req = constant_table().size();
1120 }
1122 // Initialize the space for the BufferBlob used to find and verify
1123 // instruction size in MachNode::emit_size()
1124 init_scratch_buffer_blob(const_req);
1125 if (failing()) return NULL; // Out of memory
1127 // Pre-compute the length of blocks and replace
1128 // long branches with short if machine supports it.
1129 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1131 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1132 int exception_handler_req = size_exception_handler();
1133 int deopt_handler_req = size_deopt_handler();
1134 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1135 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1136 stub_req += MAX_stubs_size; // ensure per-stub margin
1137 code_req += MAX_inst_size; // ensure per-instruction margin
1139 if (StressCodeBuffers)
1140 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1142 int total_req =
1143 const_req +
1144 code_req +
1145 pad_req +
1146 stub_req +
1147 exception_handler_req +
1148 deopt_handler_req; // deopt handler
1150 if (has_method_handle_invokes())
1151 total_req += deopt_handler_req; // deopt MH handler
1153 CodeBuffer* cb = code_buffer();
1154 cb->initialize(total_req, locs_req);
1156 // Have we run out of code space?
1157 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1158 C->record_failure("CodeCache is full");
1159 return NULL;
1160 }
1161 // Configure the code buffer.
1162 cb->initialize_consts_size(const_req);
1163 cb->initialize_stubs_size(stub_req);
1164 cb->initialize_oop_recorder(env()->oop_recorder());
1166 // fill in the nop array for bundling computations
1167 MachNode *_nop_list[Bundle::_nop_count];
1168 Bundle::initialize_nops(_nop_list, this);
1170 return cb;
1171 }
1173 //------------------------------fill_buffer------------------------------------
1174 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1175 // blk_starts[] contains offsets calculated during short branches processing,
1176 // offsets should not be increased during following steps.
1178 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1179 // of a loop. It is used to determine the padding for loop alignment.
1180 compute_loop_first_inst_sizes();
1182 // Create oopmap set.
1183 _oop_map_set = new OopMapSet();
1185 // !!!!! This preserves old handling of oopmaps for now
1186 debug_info()->set_oopmaps(_oop_map_set);
1188 uint nblocks = _cfg->number_of_blocks();
1189 // Count and start of implicit null check instructions
1190 uint inct_cnt = 0;
1191 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1193 // Count and start of calls
1194 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1196 uint return_offset = 0;
1197 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1199 int previous_offset = 0;
1200 int current_offset = 0;
1201 int last_call_offset = -1;
1202 int last_avoid_back_to_back_offset = -1;
1203 #ifdef ASSERT
1204 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1205 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1206 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1207 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1208 #endif
1210 // Create an array of unused labels, one for each basic block, if printing is enabled
1211 #ifndef PRODUCT
1212 int *node_offsets = NULL;
1213 uint node_offset_limit = unique();
1215 if (print_assembly())
1216 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1217 #endif
1219 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1221 // Emit the constant table.
1222 if (has_mach_constant_base_node()) {
1223 constant_table().emit(*cb);
1224 }
1226 // Create an array of labels, one for each basic block
1227 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1228 for (uint i=0; i <= nblocks; i++) {
1229 blk_labels[i].init();
1230 }
1232 // ------------------
1233 // Now fill in the code buffer
1234 Node *delay_slot = NULL;
1236 for (uint i = 0; i < nblocks; i++) {
1237 Block* block = _cfg->get_block(i);
1238 Node* head = block->head();
1240 // If this block needs to start aligned (i.e, can be reached other
1241 // than by falling-thru from the previous block), then force the
1242 // start of a new bundle.
1243 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1244 cb->flush_bundle(true);
1245 }
1247 #ifdef ASSERT
1248 if (!block->is_connector()) {
1249 stringStream st;
1250 block->dump_head(_cfg, &st);
1251 MacroAssembler(cb).block_comment(st.as_string());
1252 }
1253 jmp_target[i] = 0;
1254 jmp_offset[i] = 0;
1255 jmp_size[i] = 0;
1256 jmp_rule[i] = 0;
1257 #endif
1258 int blk_offset = current_offset;
1260 // Define the label at the beginning of the basic block
1261 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1263 uint last_inst = block->number_of_nodes();
1265 // Emit block normally, except for last instruction.
1266 // Emit means "dump code bits into code buffer".
1267 for (uint j = 0; j<last_inst; j++) {
1269 // Get the node
1270 Node* n = block->get_node(j);
1272 // See if delay slots are supported
1273 if (valid_bundle_info(n) &&
1274 node_bundling(n)->used_in_unconditional_delay()) {
1275 assert(delay_slot == NULL, "no use of delay slot node");
1276 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1278 delay_slot = n;
1279 continue;
1280 }
1282 // If this starts a new instruction group, then flush the current one
1283 // (but allow split bundles)
1284 if (Pipeline::requires_bundling() && starts_bundle(n))
1285 cb->flush_bundle(false);
1287 // The following logic is duplicated in the code ifdeffed for
1288 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1289 // should be factored out. Or maybe dispersed to the nodes?
1291 // Special handling for SafePoint/Call Nodes
1292 bool is_mcall = false;
1293 if (n->is_Mach()) {
1294 MachNode *mach = n->as_Mach();
1295 is_mcall = n->is_MachCall();
1296 bool is_sfn = n->is_MachSafePoint();
1298 // If this requires all previous instructions be flushed, then do so
1299 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1300 cb->flush_bundle(true);
1301 current_offset = cb->insts_size();
1302 }
1304 // A padding may be needed again since a previous instruction
1305 // could be moved to delay slot.
1307 // align the instruction if necessary
1308 int padding = mach->compute_padding(current_offset);
1309 // Make sure safepoint node for polling is distinct from a call's
1310 // return by adding a nop if needed.
1311 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1312 padding = nop_size;
1313 }
1314 if (padding == 0 && mach->avoid_back_to_back() &&
1315 current_offset == last_avoid_back_to_back_offset) {
1316 // Avoid back to back some instructions.
1317 padding = nop_size;
1318 }
1320 if(padding > 0) {
1321 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1322 int nops_cnt = padding / nop_size;
1323 MachNode *nop = new (this) MachNopNode(nops_cnt);
1324 block->insert_node(nop, j++);
1325 last_inst++;
1326 _cfg->map_node_to_block(nop, block);
1327 nop->emit(*cb, _regalloc);
1328 cb->flush_bundle(true);
1329 current_offset = cb->insts_size();
1330 }
1332 // Remember the start of the last call in a basic block
1333 if (is_mcall) {
1334 MachCallNode *mcall = mach->as_MachCall();
1336 // This destination address is NOT PC-relative
1337 mcall->method_set((intptr_t)mcall->entry_point());
1339 // Save the return address
1340 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1342 if (mcall->is_MachCallLeaf()) {
1343 is_mcall = false;
1344 is_sfn = false;
1345 }
1346 }
1348 // sfn will be valid whenever mcall is valid now because of inheritance
1349 if (is_sfn || is_mcall) {
1351 // Handle special safepoint nodes for synchronization
1352 if (!is_mcall) {
1353 MachSafePointNode *sfn = mach->as_MachSafePoint();
1354 // !!!!! Stubs only need an oopmap right now, so bail out
1355 if (sfn->jvms()->method() == NULL) {
1356 // Write the oopmap directly to the code blob??!!
1357 # ifdef ENABLE_ZAP_DEAD_LOCALS
1358 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1359 # endif
1360 continue;
1361 }
1362 } // End synchronization
1364 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1365 current_offset);
1366 Process_OopMap_Node(mach, current_offset);
1367 } // End if safepoint
1369 // If this is a null check, then add the start of the previous instruction to the list
1370 else if( mach->is_MachNullCheck() ) {
1371 inct_starts[inct_cnt++] = previous_offset;
1372 }
1374 // If this is a branch, then fill in the label with the target BB's label
1375 else if (mach->is_MachBranch()) {
1376 // This requires the TRUE branch target be in succs[0]
1377 uint block_num = block->non_connector_successor(0)->_pre_order;
1379 // Try to replace long branch if delay slot is not used,
1380 // it is mostly for back branches since forward branch's
1381 // distance is not updated yet.
1382 bool delay_slot_is_used = valid_bundle_info(n) &&
1383 node_bundling(n)->use_unconditional_delay();
1384 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1385 assert(delay_slot == NULL, "not expecting delay slot node");
1386 int br_size = n->size(_regalloc);
1387 int offset = blk_starts[block_num] - current_offset;
1388 if (block_num >= i) {
1389 // Current and following block's offset are not
1390 // finilized yet, adjust distance by the difference
1391 // between calculated and final offsets of current block.
1392 offset -= (blk_starts[i] - blk_offset);
1393 }
1394 // In the following code a nop could be inserted before
1395 // the branch which will increase the backward distance.
1396 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1397 if (needs_padding && offset <= 0)
1398 offset -= nop_size;
1400 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1401 // We've got a winner. Replace this branch.
1402 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1404 // Update the jmp_size.
1405 int new_size = replacement->size(_regalloc);
1406 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1407 // Insert padding between avoid_back_to_back branches.
1408 if (needs_padding && replacement->avoid_back_to_back()) {
1409 MachNode *nop = new (this) MachNopNode();
1410 block->insert_node(nop, j++);
1411 _cfg->map_node_to_block(nop, block);
1412 last_inst++;
1413 nop->emit(*cb, _regalloc);
1414 cb->flush_bundle(true);
1415 current_offset = cb->insts_size();
1416 }
1417 #ifdef ASSERT
1418 jmp_target[i] = block_num;
1419 jmp_offset[i] = current_offset - blk_offset;
1420 jmp_size[i] = new_size;
1421 jmp_rule[i] = mach->rule();
1422 #endif
1423 block->map_node(replacement, j);
1424 mach->subsume_by(replacement, C);
1425 n = replacement;
1426 mach = replacement;
1427 }
1428 }
1429 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1430 } else if (mach->ideal_Opcode() == Op_Jump) {
1431 for (uint h = 0; h < block->_num_succs; h++) {
1432 Block* succs_block = block->_succs[h];
1433 for (uint j = 1; j < succs_block->num_preds(); j++) {
1434 Node* jpn = succs_block->pred(j);
1435 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1436 uint block_num = succs_block->non_connector()->_pre_order;
1437 Label *blkLabel = &blk_labels[block_num];
1438 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1439 }
1440 }
1441 }
1442 }
1443 #ifdef ASSERT
1444 // Check that oop-store precedes the card-mark
1445 else if (mach->ideal_Opcode() == Op_StoreCM) {
1446 uint storeCM_idx = j;
1447 int count = 0;
1448 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1449 Node *oop_store = mach->in(prec); // Precedence edge
1450 if (oop_store == NULL) continue;
1451 count++;
1452 uint i4;
1453 for (i4 = 0; i4 < last_inst; ++i4) {
1454 if (block->get_node(i4) == oop_store) {
1455 break;
1456 }
1457 }
1458 // Note: This test can provide a false failure if other precedence
1459 // edges have been added to the storeCMNode.
1460 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1461 }
1462 assert(count > 0, "storeCM expects at least one precedence edge");
1463 }
1464 #endif
1465 else if (!n->is_Proj()) {
1466 // Remember the beginning of the previous instruction, in case
1467 // it's followed by a flag-kill and a null-check. Happens on
1468 // Intel all the time, with add-to-memory kind of opcodes.
1469 previous_offset = current_offset;
1470 }
1471 }
1473 // Verify that there is sufficient space remaining
1474 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1475 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1476 C->record_failure("CodeCache is full");
1477 return;
1478 }
1480 // Save the offset for the listing
1481 #ifndef PRODUCT
1482 if (node_offsets && n->_idx < node_offset_limit)
1483 node_offsets[n->_idx] = cb->insts_size();
1484 #endif
1486 // "Normal" instruction case
1487 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1488 n->emit(*cb, _regalloc);
1489 current_offset = cb->insts_size();
1491 #ifdef ASSERT
1492 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1493 n->dump();
1494 assert(false, "wrong size of mach node");
1495 }
1496 #endif
1497 non_safepoints.observe_instruction(n, current_offset);
1499 // mcall is last "call" that can be a safepoint
1500 // record it so we can see if a poll will directly follow it
1501 // in which case we'll need a pad to make the PcDesc sites unique
1502 // see 5010568. This can be slightly inaccurate but conservative
1503 // in the case that return address is not actually at current_offset.
1504 // This is a small price to pay.
1506 if (is_mcall) {
1507 last_call_offset = current_offset;
1508 }
1510 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
1511 // Avoid back to back some instructions.
1512 last_avoid_back_to_back_offset = current_offset;
1513 }
1515 // See if this instruction has a delay slot
1516 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1517 assert(delay_slot != NULL, "expecting delay slot node");
1519 // Back up 1 instruction
1520 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1522 // Save the offset for the listing
1523 #ifndef PRODUCT
1524 if (node_offsets && delay_slot->_idx < node_offset_limit)
1525 node_offsets[delay_slot->_idx] = cb->insts_size();
1526 #endif
1528 // Support a SafePoint in the delay slot
1529 if (delay_slot->is_MachSafePoint()) {
1530 MachNode *mach = delay_slot->as_Mach();
1531 // !!!!! Stubs only need an oopmap right now, so bail out
1532 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1533 // Write the oopmap directly to the code blob??!!
1534 # ifdef ENABLE_ZAP_DEAD_LOCALS
1535 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1536 # endif
1537 delay_slot = NULL;
1538 continue;
1539 }
1541 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1542 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1543 adjusted_offset);
1544 // Generate an OopMap entry
1545 Process_OopMap_Node(mach, adjusted_offset);
1546 }
1548 // Insert the delay slot instruction
1549 delay_slot->emit(*cb, _regalloc);
1551 // Don't reuse it
1552 delay_slot = NULL;
1553 }
1555 } // End for all instructions in block
1557 // If the next block is the top of a loop, pad this block out to align
1558 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1559 if (i < nblocks-1) {
1560 Block *nb = _cfg->get_block(i + 1);
1561 int padding = nb->alignment_padding(current_offset);
1562 if( padding > 0 ) {
1563 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1564 block->insert_node(nop, block->number_of_nodes());
1565 _cfg->map_node_to_block(nop, block);
1566 nop->emit(*cb, _regalloc);
1567 current_offset = cb->insts_size();
1568 }
1569 }
1570 // Verify that the distance for generated before forward
1571 // short branches is still valid.
1572 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1574 // Save new block start offset
1575 blk_starts[i] = blk_offset;
1576 } // End of for all blocks
1577 blk_starts[nblocks] = current_offset;
1579 non_safepoints.flush_at_end();
1581 // Offset too large?
1582 if (failing()) return;
1584 // Define a pseudo-label at the end of the code
1585 MacroAssembler(cb).bind( blk_labels[nblocks] );
1587 // Compute the size of the first block
1588 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1590 assert(cb->insts_size() < 500000, "method is unreasonably large");
1592 #ifdef ASSERT
1593 for (uint i = 0; i < nblocks; i++) { // For all blocks
1594 if (jmp_target[i] != 0) {
1595 int br_size = jmp_size[i];
1596 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1597 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1598 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]);
1599 assert(false, "Displacement too large for short jmp");
1600 }
1601 }
1602 }
1603 #endif
1605 #ifndef PRODUCT
1606 // Information on the size of the method, without the extraneous code
1607 Scheduling::increment_method_size(cb->insts_size());
1608 #endif
1610 // ------------------
1611 // Fill in exception table entries.
1612 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1614 // Only java methods have exception handlers and deopt handlers
1615 if (_method) {
1616 // Emit the exception handler code.
1617 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1618 // Emit the deopt handler code.
1619 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1621 // Emit the MethodHandle deopt handler code (if required).
1622 if (has_method_handle_invokes()) {
1623 // We can use the same code as for the normal deopt handler, we
1624 // just need a different entry point address.
1625 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1626 }
1627 }
1629 // One last check for failed CodeBuffer::expand:
1630 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1631 C->record_failure("CodeCache is full");
1632 return;
1633 }
1635 #ifndef PRODUCT
1636 // Dump the assembly code, including basic-block numbers
1637 if (print_assembly()) {
1638 ttyLocker ttyl; // keep the following output all in one block
1639 if (!VMThread::should_terminate()) { // test this under the tty lock
1640 // This output goes directly to the tty, not the compiler log.
1641 // To enable tools to match it up with the compilation activity,
1642 // be sure to tag this tty output with the compile ID.
1643 if (xtty != NULL) {
1644 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1645 is_osr_compilation() ? " compile_kind='osr'" :
1646 "");
1647 }
1648 if (method() != NULL) {
1649 method()->print_metadata();
1650 }
1651 dump_asm(node_offsets, node_offset_limit);
1652 if (xtty != NULL) {
1653 xtty->tail("opto_assembly");
1654 }
1655 }
1656 }
1657 #endif
1659 }
1661 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1662 _inc_table.set_size(cnt);
1664 uint inct_cnt = 0;
1665 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1666 Block* block = _cfg->get_block(i);
1667 Node *n = NULL;
1668 int j;
1670 // Find the branch; ignore trailing NOPs.
1671 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1672 n = block->get_node(j);
1673 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1674 break;
1675 }
1676 }
1678 // If we didn't find anything, continue
1679 if (j < 0) {
1680 continue;
1681 }
1683 // Compute ExceptionHandlerTable subtable entry and add it
1684 // (skip empty blocks)
1685 if (n->is_Catch()) {
1687 // Get the offset of the return from the call
1688 uint call_return = call_returns[block->_pre_order];
1689 #ifdef ASSERT
1690 assert( call_return > 0, "no call seen for this basic block" );
1691 while (block->get_node(--j)->is_MachProj()) ;
1692 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1693 #endif
1694 // last instruction is a CatchNode, find it's CatchProjNodes
1695 int nof_succs = block->_num_succs;
1696 // allocate space
1697 GrowableArray<intptr_t> handler_bcis(nof_succs);
1698 GrowableArray<intptr_t> handler_pcos(nof_succs);
1699 // iterate through all successors
1700 for (int j = 0; j < nof_succs; j++) {
1701 Block* s = block->_succs[j];
1702 bool found_p = false;
1703 for (uint k = 1; k < s->num_preds(); k++) {
1704 Node* pk = s->pred(k);
1705 if (pk->is_CatchProj() && pk->in(0) == n) {
1706 const CatchProjNode* p = pk->as_CatchProj();
1707 found_p = true;
1708 // add the corresponding handler bci & pco information
1709 if (p->_con != CatchProjNode::fall_through_index) {
1710 // p leads to an exception handler (and is not fall through)
1711 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1712 // no duplicates, please
1713 if (!handler_bcis.contains(p->handler_bci())) {
1714 uint block_num = s->non_connector()->_pre_order;
1715 handler_bcis.append(p->handler_bci());
1716 handler_pcos.append(blk_labels[block_num].loc_pos());
1717 }
1718 }
1719 }
1720 }
1721 assert(found_p, "no matching predecessor found");
1722 // Note: Due to empty block removal, one block may have
1723 // several CatchProj inputs, from the same Catch.
1724 }
1726 // Set the offset of the return from the call
1727 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1728 continue;
1729 }
1731 // Handle implicit null exception table updates
1732 if (n->is_MachNullCheck()) {
1733 uint block_num = block->non_connector_successor(0)->_pre_order;
1734 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1735 continue;
1736 }
1737 } // End of for all blocks fill in exception table entries
1738 }
1740 // Static Variables
1741 #ifndef PRODUCT
1742 uint Scheduling::_total_nop_size = 0;
1743 uint Scheduling::_total_method_size = 0;
1744 uint Scheduling::_total_branches = 0;
1745 uint Scheduling::_total_unconditional_delays = 0;
1746 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1747 #endif
1749 // Initializer for class Scheduling
1751 Scheduling::Scheduling(Arena *arena, Compile &compile)
1752 : _arena(arena),
1753 _cfg(compile.cfg()),
1754 _regalloc(compile.regalloc()),
1755 _reg_node(arena),
1756 _bundle_instr_count(0),
1757 _bundle_cycle_number(0),
1758 _scheduled(arena),
1759 _available(arena),
1760 _next_node(NULL),
1761 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1762 _pinch_free_list(arena)
1763 #ifndef PRODUCT
1764 , _branches(0)
1765 , _unconditional_delays(0)
1766 #endif
1767 {
1768 // Create a MachNopNode
1769 _nop = new (&compile) MachNopNode();
1771 // Now that the nops are in the array, save the count
1772 // (but allow entries for the nops)
1773 _node_bundling_limit = compile.unique();
1774 uint node_max = _regalloc->node_regs_max_index();
1776 compile.set_node_bundling_limit(_node_bundling_limit);
1778 // This one is persistent within the Compile class
1779 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1781 // Allocate space for fixed-size arrays
1782 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1783 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1784 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1786 // Clear the arrays
1787 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1788 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1789 memset(_uses, 0, node_max * sizeof(short));
1790 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1792 // Clear the bundling information
1793 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1795 // Get the last node
1796 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1798 _next_node = block->get_node(block->number_of_nodes() - 1);
1799 }
1801 #ifndef PRODUCT
1802 // Scheduling destructor
1803 Scheduling::~Scheduling() {
1804 _total_branches += _branches;
1805 _total_unconditional_delays += _unconditional_delays;
1806 }
1807 #endif
1809 // Step ahead "i" cycles
1810 void Scheduling::step(uint i) {
1812 Bundle *bundle = node_bundling(_next_node);
1813 bundle->set_starts_bundle();
1815 // Update the bundle record, but leave the flags information alone
1816 if (_bundle_instr_count > 0) {
1817 bundle->set_instr_count(_bundle_instr_count);
1818 bundle->set_resources_used(_bundle_use.resourcesUsed());
1819 }
1821 // Update the state information
1822 _bundle_instr_count = 0;
1823 _bundle_cycle_number += i;
1824 _bundle_use.step(i);
1825 }
1827 void Scheduling::step_and_clear() {
1828 Bundle *bundle = node_bundling(_next_node);
1829 bundle->set_starts_bundle();
1831 // Update the bundle record
1832 if (_bundle_instr_count > 0) {
1833 bundle->set_instr_count(_bundle_instr_count);
1834 bundle->set_resources_used(_bundle_use.resourcesUsed());
1836 _bundle_cycle_number += 1;
1837 }
1839 // Clear the bundling information
1840 _bundle_instr_count = 0;
1841 _bundle_use.reset();
1843 memcpy(_bundle_use_elements,
1844 Pipeline_Use::elaborated_elements,
1845 sizeof(Pipeline_Use::elaborated_elements));
1846 }
1848 // Perform instruction scheduling and bundling over the sequence of
1849 // instructions in backwards order.
1850 void Compile::ScheduleAndBundle() {
1852 // Don't optimize this if it isn't a method
1853 if (!_method)
1854 return;
1856 // Don't optimize this if scheduling is disabled
1857 if (!do_scheduling())
1858 return;
1860 // Scheduling code works only with pairs (8 bytes) maximum.
1861 if (max_vector_size() > 8)
1862 return;
1864 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1866 // Create a data structure for all the scheduling information
1867 Scheduling scheduling(Thread::current()->resource_area(), *this);
1869 // Walk backwards over each basic block, computing the needed alignment
1870 // Walk over all the basic blocks
1871 scheduling.DoScheduling();
1872 }
1874 // Compute the latency of all the instructions. This is fairly simple,
1875 // because we already have a legal ordering. Walk over the instructions
1876 // from first to last, and compute the latency of the instruction based
1877 // on the latency of the preceding instruction(s).
1878 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1879 #ifndef PRODUCT
1880 if (_cfg->C->trace_opto_output())
1881 tty->print("# -> ComputeLocalLatenciesForward\n");
1882 #endif
1884 // Walk over all the schedulable instructions
1885 for( uint j=_bb_start; j < _bb_end; j++ ) {
1887 // This is a kludge, forcing all latency calculations to start at 1.
1888 // Used to allow latency 0 to force an instruction to the beginning
1889 // of the bb
1890 uint latency = 1;
1891 Node *use = bb->get_node(j);
1892 uint nlen = use->len();
1894 // Walk over all the inputs
1895 for ( uint k=0; k < nlen; k++ ) {
1896 Node *def = use->in(k);
1897 if (!def)
1898 continue;
1900 uint l = _node_latency[def->_idx] + use->latency(k);
1901 if (latency < l)
1902 latency = l;
1903 }
1905 _node_latency[use->_idx] = latency;
1907 #ifndef PRODUCT
1908 if (_cfg->C->trace_opto_output()) {
1909 tty->print("# latency %4d: ", latency);
1910 use->dump();
1911 }
1912 #endif
1913 }
1915 #ifndef PRODUCT
1916 if (_cfg->C->trace_opto_output())
1917 tty->print("# <- ComputeLocalLatenciesForward\n");
1918 #endif
1920 } // end ComputeLocalLatenciesForward
1922 // See if this node fits into the present instruction bundle
1923 bool Scheduling::NodeFitsInBundle(Node *n) {
1924 uint n_idx = n->_idx;
1926 // If this is the unconditional delay instruction, then it fits
1927 if (n == _unconditional_delay_slot) {
1928 #ifndef PRODUCT
1929 if (_cfg->C->trace_opto_output())
1930 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1931 #endif
1932 return (true);
1933 }
1935 // If the node cannot be scheduled this cycle, skip it
1936 if (_current_latency[n_idx] > _bundle_cycle_number) {
1937 #ifndef PRODUCT
1938 if (_cfg->C->trace_opto_output())
1939 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1940 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1941 #endif
1942 return (false);
1943 }
1945 const Pipeline *node_pipeline = n->pipeline();
1947 uint instruction_count = node_pipeline->instructionCount();
1948 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1949 instruction_count = 0;
1950 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1951 instruction_count++;
1953 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1954 #ifndef PRODUCT
1955 if (_cfg->C->trace_opto_output())
1956 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1957 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1958 #endif
1959 return (false);
1960 }
1962 // Don't allow non-machine nodes to be handled this way
1963 if (!n->is_Mach() && instruction_count == 0)
1964 return (false);
1966 // See if there is any overlap
1967 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1969 if (delay > 0) {
1970 #ifndef PRODUCT
1971 if (_cfg->C->trace_opto_output())
1972 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1973 #endif
1974 return false;
1975 }
1977 #ifndef PRODUCT
1978 if (_cfg->C->trace_opto_output())
1979 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1980 #endif
1982 return true;
1983 }
1985 Node * Scheduling::ChooseNodeToBundle() {
1986 uint siz = _available.size();
1988 if (siz == 0) {
1990 #ifndef PRODUCT
1991 if (_cfg->C->trace_opto_output())
1992 tty->print("# ChooseNodeToBundle: NULL\n");
1993 #endif
1994 return (NULL);
1995 }
1997 // Fast path, if only 1 instruction in the bundle
1998 if (siz == 1) {
1999 #ifndef PRODUCT
2000 if (_cfg->C->trace_opto_output()) {
2001 tty->print("# ChooseNodeToBundle (only 1): ");
2002 _available[0]->dump();
2003 }
2004 #endif
2005 return (_available[0]);
2006 }
2008 // Don't bother, if the bundle is already full
2009 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2010 for ( uint i = 0; i < siz; i++ ) {
2011 Node *n = _available[i];
2013 // Skip projections, we'll handle them another way
2014 if (n->is_Proj())
2015 continue;
2017 // This presupposed that instructions are inserted into the
2018 // available list in a legality order; i.e. instructions that
2019 // must be inserted first are at the head of the list
2020 if (NodeFitsInBundle(n)) {
2021 #ifndef PRODUCT
2022 if (_cfg->C->trace_opto_output()) {
2023 tty->print("# ChooseNodeToBundle: ");
2024 n->dump();
2025 }
2026 #endif
2027 return (n);
2028 }
2029 }
2030 }
2032 // Nothing fits in this bundle, choose the highest priority
2033 #ifndef PRODUCT
2034 if (_cfg->C->trace_opto_output()) {
2035 tty->print("# ChooseNodeToBundle: ");
2036 _available[0]->dump();
2037 }
2038 #endif
2040 return _available[0];
2041 }
2043 void Scheduling::AddNodeToAvailableList(Node *n) {
2044 assert( !n->is_Proj(), "projections never directly made available" );
2045 #ifndef PRODUCT
2046 if (_cfg->C->trace_opto_output()) {
2047 tty->print("# AddNodeToAvailableList: ");
2048 n->dump();
2049 }
2050 #endif
2052 int latency = _current_latency[n->_idx];
2054 // Insert in latency order (insertion sort)
2055 uint i;
2056 for ( i=0; i < _available.size(); i++ )
2057 if (_current_latency[_available[i]->_idx] > latency)
2058 break;
2060 // Special Check for compares following branches
2061 if( n->is_Mach() && _scheduled.size() > 0 ) {
2062 int op = n->as_Mach()->ideal_Opcode();
2063 Node *last = _scheduled[0];
2064 if( last->is_MachIf() && last->in(1) == n &&
2065 ( op == Op_CmpI ||
2066 op == Op_CmpU ||
2067 op == Op_CmpP ||
2068 op == Op_CmpF ||
2069 op == Op_CmpD ||
2070 op == Op_CmpL ) ) {
2072 // Recalculate position, moving to front of same latency
2073 for ( i=0 ; i < _available.size(); i++ )
2074 if (_current_latency[_available[i]->_idx] >= latency)
2075 break;
2076 }
2077 }
2079 // Insert the node in the available list
2080 _available.insert(i, n);
2082 #ifndef PRODUCT
2083 if (_cfg->C->trace_opto_output())
2084 dump_available();
2085 #endif
2086 }
2088 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2089 for ( uint i=0; i < n->len(); i++ ) {
2090 Node *def = n->in(i);
2091 if (!def) continue;
2092 if( def->is_Proj() ) // If this is a machine projection, then
2093 def = def->in(0); // propagate usage thru to the base instruction
2095 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2096 continue;
2097 }
2099 // Compute the latency
2100 uint l = _bundle_cycle_number + n->latency(i);
2101 if (_current_latency[def->_idx] < l)
2102 _current_latency[def->_idx] = l;
2104 // If this does not have uses then schedule it
2105 if ((--_uses[def->_idx]) == 0)
2106 AddNodeToAvailableList(def);
2107 }
2108 }
2110 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2111 #ifndef PRODUCT
2112 if (_cfg->C->trace_opto_output()) {
2113 tty->print("# AddNodeToBundle: ");
2114 n->dump();
2115 }
2116 #endif
2118 // Remove this from the available list
2119 uint i;
2120 for (i = 0; i < _available.size(); i++)
2121 if (_available[i] == n)
2122 break;
2123 assert(i < _available.size(), "entry in _available list not found");
2124 _available.remove(i);
2126 // See if this fits in the current bundle
2127 const Pipeline *node_pipeline = n->pipeline();
2128 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2130 // Check for instructions to be placed in the delay slot. We
2131 // do this before we actually schedule the current instruction,
2132 // because the delay slot follows the current instruction.
2133 if (Pipeline::_branch_has_delay_slot &&
2134 node_pipeline->hasBranchDelay() &&
2135 !_unconditional_delay_slot) {
2137 uint siz = _available.size();
2139 // Conditional branches can support an instruction that
2140 // is unconditionally executed and not dependent by the
2141 // branch, OR a conditionally executed instruction if
2142 // the branch is taken. In practice, this means that
2143 // the first instruction at the branch target is
2144 // copied to the delay slot, and the branch goes to
2145 // the instruction after that at the branch target
2146 if ( n->is_MachBranch() ) {
2148 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2149 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2151 #ifndef PRODUCT
2152 _branches++;
2153 #endif
2155 // At least 1 instruction is on the available list
2156 // that is not dependent on the branch
2157 for (uint i = 0; i < siz; i++) {
2158 Node *d = _available[i];
2159 const Pipeline *avail_pipeline = d->pipeline();
2161 // Don't allow safepoints in the branch shadow, that will
2162 // cause a number of difficulties
2163 if ( avail_pipeline->instructionCount() == 1 &&
2164 !avail_pipeline->hasMultipleBundles() &&
2165 !avail_pipeline->hasBranchDelay() &&
2166 Pipeline::instr_has_unit_size() &&
2167 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2168 NodeFitsInBundle(d) &&
2169 !node_bundling(d)->used_in_delay()) {
2171 if (d->is_Mach() && !d->is_MachSafePoint()) {
2172 // A node that fits in the delay slot was found, so we need to
2173 // set the appropriate bits in the bundle pipeline information so
2174 // that it correctly indicates resource usage. Later, when we
2175 // attempt to add this instruction to the bundle, we will skip
2176 // setting the resource usage.
2177 _unconditional_delay_slot = d;
2178 node_bundling(n)->set_use_unconditional_delay();
2179 node_bundling(d)->set_used_in_unconditional_delay();
2180 _bundle_use.add_usage(avail_pipeline->resourceUse());
2181 _current_latency[d->_idx] = _bundle_cycle_number;
2182 _next_node = d;
2183 ++_bundle_instr_count;
2184 #ifndef PRODUCT
2185 _unconditional_delays++;
2186 #endif
2187 break;
2188 }
2189 }
2190 }
2191 }
2193 // No delay slot, add a nop to the usage
2194 if (!_unconditional_delay_slot) {
2195 // See if adding an instruction in the delay slot will overflow
2196 // the bundle.
2197 if (!NodeFitsInBundle(_nop)) {
2198 #ifndef PRODUCT
2199 if (_cfg->C->trace_opto_output())
2200 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2201 #endif
2202 step(1);
2203 }
2205 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2206 _next_node = _nop;
2207 ++_bundle_instr_count;
2208 }
2210 // See if the instruction in the delay slot requires a
2211 // step of the bundles
2212 if (!NodeFitsInBundle(n)) {
2213 #ifndef PRODUCT
2214 if (_cfg->C->trace_opto_output())
2215 tty->print("# *** STEP(branch won't fit) ***\n");
2216 #endif
2217 // Update the state information
2218 _bundle_instr_count = 0;
2219 _bundle_cycle_number += 1;
2220 _bundle_use.step(1);
2221 }
2222 }
2224 // Get the number of instructions
2225 uint instruction_count = node_pipeline->instructionCount();
2226 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2227 instruction_count = 0;
2229 // Compute the latency information
2230 uint delay = 0;
2232 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2233 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2234 if (relative_latency < 0)
2235 relative_latency = 0;
2237 delay = _bundle_use.full_latency(relative_latency, node_usage);
2239 // Does not fit in this bundle, start a new one
2240 if (delay > 0) {
2241 step(delay);
2243 #ifndef PRODUCT
2244 if (_cfg->C->trace_opto_output())
2245 tty->print("# *** STEP(%d) ***\n", delay);
2246 #endif
2247 }
2248 }
2250 // If this was placed in the delay slot, ignore it
2251 if (n != _unconditional_delay_slot) {
2253 if (delay == 0) {
2254 if (node_pipeline->hasMultipleBundles()) {
2255 #ifndef PRODUCT
2256 if (_cfg->C->trace_opto_output())
2257 tty->print("# *** STEP(multiple instructions) ***\n");
2258 #endif
2259 step(1);
2260 }
2262 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2263 #ifndef PRODUCT
2264 if (_cfg->C->trace_opto_output())
2265 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2266 instruction_count + _bundle_instr_count,
2267 Pipeline::_max_instrs_per_cycle);
2268 #endif
2269 step(1);
2270 }
2271 }
2273 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2274 _bundle_instr_count++;
2276 // Set the node's latency
2277 _current_latency[n->_idx] = _bundle_cycle_number;
2279 // Now merge the functional unit information
2280 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2281 _bundle_use.add_usage(node_usage);
2283 // Increment the number of instructions in this bundle
2284 _bundle_instr_count += instruction_count;
2286 // Remember this node for later
2287 if (n->is_Mach())
2288 _next_node = n;
2289 }
2291 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2292 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2293 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2294 // into the block. All other scheduled nodes get put in the schedule here.
2295 int op = n->Opcode();
2296 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2297 (op != Op_Node && // Not an unused antidepedence node and
2298 // not an unallocated boxlock
2299 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2301 // Push any trailing projections
2302 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2303 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2304 Node *foi = n->fast_out(i);
2305 if( foi->is_Proj() )
2306 _scheduled.push(foi);
2307 }
2308 }
2310 // Put the instruction in the schedule list
2311 _scheduled.push(n);
2312 }
2314 #ifndef PRODUCT
2315 if (_cfg->C->trace_opto_output())
2316 dump_available();
2317 #endif
2319 // Walk all the definitions, decrementing use counts, and
2320 // if a definition has a 0 use count, place it in the available list.
2321 DecrementUseCounts(n,bb);
2322 }
2324 // This method sets the use count within a basic block. We will ignore all
2325 // uses outside the current basic block. As we are doing a backwards walk,
2326 // any node we reach that has a use count of 0 may be scheduled. This also
2327 // avoids the problem of cyclic references from phi nodes, as long as phi
2328 // nodes are at the front of the basic block. This method also initializes
2329 // the available list to the set of instructions that have no uses within this
2330 // basic block.
2331 void Scheduling::ComputeUseCount(const Block *bb) {
2332 #ifndef PRODUCT
2333 if (_cfg->C->trace_opto_output())
2334 tty->print("# -> ComputeUseCount\n");
2335 #endif
2337 // Clear the list of available and scheduled instructions, just in case
2338 _available.clear();
2339 _scheduled.clear();
2341 // No delay slot specified
2342 _unconditional_delay_slot = NULL;
2344 #ifdef ASSERT
2345 for( uint i=0; i < bb->number_of_nodes(); i++ )
2346 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2347 #endif
2349 // Force the _uses count to never go to zero for unscheduable pieces
2350 // of the block
2351 for( uint k = 0; k < _bb_start; k++ )
2352 _uses[bb->get_node(k)->_idx] = 1;
2353 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2354 _uses[bb->get_node(l)->_idx] = 1;
2356 // Iterate backwards over the instructions in the block. Don't count the
2357 // branch projections at end or the block header instructions.
2358 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2359 Node *n = bb->get_node(j);
2360 if( n->is_Proj() ) continue; // Projections handled another way
2362 // Account for all uses
2363 for ( uint k = 0; k < n->len(); k++ ) {
2364 Node *inp = n->in(k);
2365 if (!inp) continue;
2366 assert(inp != n, "no cycles allowed" );
2367 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2368 if (inp->is_Proj()) { // Skip through Proj's
2369 inp = inp->in(0);
2370 }
2371 ++_uses[inp->_idx]; // Count 1 block-local use
2372 }
2373 }
2375 // If this instruction has a 0 use count, then it is available
2376 if (!_uses[n->_idx]) {
2377 _current_latency[n->_idx] = _bundle_cycle_number;
2378 AddNodeToAvailableList(n);
2379 }
2381 #ifndef PRODUCT
2382 if (_cfg->C->trace_opto_output()) {
2383 tty->print("# uses: %3d: ", _uses[n->_idx]);
2384 n->dump();
2385 }
2386 #endif
2387 }
2389 #ifndef PRODUCT
2390 if (_cfg->C->trace_opto_output())
2391 tty->print("# <- ComputeUseCount\n");
2392 #endif
2393 }
2395 // This routine performs scheduling on each basic block in reverse order,
2396 // using instruction latencies and taking into account function unit
2397 // availability.
2398 void Scheduling::DoScheduling() {
2399 #ifndef PRODUCT
2400 if (_cfg->C->trace_opto_output())
2401 tty->print("# -> DoScheduling\n");
2402 #endif
2404 Block *succ_bb = NULL;
2405 Block *bb;
2407 // Walk over all the basic blocks in reverse order
2408 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2409 bb = _cfg->get_block(i);
2411 #ifndef PRODUCT
2412 if (_cfg->C->trace_opto_output()) {
2413 tty->print("# Schedule BB#%03d (initial)\n", i);
2414 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2415 bb->get_node(j)->dump();
2416 }
2417 }
2418 #endif
2420 // On the head node, skip processing
2421 if (bb == _cfg->get_root_block()) {
2422 continue;
2423 }
2425 // Skip empty, connector blocks
2426 if (bb->is_connector())
2427 continue;
2429 // If the following block is not the sole successor of
2430 // this one, then reset the pipeline information
2431 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2432 #ifndef PRODUCT
2433 if (_cfg->C->trace_opto_output()) {
2434 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2435 _next_node->_idx, _bundle_instr_count);
2436 }
2437 #endif
2438 step_and_clear();
2439 }
2441 // Leave untouched the starting instruction, any Phis, a CreateEx node
2442 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2443 _bb_end = bb->number_of_nodes()-1;
2444 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2445 Node *n = bb->get_node(_bb_start);
2446 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2447 // Also, MachIdealNodes do not get scheduled
2448 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2449 MachNode *mach = n->as_Mach();
2450 int iop = mach->ideal_Opcode();
2451 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2452 if( iop == Op_Con ) continue; // Do not schedule Top
2453 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2454 mach->pipeline() == MachNode::pipeline_class() &&
2455 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2456 continue;
2457 break; // Funny loop structure to be sure...
2458 }
2459 // Compute last "interesting" instruction in block - last instruction we
2460 // might schedule. _bb_end points just after last schedulable inst. We
2461 // normally schedule conditional branches (despite them being forced last
2462 // in the block), because they have delay slots we can fill. Calls all
2463 // have their delay slots filled in the template expansions, so we don't
2464 // bother scheduling them.
2465 Node *last = bb->get_node(_bb_end);
2466 // Ignore trailing NOPs.
2467 while (_bb_end > 0 && last->is_Mach() &&
2468 last->as_Mach()->ideal_Opcode() == Op_Con) {
2469 last = bb->get_node(--_bb_end);
2470 }
2471 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2472 if( last->is_Catch() ||
2473 // Exclude unreachable path case when Halt node is in a separate block.
2474 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2475 // There must be a prior call. Skip it.
2476 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2477 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2478 }
2479 } else if( last->is_MachNullCheck() ) {
2480 // Backup so the last null-checked memory instruction is
2481 // outside the schedulable range. Skip over the nullcheck,
2482 // projection, and the memory nodes.
2483 Node *mem = last->in(1);
2484 do {
2485 _bb_end--;
2486 } while (mem != bb->get_node(_bb_end));
2487 } else {
2488 // Set _bb_end to point after last schedulable inst.
2489 _bb_end++;
2490 }
2492 assert( _bb_start <= _bb_end, "inverted block ends" );
2494 // Compute the register antidependencies for the basic block
2495 ComputeRegisterAntidependencies(bb);
2496 if (_cfg->C->failing()) return; // too many D-U pinch points
2498 // Compute intra-bb latencies for the nodes
2499 ComputeLocalLatenciesForward(bb);
2501 // Compute the usage within the block, and set the list of all nodes
2502 // in the block that have no uses within the block.
2503 ComputeUseCount(bb);
2505 // Schedule the remaining instructions in the block
2506 while ( _available.size() > 0 ) {
2507 Node *n = ChooseNodeToBundle();
2508 guarantee(n != NULL, "no nodes available");
2509 AddNodeToBundle(n,bb);
2510 }
2512 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2513 #ifdef ASSERT
2514 for( uint l = _bb_start; l < _bb_end; l++ ) {
2515 Node *n = bb->get_node(l);
2516 uint m;
2517 for( m = 0; m < _bb_end-_bb_start; m++ )
2518 if( _scheduled[m] == n )
2519 break;
2520 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2521 }
2522 #endif
2524 // Now copy the instructions (in reverse order) back to the block
2525 for ( uint k = _bb_start; k < _bb_end; k++ )
2526 bb->map_node(_scheduled[_bb_end-k-1], k);
2528 #ifndef PRODUCT
2529 if (_cfg->C->trace_opto_output()) {
2530 tty->print("# Schedule BB#%03d (final)\n", i);
2531 uint current = 0;
2532 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2533 Node *n = bb->get_node(j);
2534 if( valid_bundle_info(n) ) {
2535 Bundle *bundle = node_bundling(n);
2536 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2537 tty->print("*** Bundle: ");
2538 bundle->dump();
2539 }
2540 n->dump();
2541 }
2542 }
2543 }
2544 #endif
2545 #ifdef ASSERT
2546 verify_good_schedule(bb,"after block local scheduling");
2547 #endif
2548 }
2550 #ifndef PRODUCT
2551 if (_cfg->C->trace_opto_output())
2552 tty->print("# <- DoScheduling\n");
2553 #endif
2555 // Record final node-bundling array location
2556 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2558 } // end DoScheduling
2560 // Verify that no live-range used in the block is killed in the block by a
2561 // wrong DEF. This doesn't verify live-ranges that span blocks.
2563 // Check for edge existence. Used to avoid adding redundant precedence edges.
2564 static bool edge_from_to( Node *from, Node *to ) {
2565 for( uint i=0; i<from->len(); i++ )
2566 if( from->in(i) == to )
2567 return true;
2568 return false;
2569 }
2571 #ifdef ASSERT
2572 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2573 // Check for bad kills
2574 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2575 Node *prior_use = _reg_node[def];
2576 if( prior_use && !edge_from_to(prior_use,n) ) {
2577 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2578 n->dump();
2579 tty->print_cr("...");
2580 prior_use->dump();
2581 assert(edge_from_to(prior_use,n),msg);
2582 }
2583 _reg_node.map(def,NULL); // Kill live USEs
2584 }
2585 }
2587 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2589 // Zap to something reasonable for the verify code
2590 _reg_node.clear();
2592 // Walk over the block backwards. Check to make sure each DEF doesn't
2593 // kill a live value (other than the one it's supposed to). Add each
2594 // USE to the live set.
2595 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2596 Node *n = b->get_node(i);
2597 int n_op = n->Opcode();
2598 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2599 // Fat-proj kills a slew of registers
2600 RegMask rm = n->out_RegMask();// Make local copy
2601 while( rm.is_NotEmpty() ) {
2602 OptoReg::Name kill = rm.find_first_elem();
2603 rm.Remove(kill);
2604 verify_do_def( n, kill, msg );
2605 }
2606 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2607 // Get DEF'd registers the normal way
2608 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2609 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2610 }
2612 // Now make all USEs live
2613 for( uint i=1; i<n->req(); i++ ) {
2614 Node *def = n->in(i);
2615 assert(def != 0, "input edge required");
2616 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2617 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2618 if( OptoReg::is_valid(reg_lo) ) {
2619 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2620 _reg_node.map(reg_lo,n);
2621 }
2622 if( OptoReg::is_valid(reg_hi) ) {
2623 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2624 _reg_node.map(reg_hi,n);
2625 }
2626 }
2628 }
2630 // Zap to something reasonable for the Antidependence code
2631 _reg_node.clear();
2632 }
2633 #endif
2635 // Conditionally add precedence edges. Avoid putting edges on Projs.
2636 static void add_prec_edge_from_to( Node *from, Node *to ) {
2637 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2638 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2639 from = from->in(0);
2640 }
2641 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2642 !edge_from_to( from, to ) ) // Avoid duplicate edge
2643 from->add_prec(to);
2644 }
2646 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2647 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2648 return;
2650 Node *pinch = _reg_node[def_reg]; // Get pinch point
2651 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2652 is_def ) { // Check for a true def (not a kill)
2653 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2654 return;
2655 }
2657 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2658 debug_only( def = (Node*)0xdeadbeef; )
2660 // After some number of kills there _may_ be a later def
2661 Node *later_def = NULL;
2663 // Finding a kill requires a real pinch-point.
2664 // Check for not already having a pinch-point.
2665 // Pinch points are Op_Node's.
2666 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2667 later_def = pinch; // Must be def/kill as optimistic pinch-point
2668 if ( _pinch_free_list.size() > 0) {
2669 pinch = _pinch_free_list.pop();
2670 } else {
2671 pinch = new (_cfg->C) Node(1); // Pinch point to-be
2672 }
2673 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2674 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2675 return;
2676 }
2677 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2678 _reg_node.map(def_reg,pinch); // Record pinch-point
2679 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2680 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2681 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2682 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2683 later_def = NULL; // and no later def
2684 }
2685 pinch->set_req(0,later_def); // Hook later def so we can find it
2686 } else { // Else have valid pinch point
2687 if( pinch->in(0) ) // If there is a later-def
2688 later_def = pinch->in(0); // Get it
2689 }
2691 // Add output-dependence edge from later def to kill
2692 if( later_def ) // If there is some original def
2693 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2695 // See if current kill is also a use, and so is forced to be the pinch-point.
2696 if( pinch->Opcode() == Op_Node ) {
2697 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2698 for( uint i=1; i<uses->req(); i++ ) {
2699 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2700 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2701 // Yes, found a use/kill pinch-point
2702 pinch->set_req(0,NULL); //
2703 pinch->replace_by(kill); // Move anti-dep edges up
2704 pinch = kill;
2705 _reg_node.map(def_reg,pinch);
2706 return;
2707 }
2708 }
2709 }
2711 // Add edge from kill to pinch-point
2712 add_prec_edge_from_to(kill,pinch);
2713 }
2715 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2716 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2717 return;
2718 Node *pinch = _reg_node[use_reg]; // Get pinch point
2719 // Check for no later def_reg/kill in block
2720 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2721 // Use has to be block-local as well
2722 _cfg->get_block_for_node(use) == b) {
2723 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2724 pinch->req() == 1 ) { // pinch not yet in block?
2725 pinch->del_req(0); // yank pointer to later-def, also set flag
2726 // Insert the pinch-point in the block just after the last use
2727 b->insert_node(pinch, b->find_node(use) + 1);
2728 _bb_end++; // Increase size scheduled region in block
2729 }
2731 add_prec_edge_from_to(pinch,use);
2732 }
2733 }
2735 // We insert antidependences between the reads and following write of
2736 // allocated registers to prevent illegal code motion. Hopefully, the
2737 // number of added references should be fairly small, especially as we
2738 // are only adding references within the current basic block.
2739 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2741 #ifdef ASSERT
2742 verify_good_schedule(b,"before block local scheduling");
2743 #endif
2745 // A valid schedule, for each register independently, is an endless cycle
2746 // of: a def, then some uses (connected to the def by true dependencies),
2747 // then some kills (defs with no uses), finally the cycle repeats with a new
2748 // def. The uses are allowed to float relative to each other, as are the
2749 // kills. No use is allowed to slide past a kill (or def). This requires
2750 // antidependencies between all uses of a single def and all kills that
2751 // follow, up to the next def. More edges are redundant, because later defs
2752 // & kills are already serialized with true or antidependencies. To keep
2753 // the edge count down, we add a 'pinch point' node if there's more than
2754 // one use or more than one kill/def.
2756 // We add dependencies in one bottom-up pass.
2758 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2760 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2761 // register. If not, we record the DEF/KILL in _reg_node, the
2762 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2763 // "pinch point", a new Node that's in the graph but not in the block.
2764 // We put edges from the prior and current DEF/KILLs to the pinch point.
2765 // We put the pinch point in _reg_node. If there's already a pinch point
2766 // we merely add an edge from the current DEF/KILL to the pinch point.
2768 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2769 // put an edge from the pinch point to the USE.
2771 // To be expedient, the _reg_node array is pre-allocated for the whole
2772 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2773 // or a valid def/kill/pinch-point, or a leftover node from some prior
2774 // block. Leftover node from some prior block is treated like a NULL (no
2775 // prior def, so no anti-dependence needed). Valid def is distinguished by
2776 // it being in the current block.
2777 bool fat_proj_seen = false;
2778 uint last_safept = _bb_end-1;
2779 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2780 Node* last_safept_node = end_node;
2781 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2782 Node *n = b->get_node(i);
2783 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2784 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2785 // Fat-proj kills a slew of registers
2786 // This can add edges to 'n' and obscure whether or not it was a def,
2787 // hence the is_def flag.
2788 fat_proj_seen = true;
2789 RegMask rm = n->out_RegMask();// Make local copy
2790 while( rm.is_NotEmpty() ) {
2791 OptoReg::Name kill = rm.find_first_elem();
2792 rm.Remove(kill);
2793 anti_do_def( b, n, kill, is_def );
2794 }
2795 } else {
2796 // Get DEF'd registers the normal way
2797 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2798 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2799 }
2801 // Kill projections on a branch should appear to occur on the
2802 // branch, not afterwards, so grab the masks from the projections
2803 // and process them.
2804 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2805 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2806 Node* use = n->fast_out(i);
2807 if (use->is_Proj()) {
2808 RegMask rm = use->out_RegMask();// Make local copy
2809 while( rm.is_NotEmpty() ) {
2810 OptoReg::Name kill = rm.find_first_elem();
2811 rm.Remove(kill);
2812 anti_do_def( b, n, kill, false );
2813 }
2814 }
2815 }
2816 }
2818 // Check each register used by this instruction for a following DEF/KILL
2819 // that must occur afterward and requires an anti-dependence edge.
2820 for( uint j=0; j<n->req(); j++ ) {
2821 Node *def = n->in(j);
2822 if( def ) {
2823 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2824 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2825 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2826 }
2827 }
2828 // Do not allow defs of new derived values to float above GC
2829 // points unless the base is definitely available at the GC point.
2831 Node *m = b->get_node(i);
2833 // Add precedence edge from following safepoint to use of derived pointer
2834 if( last_safept_node != end_node &&
2835 m != last_safept_node) {
2836 for (uint k = 1; k < m->req(); k++) {
2837 const Type *t = m->in(k)->bottom_type();
2838 if( t->isa_oop_ptr() &&
2839 t->is_ptr()->offset() != 0 ) {
2840 last_safept_node->add_prec( m );
2841 break;
2842 }
2843 }
2844 }
2846 if( n->jvms() ) { // Precedence edge from derived to safept
2847 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2848 if( b->get_node(last_safept) != last_safept_node ) {
2849 last_safept = b->find_node(last_safept_node);
2850 }
2851 for( uint j=last_safept; j > i; j-- ) {
2852 Node *mach = b->get_node(j);
2853 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2854 mach->add_prec( n );
2855 }
2856 last_safept = i;
2857 last_safept_node = m;
2858 }
2859 }
2861 if (fat_proj_seen) {
2862 // Garbage collect pinch nodes that were not consumed.
2863 // They are usually created by a fat kill MachProj for a call.
2864 garbage_collect_pinch_nodes();
2865 }
2866 }
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 #ifndef PRODUCT
2944 void Scheduling::dump_available() const {
2945 tty->print("#Availist ");
2946 for (uint i = 0; i < _available.size(); i++)
2947 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2948 tty->cr();
2949 }
2951 // Print Scheduling Statistics
2952 void Scheduling::print_statistics() {
2953 // Print the size added by nops for bundling
2954 tty->print("Nops added %d bytes to total of %d bytes",
2955 _total_nop_size, _total_method_size);
2956 if (_total_method_size > 0)
2957 tty->print(", for %.2f%%",
2958 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2959 tty->print("\n");
2961 // Print the number of branch shadows filled
2962 if (Pipeline::_branch_has_delay_slot) {
2963 tty->print("Of %d branches, %d had unconditional delay slots filled",
2964 _total_branches, _total_unconditional_delays);
2965 if (_total_branches > 0)
2966 tty->print(", for %.2f%%",
2967 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2968 tty->print("\n");
2969 }
2971 uint total_instructions = 0, total_bundles = 0;
2973 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2974 uint bundle_count = _total_instructions_per_bundle[i];
2975 total_instructions += bundle_count * i;
2976 total_bundles += bundle_count;
2977 }
2979 if (total_bundles > 0)
2980 tty->print("Average ILP (excluding nops) is %.2f\n",
2981 ((double)total_instructions) / ((double)total_bundles));
2982 }
2983 #endif