Fri, 25 May 2012 07:53:11 -0700
7170463: C2 should recognize "obj.getClass() == A.class" code pattern
Summary: optimize this code pattern obj.getClass() == A.class.
Reviewed-by: jrose, kvn
Contributed-by: Krystal Mok <sajia@taobao.com>
1 /*
2 * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "code/debugInfo.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/oopMap.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/locknode.hpp"
35 #include "opto/machnode.hpp"
36 #include "opto/output.hpp"
37 #include "opto/regalloc.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
40 #include "opto/type.hpp"
41 #include "runtime/handles.inline.hpp"
42 #include "utilities/xmlstream.hpp"
44 extern uint size_java_to_interp();
45 extern uint reloc_java_to_interp();
46 extern uint size_exception_handler();
47 extern uint size_deopt_handler();
49 #ifndef PRODUCT
50 #define DEBUG_ARG(x) , x
51 #else
52 #define DEBUG_ARG(x)
53 #endif
55 extern int emit_exception_handler(CodeBuffer &cbuf);
56 extern int emit_deopt_handler(CodeBuffer &cbuf);
58 //------------------------------Output-----------------------------------------
59 // Convert Nodes to instruction bits and pass off to the VM
60 void Compile::Output() {
61 // RootNode goes
62 assert( _cfg->_broot->_nodes.size() == 0, "" );
64 // The number of new nodes (mostly MachNop) is proportional to
65 // the number of java calls and inner loops which are aligned.
66 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
67 C->inner_loops()*(OptoLoopAlignment-1)),
68 "out of nodes before code generation" ) ) {
69 return;
70 }
71 // Make sure I can find the Start Node
72 Block_Array& bbs = _cfg->_bbs;
73 Block *entry = _cfg->_blocks[1];
74 Block *broot = _cfg->_broot;
76 const StartNode *start = entry->_nodes[0]->as_Start();
78 // Replace StartNode with prolog
79 MachPrologNode *prolog = new (this) MachPrologNode();
80 entry->_nodes.map( 0, prolog );
81 bbs.map( prolog->_idx, entry );
82 bbs.map( start->_idx, NULL ); // start is no longer in any block
84 // Virtual methods need an unverified entry point
86 if( is_osr_compilation() ) {
87 if( PoisonOSREntry ) {
88 // TODO: Should use a ShouldNotReachHereNode...
89 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
90 }
91 } else {
92 if( _method && !_method->flags().is_static() ) {
93 // Insert unvalidated entry point
94 _cfg->insert( broot, 0, new (this) MachUEPNode() );
95 }
97 }
100 // Break before main entry point
101 if( (_method && _method->break_at_execute())
102 #ifndef PRODUCT
103 ||(OptoBreakpoint && is_method_compilation())
104 ||(OptoBreakpointOSR && is_osr_compilation())
105 ||(OptoBreakpointC2R && !_method)
106 #endif
107 ) {
108 // checking for _method means that OptoBreakpoint does not apply to
109 // runtime stubs or frame converters
110 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
111 }
113 // Insert epilogs before every return
114 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
115 Block *b = _cfg->_blocks[i];
116 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point?
117 Node *m = b->end();
118 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) {
119 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
120 b->add_inst( epilog );
121 bbs.map(epilog->_idx, b);
122 //_regalloc->set_bad(epilog->_idx); // Already initialized this way.
123 }
124 }
125 }
127 # ifdef ENABLE_ZAP_DEAD_LOCALS
128 if ( ZapDeadCompiledLocals ) Insert_zap_nodes();
129 # endif
131 uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
132 blk_starts[0] = 0;
134 // Initialize code buffer and process short branches.
135 CodeBuffer* cb = init_buffer(blk_starts);
137 if (cb == NULL || failing()) return;
139 ScheduleAndBundle();
141 #ifndef PRODUCT
142 if (trace_opto_output()) {
143 tty->print("\n---- After ScheduleAndBundle ----\n");
144 for (uint i = 0; i < _cfg->_num_blocks; i++) {
145 tty->print("\nBB#%03d:\n", i);
146 Block *bb = _cfg->_blocks[i];
147 for (uint j = 0; j < bb->_nodes.size(); j++) {
148 Node *n = bb->_nodes[j];
149 OptoReg::Name reg = _regalloc->get_reg_first(n);
150 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
151 n->dump();
152 }
153 }
154 }
155 #endif
157 if (failing()) return;
159 BuildOopMaps();
161 if (failing()) return;
163 fill_buffer(cb, blk_starts);
164 }
166 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
167 // Determine if we need to generate a stack overflow check.
168 // Do it if the method is not a stub function and
169 // has java calls or has frame size > vm_page_size/8.
170 return (UseStackBanging && stub_function() == NULL &&
171 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
172 }
174 bool Compile::need_register_stack_bang() const {
175 // Determine if we need to generate a register stack overflow check.
176 // This is only used on architectures which have split register
177 // and memory stacks (ie. IA64).
178 // Bang if the method is not a stub function and has java calls
179 return (stub_function() == NULL && has_java_calls());
180 }
182 # ifdef ENABLE_ZAP_DEAD_LOCALS
185 // In order to catch compiler oop-map bugs, we have implemented
186 // a debugging mode called ZapDeadCompilerLocals.
187 // This mode causes the compiler to insert a call to a runtime routine,
188 // "zap_dead_locals", right before each place in compiled code
189 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
190 // The runtime routine checks that locations mapped as oops are really
191 // oops, that locations mapped as values do not look like oops,
192 // and that locations mapped as dead are not used later
193 // (by zapping them to an invalid address).
195 int Compile::_CompiledZap_count = 0;
197 void Compile::Insert_zap_nodes() {
198 bool skip = false;
201 // Dink with static counts because code code without the extra
202 // runtime calls is MUCH faster for debugging purposes
204 if ( CompileZapFirst == 0 ) ; // nothing special
205 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
206 else if ( CompileZapFirst == CompiledZap_count() )
207 warning("starting zap compilation after skipping");
209 if ( CompileZapLast == -1 ) ; // nothing special
210 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
211 else if ( CompileZapLast == CompiledZap_count() )
212 warning("about to compile last zap");
214 ++_CompiledZap_count; // counts skipped zaps, too
216 if ( skip ) return;
219 if ( _method == NULL )
220 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
222 // Insert call to zap runtime stub before every node with an oop map
223 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
224 Block *b = _cfg->_blocks[i];
225 for ( uint j = 0; j < b->_nodes.size(); ++j ) {
226 Node *n = b->_nodes[j];
228 // Determining if we should insert a zap-a-lot node in output.
229 // We do that for all nodes that has oopmap info, except for calls
230 // to allocation. Calls to allocation passes in the old top-of-eden pointer
231 // and expect the C code to reset it. Hence, there can be no safepoints between
232 // the inlined-allocation and the call to new_Java, etc.
233 // We also cannot zap monitor calls, as they must hold the microlock
234 // during the call to Zap, which also wants to grab the microlock.
235 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
236 if ( insert ) { // it is MachSafePoint
237 if ( !n->is_MachCall() ) {
238 insert = false;
239 } else if ( n->is_MachCall() ) {
240 MachCallNode* call = n->as_MachCall();
241 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
242 call->entry_point() == OptoRuntime::new_array_Java() ||
243 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
244 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
245 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
246 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
247 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
248 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
249 ) {
250 insert = false;
251 }
252 }
253 if (insert) {
254 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
255 b->_nodes.insert( j, zap );
256 _cfg->_bbs.map( zap->_idx, b );
257 ++j;
258 }
259 }
260 }
261 }
262 }
265 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
266 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
267 CallStaticJavaNode* ideal_node =
268 new (this, tf->domain()->cnt()) CallStaticJavaNode( tf,
269 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
270 "call zap dead locals stub", 0, TypePtr::BOTTOM);
271 // We need to copy the OopMap from the site we're zapping at.
272 // We have to make a copy, because the zap site might not be
273 // a call site, and zap_dead is a call site.
274 OopMap* clone = node_to_check->oop_map()->deep_copy();
276 // Add the cloned OopMap to the zap node
277 ideal_node->set_oop_map(clone);
278 return _matcher->match_sfpt(ideal_node);
279 }
281 //------------------------------is_node_getting_a_safepoint--------------------
282 bool Compile::is_node_getting_a_safepoint( Node* n) {
283 // This code duplicates the logic prior to the call of add_safepoint
284 // below in this file.
285 if( n->is_MachSafePoint() ) return true;
286 return false;
287 }
289 # endif // ENABLE_ZAP_DEAD_LOCALS
291 //------------------------------compute_loop_first_inst_sizes------------------
292 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
293 // of a loop. When aligning a loop we need to provide enough instructions
294 // in cpu's fetch buffer to feed decoders. The loop alignment could be
295 // avoided if we have enough instructions in fetch buffer at the head of a loop.
296 // By default, the size is set to 999999 by Block's constructor so that
297 // a loop will be aligned if the size is not reset here.
298 //
299 // Note: Mach instructions could contain several HW instructions
300 // so the size is estimated only.
301 //
302 void Compile::compute_loop_first_inst_sizes() {
303 // The next condition is used to gate the loop alignment optimization.
304 // Don't aligned a loop if there are enough instructions at the head of a loop
305 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
306 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
307 // equal to 11 bytes which is the largest address NOP instruction.
308 if( MaxLoopPad < OptoLoopAlignment-1 ) {
309 uint last_block = _cfg->_num_blocks-1;
310 for( uint i=1; i <= last_block; i++ ) {
311 Block *b = _cfg->_blocks[i];
312 // Check the first loop's block which requires an alignment.
313 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {
314 uint sum_size = 0;
315 uint inst_cnt = NumberOfLoopInstrToAlign;
316 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
318 // Check subsequent fallthrough blocks if the loop's first
319 // block(s) does not have enough instructions.
320 Block *nb = b;
321 while( inst_cnt > 0 &&
322 i < last_block &&
323 !_cfg->_blocks[i+1]->has_loop_alignment() &&
324 !nb->has_successor(b) ) {
325 i++;
326 nb = _cfg->_blocks[i];
327 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
328 } // while( inst_cnt > 0 && i < last_block )
330 b->set_first_inst_size(sum_size);
331 } // f( b->head()->is_Loop() )
332 } // for( i <= last_block )
333 } // if( MaxLoopPad < OptoLoopAlignment-1 )
334 }
336 //----------------------shorten_branches---------------------------------------
337 // The architecture description provides short branch variants for some long
338 // branch instructions. Replace eligible long branches with short branches.
339 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
341 // ------------------
342 // Compute size of each block, method size, and relocation information size
343 uint nblocks = _cfg->_num_blocks;
345 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
346 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
347 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
348 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
349 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
351 bool has_short_branch_candidate = false;
353 // Initialize the sizes to 0
354 code_size = 0; // Size in bytes of generated code
355 stub_size = 0; // Size in bytes of all stub entries
356 // Size in bytes of all relocation entries, including those in local stubs.
357 // Start with 2-bytes of reloc info for the unvalidated entry point
358 reloc_size = 1; // Number of relocation entries
360 // Make three passes. The first computes pessimistic blk_starts,
361 // relative jmp_offset and reloc_size information. The second performs
362 // short branch substitution using the pessimistic sizing. The
363 // third inserts nops where needed.
365 // Step one, perform a pessimistic sizing pass.
366 uint last_call_adr = max_uint;
367 uint last_avoid_back_to_back_adr = max_uint;
368 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
369 for (uint i = 0; i < nblocks; i++) { // For all blocks
370 Block *b = _cfg->_blocks[i];
372 // During short branch replacement, we store the relative (to blk_starts)
373 // offset of jump in jmp_offset, rather than the absolute offset of jump.
374 // This is so that we do not need to recompute sizes of all nodes when
375 // we compute correct blk_starts in our next sizing pass.
376 jmp_offset[i] = 0;
377 jmp_size[i] = 0;
378 jmp_nidx[i] = -1;
379 DEBUG_ONLY( jmp_target[i] = 0; )
380 DEBUG_ONLY( jmp_rule[i] = 0; )
382 // Sum all instruction sizes to compute block size
383 uint last_inst = b->_nodes.size();
384 uint blk_size = 0;
385 for (uint j = 0; j < last_inst; j++) {
386 Node* nj = b->_nodes[j];
387 // Handle machine instruction nodes
388 if (nj->is_Mach()) {
389 MachNode *mach = nj->as_Mach();
390 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
391 reloc_size += mach->reloc();
392 if( mach->is_MachCall() ) {
393 MachCallNode *mcall = mach->as_MachCall();
394 // This destination address is NOT PC-relative
396 mcall->method_set((intptr_t)mcall->entry_point());
398 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) {
399 stub_size += size_java_to_interp();
400 reloc_size += reloc_java_to_interp();
401 }
402 } else if (mach->is_MachSafePoint()) {
403 // If call/safepoint are adjacent, account for possible
404 // nop to disambiguate the two safepoints.
405 // ScheduleAndBundle() can rearrange nodes in a block,
406 // check for all offsets inside this block.
407 if (last_call_adr >= blk_starts[i]) {
408 blk_size += nop_size;
409 }
410 }
411 if (mach->avoid_back_to_back()) {
412 // Nop is inserted between "avoid back to back" instructions.
413 // ScheduleAndBundle() can rearrange nodes in a block,
414 // check for all offsets inside this block.
415 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
416 blk_size += nop_size;
417 }
418 }
419 if (mach->may_be_short_branch()) {
420 if (!nj->is_MachBranch()) {
421 #ifndef PRODUCT
422 nj->dump(3);
423 #endif
424 Unimplemented();
425 }
426 assert(jmp_nidx[i] == -1, "block should have only one branch");
427 jmp_offset[i] = blk_size;
428 jmp_size[i] = nj->size(_regalloc);
429 jmp_nidx[i] = j;
430 has_short_branch_candidate = true;
431 }
432 }
433 blk_size += nj->size(_regalloc);
434 // Remember end of call offset
435 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
436 last_call_adr = blk_starts[i]+blk_size;
437 }
438 // Remember end of avoid_back_to_back offset
439 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) {
440 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
441 }
442 }
444 // When the next block starts a loop, we may insert pad NOP
445 // instructions. Since we cannot know our future alignment,
446 // assume the worst.
447 if (i< nblocks-1) {
448 Block *nb = _cfg->_blocks[i+1];
449 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
450 if (max_loop_pad > 0) {
451 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
452 blk_size += max_loop_pad;
453 }
454 }
456 // Save block size; update total method size
457 blk_starts[i+1] = blk_starts[i]+blk_size;
458 }
460 // Step two, replace eligible long jumps.
461 bool progress = true;
462 uint last_may_be_short_branch_adr = max_uint;
463 while (has_short_branch_candidate && progress) {
464 progress = false;
465 has_short_branch_candidate = false;
466 int adjust_block_start = 0;
467 for (uint i = 0; i < nblocks; i++) {
468 Block *b = _cfg->_blocks[i];
469 int idx = jmp_nidx[i];
470 MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach();
471 if (mach != NULL && mach->may_be_short_branch()) {
472 #ifdef ASSERT
473 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
474 int j;
475 // Find the branch; ignore trailing NOPs.
476 for (j = b->_nodes.size()-1; j>=0; j--) {
477 Node* n = b->_nodes[j];
478 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
479 break;
480 }
481 assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity");
482 #endif
483 int br_size = jmp_size[i];
484 int br_offs = blk_starts[i] + jmp_offset[i];
486 // This requires the TRUE branch target be in succs[0]
487 uint bnum = b->non_connector_successor(0)->_pre_order;
488 int offset = blk_starts[bnum] - br_offs;
489 if (bnum > i) { // adjust following block's offset
490 offset -= adjust_block_start;
491 }
492 // In the following code a nop could be inserted before
493 // the branch which will increase the backward distance.
494 bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr);
495 if (needs_padding && offset <= 0)
496 offset -= nop_size;
498 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
499 // We've got a winner. Replace this branch.
500 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
502 // Update the jmp_size.
503 int new_size = replacement->size(_regalloc);
504 int diff = br_size - new_size;
505 assert(diff >= (int)nop_size, "short_branch size should be smaller");
506 // Conservatively take into accound padding between
507 // avoid_back_to_back branches. Previous branch could be
508 // converted into avoid_back_to_back branch during next
509 // rounds.
510 if (needs_padding && replacement->avoid_back_to_back()) {
511 jmp_offset[i] += nop_size;
512 diff -= nop_size;
513 }
514 adjust_block_start += diff;
515 b->_nodes.map(idx, replacement);
516 mach->subsume_by(replacement);
517 mach = replacement;
518 progress = true;
520 jmp_size[i] = new_size;
521 DEBUG_ONLY( jmp_target[i] = bnum; );
522 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
523 } else {
524 // The jump distance is not short, try again during next iteration.
525 has_short_branch_candidate = true;
526 }
527 } // (mach->may_be_short_branch())
528 if (mach != NULL && (mach->may_be_short_branch() ||
529 mach->avoid_back_to_back())) {
530 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
531 }
532 blk_starts[i+1] -= adjust_block_start;
533 }
534 }
536 #ifdef ASSERT
537 for (uint i = 0; i < nblocks; i++) { // For all blocks
538 if (jmp_target[i] != 0) {
539 int br_size = jmp_size[i];
540 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
541 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
542 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]);
543 }
544 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
545 }
546 }
547 #endif
549 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
550 // after ScheduleAndBundle().
552 // ------------------
553 // Compute size for code buffer
554 code_size = blk_starts[nblocks];
556 // Relocation records
557 reloc_size += 1; // Relo entry for exception handler
559 // Adjust reloc_size to number of record of relocation info
560 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
561 // a relocation index.
562 // The CodeBuffer will expand the locs array if this estimate is too low.
563 reloc_size *= 10 / sizeof(relocInfo);
564 }
566 //------------------------------FillLocArray-----------------------------------
567 // Create a bit of debug info and append it to the array. The mapping is from
568 // Java local or expression stack to constant, register or stack-slot. For
569 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
570 // entry has been taken care of and caller should skip it).
571 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
572 // This should never have accepted Bad before
573 assert(OptoReg::is_valid(regnum), "location must be valid");
574 return (OptoReg::is_reg(regnum))
575 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
576 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
577 }
580 ObjectValue*
581 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
582 for (int i = 0; i < objs->length(); i++) {
583 assert(objs->at(i)->is_object(), "corrupt object cache");
584 ObjectValue* sv = (ObjectValue*) objs->at(i);
585 if (sv->id() == id) {
586 return sv;
587 }
588 }
589 // Otherwise..
590 return NULL;
591 }
593 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
594 ObjectValue* sv ) {
595 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
596 objs->append(sv);
597 }
600 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
601 GrowableArray<ScopeValue*> *array,
602 GrowableArray<ScopeValue*> *objs ) {
603 assert( local, "use _top instead of null" );
604 if (array->length() != idx) {
605 assert(array->length() == idx + 1, "Unexpected array count");
606 // Old functionality:
607 // return
608 // New functionality:
609 // Assert if the local is not top. In product mode let the new node
610 // override the old entry.
611 assert(local == top(), "LocArray collision");
612 if (local == top()) {
613 return;
614 }
615 array->pop();
616 }
617 const Type *t = local->bottom_type();
619 // Is it a safepoint scalar object node?
620 if (local->is_SafePointScalarObject()) {
621 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
623 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
624 if (sv == NULL) {
625 ciKlass* cik = t->is_oopptr()->klass();
626 assert(cik->is_instance_klass() ||
627 cik->is_array_klass(), "Not supported allocation.");
628 sv = new ObjectValue(spobj->_idx,
629 new ConstantOopWriteValue(cik->constant_encoding()));
630 Compile::set_sv_for_object_node(objs, sv);
632 uint first_ind = spobj->first_index();
633 for (uint i = 0; i < spobj->n_fields(); i++) {
634 Node* fld_node = sfpt->in(first_ind+i);
635 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
636 }
637 }
638 array->append(sv);
639 return;
640 }
642 // Grab the register number for the local
643 OptoReg::Name regnum = _regalloc->get_reg_first(local);
644 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
645 // Record the double as two float registers.
646 // The register mask for such a value always specifies two adjacent
647 // float registers, with the lower register number even.
648 // Normally, the allocation of high and low words to these registers
649 // is irrelevant, because nearly all operations on register pairs
650 // (e.g., StoreD) treat them as a single unit.
651 // Here, we assume in addition that the words in these two registers
652 // stored "naturally" (by operations like StoreD and double stores
653 // within the interpreter) such that the lower-numbered register
654 // is written to the lower memory address. This may seem like
655 // a machine dependency, but it is not--it is a requirement on
656 // the author of the <arch>.ad file to ensure that, for every
657 // even/odd double-register pair to which a double may be allocated,
658 // the word in the even single-register is stored to the first
659 // memory word. (Note that register numbers are completely
660 // arbitrary, and are not tied to any machine-level encodings.)
661 #ifdef _LP64
662 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
663 array->append(new ConstantIntValue(0));
664 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
665 } else if ( t->base() == Type::Long ) {
666 array->append(new ConstantIntValue(0));
667 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
668 } else if ( t->base() == Type::RawPtr ) {
669 // jsr/ret return address which must be restored into a the full
670 // width 64-bit stack slot.
671 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
672 }
673 #else //_LP64
674 #ifdef SPARC
675 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
676 // For SPARC we have to swap high and low words for
677 // long values stored in a single-register (g0-g7).
678 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
679 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
680 } else
681 #endif //SPARC
682 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
683 // Repack the double/long as two jints.
684 // The convention the interpreter uses is that the second local
685 // holds the first raw word of the native double representation.
686 // This is actually reasonable, since locals and stack arrays
687 // grow downwards in all implementations.
688 // (If, on some machine, the interpreter's Java locals or stack
689 // were to grow upwards, the embedded doubles would be word-swapped.)
690 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
691 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
692 }
693 #endif //_LP64
694 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
695 OptoReg::is_reg(regnum) ) {
696 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
697 ? Location::float_in_dbl : Location::normal ));
698 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
699 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
700 ? Location::int_in_long : Location::normal ));
701 } else if( t->base() == Type::NarrowOop ) {
702 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
703 } else {
704 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
705 }
706 return;
707 }
709 // No register. It must be constant data.
710 switch (t->base()) {
711 case Type::Half: // Second half of a double
712 ShouldNotReachHere(); // Caller should skip 2nd halves
713 break;
714 case Type::AnyPtr:
715 array->append(new ConstantOopWriteValue(NULL));
716 break;
717 case Type::AryPtr:
718 case Type::InstPtr:
719 case Type::KlassPtr: // fall through
720 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
721 break;
722 case Type::NarrowOop:
723 if (t == TypeNarrowOop::NULL_PTR) {
724 array->append(new ConstantOopWriteValue(NULL));
725 } else {
726 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
727 }
728 break;
729 case Type::Int:
730 array->append(new ConstantIntValue(t->is_int()->get_con()));
731 break;
732 case Type::RawPtr:
733 // A return address (T_ADDRESS).
734 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
735 #ifdef _LP64
736 // Must be restored to the full-width 64-bit stack slot.
737 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
738 #else
739 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
740 #endif
741 break;
742 case Type::FloatCon: {
743 float f = t->is_float_constant()->getf();
744 array->append(new ConstantIntValue(jint_cast(f)));
745 break;
746 }
747 case Type::DoubleCon: {
748 jdouble d = t->is_double_constant()->getd();
749 #ifdef _LP64
750 array->append(new ConstantIntValue(0));
751 array->append(new ConstantDoubleValue(d));
752 #else
753 // Repack the double as two jints.
754 // The convention the interpreter uses is that the second local
755 // holds the first raw word of the native double representation.
756 // This is actually reasonable, since locals and stack arrays
757 // grow downwards in all implementations.
758 // (If, on some machine, the interpreter's Java locals or stack
759 // were to grow upwards, the embedded doubles would be word-swapped.)
760 jint *dp = (jint*)&d;
761 array->append(new ConstantIntValue(dp[1]));
762 array->append(new ConstantIntValue(dp[0]));
763 #endif
764 break;
765 }
766 case Type::Long: {
767 jlong d = t->is_long()->get_con();
768 #ifdef _LP64
769 array->append(new ConstantIntValue(0));
770 array->append(new ConstantLongValue(d));
771 #else
772 // Repack the long as two jints.
773 // The convention the interpreter uses is that the second local
774 // holds the first raw word of the native double representation.
775 // This is actually reasonable, since locals and stack arrays
776 // grow downwards in all implementations.
777 // (If, on some machine, the interpreter's Java locals or stack
778 // were to grow upwards, the embedded doubles would be word-swapped.)
779 jint *dp = (jint*)&d;
780 array->append(new ConstantIntValue(dp[1]));
781 array->append(new ConstantIntValue(dp[0]));
782 #endif
783 break;
784 }
785 case Type::Top: // Add an illegal value here
786 array->append(new LocationValue(Location()));
787 break;
788 default:
789 ShouldNotReachHere();
790 break;
791 }
792 }
794 // Determine if this node starts a bundle
795 bool Compile::starts_bundle(const Node *n) const {
796 return (_node_bundling_limit > n->_idx &&
797 _node_bundling_base[n->_idx].starts_bundle());
798 }
800 //--------------------------Process_OopMap_Node--------------------------------
801 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
803 // Handle special safepoint nodes for synchronization
804 MachSafePointNode *sfn = mach->as_MachSafePoint();
805 MachCallNode *mcall;
807 #ifdef ENABLE_ZAP_DEAD_LOCALS
808 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
809 #endif
811 int safepoint_pc_offset = current_offset;
812 bool is_method_handle_invoke = false;
813 bool return_oop = false;
815 // Add the safepoint in the DebugInfoRecorder
816 if( !mach->is_MachCall() ) {
817 mcall = NULL;
818 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
819 } else {
820 mcall = mach->as_MachCall();
822 // Is the call a MethodHandle call?
823 if (mcall->is_MachCallJava()) {
824 if (mcall->as_MachCallJava()->_method_handle_invoke) {
825 assert(has_method_handle_invokes(), "must have been set during call generation");
826 is_method_handle_invoke = true;
827 }
828 }
830 // Check if a call returns an object.
831 if (mcall->return_value_is_used() &&
832 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
833 return_oop = true;
834 }
835 safepoint_pc_offset += mcall->ret_addr_offset();
836 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
837 }
839 // Loop over the JVMState list to add scope information
840 // Do not skip safepoints with a NULL method, they need monitor info
841 JVMState* youngest_jvms = sfn->jvms();
842 int max_depth = youngest_jvms->depth();
844 // Allocate the object pool for scalar-replaced objects -- the map from
845 // small-integer keys (which can be recorded in the local and ostack
846 // arrays) to descriptions of the object state.
847 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
849 // Visit scopes from oldest to youngest.
850 for (int depth = 1; depth <= max_depth; depth++) {
851 JVMState* jvms = youngest_jvms->of_depth(depth);
852 int idx;
853 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
854 // Safepoints that do not have method() set only provide oop-map and monitor info
855 // to support GC; these do not support deoptimization.
856 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
857 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
858 int num_mon = jvms->nof_monitors();
859 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
860 "JVMS local count must match that of the method");
862 // Add Local and Expression Stack Information
864 // Insert locals into the locarray
865 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
866 for( idx = 0; idx < num_locs; idx++ ) {
867 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
868 }
870 // Insert expression stack entries into the exparray
871 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
872 for( idx = 0; idx < num_exps; idx++ ) {
873 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
874 }
876 // Add in mappings of the monitors
877 assert( !method ||
878 !method->is_synchronized() ||
879 method->is_native() ||
880 num_mon > 0 ||
881 !GenerateSynchronizationCode,
882 "monitors must always exist for synchronized methods");
884 // Build the growable array of ScopeValues for exp stack
885 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
887 // Loop over monitors and insert into array
888 for(idx = 0; idx < num_mon; idx++) {
889 // Grab the node that defines this monitor
890 Node* box_node = sfn->monitor_box(jvms, idx);
891 Node* obj_node = sfn->monitor_obj(jvms, idx);
893 // Create ScopeValue for object
894 ScopeValue *scval = NULL;
896 if( obj_node->is_SafePointScalarObject() ) {
897 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
898 scval = Compile::sv_for_node_id(objs, spobj->_idx);
899 if (scval == NULL) {
900 const Type *t = obj_node->bottom_type();
901 ciKlass* cik = t->is_oopptr()->klass();
902 assert(cik->is_instance_klass() ||
903 cik->is_array_klass(), "Not supported allocation.");
904 ObjectValue* sv = new ObjectValue(spobj->_idx,
905 new ConstantOopWriteValue(cik->constant_encoding()));
906 Compile::set_sv_for_object_node(objs, sv);
908 uint first_ind = spobj->first_index();
909 for (uint i = 0; i < spobj->n_fields(); i++) {
910 Node* fld_node = sfn->in(first_ind+i);
911 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
912 }
913 scval = sv;
914 }
915 } else if( !obj_node->is_Con() ) {
916 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
917 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
918 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
919 } else {
920 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
921 }
922 } else {
923 const TypePtr *tp = obj_node->bottom_type()->make_ptr();
924 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
925 }
927 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
928 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
929 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
930 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
931 }
933 // We dump the object pool first, since deoptimization reads it in first.
934 debug_info()->dump_object_pool(objs);
936 // Build first class objects to pass to scope
937 DebugToken *locvals = debug_info()->create_scope_values(locarray);
938 DebugToken *expvals = debug_info()->create_scope_values(exparray);
939 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
941 // Make method available for all Safepoints
942 ciMethod* scope_method = method ? method : _method;
943 // Describe the scope here
944 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
945 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
946 // Now we can describe the scope.
947 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
948 } // End jvms loop
950 // Mark the end of the scope set.
951 debug_info()->end_safepoint(safepoint_pc_offset);
952 }
956 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
957 class NonSafepointEmitter {
958 Compile* C;
959 JVMState* _pending_jvms;
960 int _pending_offset;
962 void emit_non_safepoint();
964 public:
965 NonSafepointEmitter(Compile* compile) {
966 this->C = compile;
967 _pending_jvms = NULL;
968 _pending_offset = 0;
969 }
971 void observe_instruction(Node* n, int pc_offset) {
972 if (!C->debug_info()->recording_non_safepoints()) return;
974 Node_Notes* nn = C->node_notes_at(n->_idx);
975 if (nn == NULL || nn->jvms() == NULL) return;
976 if (_pending_jvms != NULL &&
977 _pending_jvms->same_calls_as(nn->jvms())) {
978 // Repeated JVMS? Stretch it up here.
979 _pending_offset = pc_offset;
980 } else {
981 if (_pending_jvms != NULL &&
982 _pending_offset < pc_offset) {
983 emit_non_safepoint();
984 }
985 _pending_jvms = NULL;
986 if (pc_offset > C->debug_info()->last_pc_offset()) {
987 // This is the only way _pending_jvms can become non-NULL:
988 _pending_jvms = nn->jvms();
989 _pending_offset = pc_offset;
990 }
991 }
992 }
994 // Stay out of the way of real safepoints:
995 void observe_safepoint(JVMState* jvms, int pc_offset) {
996 if (_pending_jvms != NULL &&
997 !_pending_jvms->same_calls_as(jvms) &&
998 _pending_offset < pc_offset) {
999 emit_non_safepoint();
1000 }
1001 _pending_jvms = NULL;
1002 }
1004 void flush_at_end() {
1005 if (_pending_jvms != NULL) {
1006 emit_non_safepoint();
1007 }
1008 _pending_jvms = NULL;
1009 }
1010 };
1012 void NonSafepointEmitter::emit_non_safepoint() {
1013 JVMState* youngest_jvms = _pending_jvms;
1014 int pc_offset = _pending_offset;
1016 // Clear it now:
1017 _pending_jvms = NULL;
1019 DebugInformationRecorder* debug_info = C->debug_info();
1020 assert(debug_info->recording_non_safepoints(), "sanity");
1022 debug_info->add_non_safepoint(pc_offset);
1023 int max_depth = youngest_jvms->depth();
1025 // Visit scopes from oldest to youngest.
1026 for (int depth = 1; depth <= max_depth; depth++) {
1027 JVMState* jvms = youngest_jvms->of_depth(depth);
1028 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1029 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1030 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1031 }
1033 // Mark the end of the scope set.
1034 debug_info->end_non_safepoint(pc_offset);
1035 }
1039 // helper for fill_buffer bailout logic
1040 static void turn_off_compiler(Compile* C) {
1041 if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) {
1042 // Do not turn off compilation if a single giant method has
1043 // blown the code cache size.
1044 C->record_failure("excessive request to CodeCache");
1045 } else {
1046 // Let CompilerBroker disable further compilations.
1047 C->record_failure("CodeCache is full");
1048 }
1049 }
1052 //------------------------------init_buffer------------------------------------
1053 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
1055 // Set the initially allocated size
1056 int code_req = initial_code_capacity;
1057 int locs_req = initial_locs_capacity;
1058 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1059 int const_req = initial_const_capacity;
1061 int pad_req = NativeCall::instruction_size;
1062 // The extra spacing after the code is necessary on some platforms.
1063 // Sometimes we need to patch in a jump after the last instruction,
1064 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1066 // Compute the byte offset where we can store the deopt pc.
1067 if (fixed_slots() != 0) {
1068 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1069 }
1071 // Compute prolog code size
1072 _method_size = 0;
1073 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1074 #ifdef IA64
1075 if (save_argument_registers()) {
1076 // 4815101: this is a stub with implicit and unknown precision fp args.
1077 // The usual spill mechanism can only generate stfd's in this case, which
1078 // doesn't work if the fp reg to spill contains a single-precision denorm.
1079 // Instead, we hack around the normal spill mechanism using stfspill's and
1080 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1081 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1082 //
1083 // If we ever implement 16-byte 'registers' == stack slots, we can
1084 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1085 // instead of stfd/stfs/ldfd/ldfs.
1086 _frame_slots += 8*(16/BytesPerInt);
1087 }
1088 #endif
1089 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1091 if (has_mach_constant_base_node()) {
1092 // Fill the constant table.
1093 // Note: This must happen before shorten_branches.
1094 for (uint i = 0; i < _cfg->_num_blocks; i++) {
1095 Block* b = _cfg->_blocks[i];
1097 for (uint j = 0; j < b->_nodes.size(); j++) {
1098 Node* n = b->_nodes[j];
1100 // If the node is a MachConstantNode evaluate the constant
1101 // value section.
1102 if (n->is_MachConstant()) {
1103 MachConstantNode* machcon = n->as_MachConstant();
1104 machcon->eval_constant(C);
1105 }
1106 }
1107 }
1109 // Calculate the offsets of the constants and the size of the
1110 // constant table (including the padding to the next section).
1111 constant_table().calculate_offsets_and_size();
1112 const_req = constant_table().size();
1113 }
1115 // Initialize the space for the BufferBlob used to find and verify
1116 // instruction size in MachNode::emit_size()
1117 init_scratch_buffer_blob(const_req);
1118 if (failing()) return NULL; // Out of memory
1120 // Pre-compute the length of blocks and replace
1121 // long branches with short if machine supports it.
1122 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1124 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1125 int exception_handler_req = size_exception_handler();
1126 int deopt_handler_req = size_deopt_handler();
1127 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1128 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1129 stub_req += MAX_stubs_size; // ensure per-stub margin
1130 code_req += MAX_inst_size; // ensure per-instruction margin
1132 if (StressCodeBuffers)
1133 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1135 int total_req =
1136 const_req +
1137 code_req +
1138 pad_req +
1139 stub_req +
1140 exception_handler_req +
1141 deopt_handler_req; // deopt handler
1143 if (has_method_handle_invokes())
1144 total_req += deopt_handler_req; // deopt MH handler
1146 CodeBuffer* cb = code_buffer();
1147 cb->initialize(total_req, locs_req);
1149 // Have we run out of code space?
1150 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1151 turn_off_compiler(this);
1152 return NULL;
1153 }
1154 // Configure the code buffer.
1155 cb->initialize_consts_size(const_req);
1156 cb->initialize_stubs_size(stub_req);
1157 cb->initialize_oop_recorder(env()->oop_recorder());
1159 // fill in the nop array for bundling computations
1160 MachNode *_nop_list[Bundle::_nop_count];
1161 Bundle::initialize_nops(_nop_list, this);
1163 return cb;
1164 }
1166 //------------------------------fill_buffer------------------------------------
1167 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1168 // blk_starts[] contains offsets calculated during short branches processing,
1169 // offsets should not be increased during following steps.
1171 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1172 // of a loop. It is used to determine the padding for loop alignment.
1173 compute_loop_first_inst_sizes();
1175 // Create oopmap set.
1176 _oop_map_set = new OopMapSet();
1178 // !!!!! This preserves old handling of oopmaps for now
1179 debug_info()->set_oopmaps(_oop_map_set);
1181 uint nblocks = _cfg->_num_blocks;
1182 // Count and start of implicit null check instructions
1183 uint inct_cnt = 0;
1184 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1186 // Count and start of calls
1187 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1189 uint return_offset = 0;
1190 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1192 int previous_offset = 0;
1193 int current_offset = 0;
1194 int last_call_offset = -1;
1195 int last_avoid_back_to_back_offset = -1;
1196 #ifdef ASSERT
1197 int block_alignment_padding = 0;
1199 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1200 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1201 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1202 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1203 #endif
1205 // Create an array of unused labels, one for each basic block, if printing is enabled
1206 #ifndef PRODUCT
1207 int *node_offsets = NULL;
1208 uint node_offset_limit = unique();
1210 if (print_assembly())
1211 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1212 #endif
1214 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1216 // Emit the constant table.
1217 if (has_mach_constant_base_node()) {
1218 constant_table().emit(*cb);
1219 }
1221 // Create an array of labels, one for each basic block
1222 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1223 for (uint i=0; i <= nblocks; i++) {
1224 blk_labels[i].init();
1225 }
1227 // ------------------
1228 // Now fill in the code buffer
1229 Node *delay_slot = NULL;
1231 for (uint i=0; i < nblocks; i++) {
1232 guarantee(blk_starts[i] >= (uint)cb->insts_size(),"should not increase size");
1234 Block *b = _cfg->_blocks[i];
1236 Node *head = b->head();
1238 // If this block needs to start aligned (i.e, can be reached other
1239 // than by falling-thru from the previous block), then force the
1240 // start of a new bundle.
1241 if (Pipeline::requires_bundling() && starts_bundle(head))
1242 cb->flush_bundle(true);
1244 #ifdef ASSERT
1245 if (!b->is_connector()) {
1246 stringStream st;
1247 b->dump_head(&_cfg->_bbs, &st);
1248 MacroAssembler(cb).block_comment(st.as_string());
1249 }
1250 jmp_target[i] = 0;
1251 jmp_offset[i] = 0;
1252 jmp_size[i] = 0;
1253 jmp_rule[i] = 0;
1255 // Maximum alignment padding for loop block was used
1256 // during first round of branches shortening, as result
1257 // padding for nodes (sfpt after call) was not added.
1258 // Take this into account for block's size change check
1259 // and allow increase block's size by the difference
1260 // of maximum and actual alignment paddings.
1261 int orig_blk_size = blk_starts[i+1] - blk_starts[i] + block_alignment_padding;
1262 #endif
1263 int blk_offset = current_offset;
1265 // Define the label at the beginning of the basic block
1266 MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
1268 uint last_inst = b->_nodes.size();
1270 // Emit block normally, except for last instruction.
1271 // Emit means "dump code bits into code buffer".
1272 for (uint j = 0; j<last_inst; j++) {
1274 // Get the node
1275 Node* n = b->_nodes[j];
1277 // See if delay slots are supported
1278 if (valid_bundle_info(n) &&
1279 node_bundling(n)->used_in_unconditional_delay()) {
1280 assert(delay_slot == NULL, "no use of delay slot node");
1281 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1283 delay_slot = n;
1284 continue;
1285 }
1287 // If this starts a new instruction group, then flush the current one
1288 // (but allow split bundles)
1289 if (Pipeline::requires_bundling() && starts_bundle(n))
1290 cb->flush_bundle(false);
1292 // The following logic is duplicated in the code ifdeffed for
1293 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1294 // should be factored out. Or maybe dispersed to the nodes?
1296 // Special handling for SafePoint/Call Nodes
1297 bool is_mcall = false;
1298 if (n->is_Mach()) {
1299 MachNode *mach = n->as_Mach();
1300 is_mcall = n->is_MachCall();
1301 bool is_sfn = n->is_MachSafePoint();
1303 // If this requires all previous instructions be flushed, then do so
1304 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1305 cb->flush_bundle(true);
1306 current_offset = cb->insts_size();
1307 }
1309 // A padding may be needed again since a previous instruction
1310 // could be moved to delay slot.
1312 // align the instruction if necessary
1313 int padding = mach->compute_padding(current_offset);
1314 // Make sure safepoint node for polling is distinct from a call's
1315 // return by adding a nop if needed.
1316 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1317 padding = nop_size;
1318 }
1319 if (padding == 0 && mach->avoid_back_to_back() &&
1320 current_offset == last_avoid_back_to_back_offset) {
1321 // Avoid back to back some instructions.
1322 padding = nop_size;
1323 }
1325 if(padding > 0) {
1326 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1327 int nops_cnt = padding / nop_size;
1328 MachNode *nop = new (this) MachNopNode(nops_cnt);
1329 b->_nodes.insert(j++, nop);
1330 last_inst++;
1331 _cfg->_bbs.map( nop->_idx, b );
1332 nop->emit(*cb, _regalloc);
1333 cb->flush_bundle(true);
1334 current_offset = cb->insts_size();
1335 }
1337 // Remember the start of the last call in a basic block
1338 if (is_mcall) {
1339 MachCallNode *mcall = mach->as_MachCall();
1341 // This destination address is NOT PC-relative
1342 mcall->method_set((intptr_t)mcall->entry_point());
1344 // Save the return address
1345 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset();
1347 if (mcall->is_MachCallLeaf()) {
1348 is_mcall = false;
1349 is_sfn = false;
1350 }
1351 }
1353 // sfn will be valid whenever mcall is valid now because of inheritance
1354 if (is_sfn || is_mcall) {
1356 // Handle special safepoint nodes for synchronization
1357 if (!is_mcall) {
1358 MachSafePointNode *sfn = mach->as_MachSafePoint();
1359 // !!!!! Stubs only need an oopmap right now, so bail out
1360 if (sfn->jvms()->method() == NULL) {
1361 // Write the oopmap directly to the code blob??!!
1362 # ifdef ENABLE_ZAP_DEAD_LOCALS
1363 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1364 # endif
1365 continue;
1366 }
1367 } // End synchronization
1369 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1370 current_offset);
1371 Process_OopMap_Node(mach, current_offset);
1372 } // End if safepoint
1374 // If this is a null check, then add the start of the previous instruction to the list
1375 else if( mach->is_MachNullCheck() ) {
1376 inct_starts[inct_cnt++] = previous_offset;
1377 }
1379 // If this is a branch, then fill in the label with the target BB's label
1380 else if (mach->is_MachBranch()) {
1381 // This requires the TRUE branch target be in succs[0]
1382 uint block_num = b->non_connector_successor(0)->_pre_order;
1384 // Try to replace long branch if delay slot is not used,
1385 // it is mostly for back branches since forward branch's
1386 // distance is not updated yet.
1387 bool delay_slot_is_used = valid_bundle_info(n) &&
1388 node_bundling(n)->use_unconditional_delay();
1389 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1390 assert(delay_slot == NULL, "not expecting delay slot node");
1391 int br_size = n->size(_regalloc);
1392 int offset = blk_starts[block_num] - current_offset;
1393 if (block_num >= i) {
1394 // Current and following block's offset are not
1395 // finilized yet, adjust distance by the difference
1396 // between calculated and final offsets of current block.
1397 offset -= (blk_starts[i] - blk_offset);
1398 }
1399 // In the following code a nop could be inserted before
1400 // the branch which will increase the backward distance.
1401 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1402 if (needs_padding && offset <= 0)
1403 offset -= nop_size;
1405 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1406 // We've got a winner. Replace this branch.
1407 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
1409 // Update the jmp_size.
1410 int new_size = replacement->size(_regalloc);
1411 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1412 // Insert padding between avoid_back_to_back branches.
1413 if (needs_padding && replacement->avoid_back_to_back()) {
1414 MachNode *nop = new (this) MachNopNode();
1415 b->_nodes.insert(j++, nop);
1416 _cfg->_bbs.map(nop->_idx, b);
1417 last_inst++;
1418 nop->emit(*cb, _regalloc);
1419 cb->flush_bundle(true);
1420 current_offset = cb->insts_size();
1421 }
1422 #ifdef ASSERT
1423 jmp_target[i] = block_num;
1424 jmp_offset[i] = current_offset - blk_offset;
1425 jmp_size[i] = new_size;
1426 jmp_rule[i] = mach->rule();
1427 #endif
1428 b->_nodes.map(j, replacement);
1429 mach->subsume_by(replacement);
1430 n = replacement;
1431 mach = replacement;
1432 }
1433 }
1434 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1435 } else if (mach->ideal_Opcode() == Op_Jump) {
1436 for (uint h = 0; h < b->_num_succs; h++) {
1437 Block* succs_block = b->_succs[h];
1438 for (uint j = 1; j < succs_block->num_preds(); j++) {
1439 Node* jpn = succs_block->pred(j);
1440 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1441 uint block_num = succs_block->non_connector()->_pre_order;
1442 Label *blkLabel = &blk_labels[block_num];
1443 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1444 }
1445 }
1446 }
1447 }
1449 #ifdef ASSERT
1450 // Check that oop-store precedes the card-mark
1451 else if (mach->ideal_Opcode() == Op_StoreCM) {
1452 uint storeCM_idx = j;
1453 int count = 0;
1454 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1455 Node *oop_store = mach->in(prec); // Precedence edge
1456 if (oop_store == NULL) continue;
1457 count++;
1458 uint i4;
1459 for( i4 = 0; i4 < last_inst; ++i4 ) {
1460 if( b->_nodes[i4] == oop_store ) break;
1461 }
1462 // Note: This test can provide a false failure if other precedence
1463 // edges have been added to the storeCMNode.
1464 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1465 }
1466 assert(count > 0, "storeCM expects at least one precedence edge");
1467 }
1468 #endif
1470 else if (!n->is_Proj()) {
1471 // Remember the beginning of the previous instruction, in case
1472 // it's followed by a flag-kill and a null-check. Happens on
1473 // Intel all the time, with add-to-memory kind of opcodes.
1474 previous_offset = current_offset;
1475 }
1476 }
1478 // Verify that there is sufficient space remaining
1479 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1480 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1481 turn_off_compiler(this);
1482 return;
1483 }
1485 // Save the offset for the listing
1486 #ifndef PRODUCT
1487 if (node_offsets && n->_idx < node_offset_limit)
1488 node_offsets[n->_idx] = cb->insts_size();
1489 #endif
1491 // "Normal" instruction case
1492 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1493 n->emit(*cb, _regalloc);
1494 current_offset = cb->insts_size();
1496 #ifdef ASSERT
1497 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1498 n->dump();
1499 assert(false, "wrong size of mach node");
1500 }
1501 #endif
1502 non_safepoints.observe_instruction(n, current_offset);
1504 // mcall is last "call" that can be a safepoint
1505 // record it so we can see if a poll will directly follow it
1506 // in which case we'll need a pad to make the PcDesc sites unique
1507 // see 5010568. This can be slightly inaccurate but conservative
1508 // in the case that return address is not actually at current_offset.
1509 // This is a small price to pay.
1511 if (is_mcall) {
1512 last_call_offset = current_offset;
1513 }
1515 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) {
1516 // Avoid back to back some instructions.
1517 last_avoid_back_to_back_offset = current_offset;
1518 }
1520 // See if this instruction has a delay slot
1521 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1522 assert(delay_slot != NULL, "expecting delay slot node");
1524 // Back up 1 instruction
1525 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1527 // Save the offset for the listing
1528 #ifndef PRODUCT
1529 if (node_offsets && delay_slot->_idx < node_offset_limit)
1530 node_offsets[delay_slot->_idx] = cb->insts_size();
1531 #endif
1533 // Support a SafePoint in the delay slot
1534 if (delay_slot->is_MachSafePoint()) {
1535 MachNode *mach = delay_slot->as_Mach();
1536 // !!!!! Stubs only need an oopmap right now, so bail out
1537 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1538 // Write the oopmap directly to the code blob??!!
1539 # ifdef ENABLE_ZAP_DEAD_LOCALS
1540 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1541 # endif
1542 delay_slot = NULL;
1543 continue;
1544 }
1546 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1547 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1548 adjusted_offset);
1549 // Generate an OopMap entry
1550 Process_OopMap_Node(mach, adjusted_offset);
1551 }
1553 // Insert the delay slot instruction
1554 delay_slot->emit(*cb, _regalloc);
1556 // Don't reuse it
1557 delay_slot = NULL;
1558 }
1560 } // End for all instructions in block
1561 assert((uint)blk_offset <= blk_starts[i], "shouldn't increase distance");
1562 blk_starts[i] = blk_offset;
1564 // If the next block is the top of a loop, pad this block out to align
1565 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1566 if (i < nblocks-1) {
1567 Block *nb = _cfg->_blocks[i+1];
1568 int padding = nb->alignment_padding(current_offset);
1569 if( padding > 0 ) {
1570 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1571 b->_nodes.insert( b->_nodes.size(), nop );
1572 _cfg->_bbs.map( nop->_idx, b );
1573 nop->emit(*cb, _regalloc);
1574 current_offset = cb->insts_size();
1575 }
1576 #ifdef ASSERT
1577 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
1578 block_alignment_padding = (max_loop_pad - padding);
1579 assert(block_alignment_padding >= 0, "sanity");
1580 #endif
1581 }
1582 // Verify that the distance for generated before forward
1583 // short branches is still valid.
1584 assert(orig_blk_size >= (current_offset - blk_offset), "shouldn't increase block size");
1586 } // End of for all blocks
1587 blk_starts[nblocks] = current_offset;
1589 non_safepoints.flush_at_end();
1591 // Offset too large?
1592 if (failing()) return;
1594 // Define a pseudo-label at the end of the code
1595 MacroAssembler(cb).bind( blk_labels[nblocks] );
1597 // Compute the size of the first block
1598 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1600 assert(cb->insts_size() < 500000, "method is unreasonably large");
1602 #ifdef ASSERT
1603 for (uint i = 0; i < nblocks; i++) { // For all blocks
1604 if (jmp_target[i] != 0) {
1605 int br_size = jmp_size[i];
1606 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1607 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1608 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]);
1609 assert(false, "Displacement too large for short jmp");
1610 }
1611 }
1612 }
1613 #endif
1615 // ------------------
1617 #ifndef PRODUCT
1618 // Information on the size of the method, without the extraneous code
1619 Scheduling::increment_method_size(cb->insts_size());
1620 #endif
1622 // ------------------
1623 // Fill in exception table entries.
1624 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1626 // Only java methods have exception handlers and deopt handlers
1627 if (_method) {
1628 // Emit the exception handler code.
1629 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1630 // Emit the deopt handler code.
1631 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1633 // Emit the MethodHandle deopt handler code (if required).
1634 if (has_method_handle_invokes()) {
1635 // We can use the same code as for the normal deopt handler, we
1636 // just need a different entry point address.
1637 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1638 }
1639 }
1641 // One last check for failed CodeBuffer::expand:
1642 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1643 turn_off_compiler(this);
1644 return;
1645 }
1647 #ifndef PRODUCT
1648 // Dump the assembly code, including basic-block numbers
1649 if (print_assembly()) {
1650 ttyLocker ttyl; // keep the following output all in one block
1651 if (!VMThread::should_terminate()) { // test this under the tty lock
1652 // This output goes directly to the tty, not the compiler log.
1653 // To enable tools to match it up with the compilation activity,
1654 // be sure to tag this tty output with the compile ID.
1655 if (xtty != NULL) {
1656 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1657 is_osr_compilation() ? " compile_kind='osr'" :
1658 "");
1659 }
1660 if (method() != NULL) {
1661 method()->print_oop();
1662 print_codes();
1663 }
1664 dump_asm(node_offsets, node_offset_limit);
1665 if (xtty != NULL) {
1666 xtty->tail("opto_assembly");
1667 }
1668 }
1669 }
1670 #endif
1672 }
1674 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1675 _inc_table.set_size(cnt);
1677 uint inct_cnt = 0;
1678 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1679 Block *b = _cfg->_blocks[i];
1680 Node *n = NULL;
1681 int j;
1683 // Find the branch; ignore trailing NOPs.
1684 for( j = b->_nodes.size()-1; j>=0; j-- ) {
1685 n = b->_nodes[j];
1686 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con )
1687 break;
1688 }
1690 // If we didn't find anything, continue
1691 if( j < 0 ) continue;
1693 // Compute ExceptionHandlerTable subtable entry and add it
1694 // (skip empty blocks)
1695 if( n->is_Catch() ) {
1697 // Get the offset of the return from the call
1698 uint call_return = call_returns[b->_pre_order];
1699 #ifdef ASSERT
1700 assert( call_return > 0, "no call seen for this basic block" );
1701 while( b->_nodes[--j]->is_MachProj() ) ;
1702 assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" );
1703 #endif
1704 // last instruction is a CatchNode, find it's CatchProjNodes
1705 int nof_succs = b->_num_succs;
1706 // allocate space
1707 GrowableArray<intptr_t> handler_bcis(nof_succs);
1708 GrowableArray<intptr_t> handler_pcos(nof_succs);
1709 // iterate through all successors
1710 for (int j = 0; j < nof_succs; j++) {
1711 Block* s = b->_succs[j];
1712 bool found_p = false;
1713 for( uint k = 1; k < s->num_preds(); k++ ) {
1714 Node *pk = s->pred(k);
1715 if( pk->is_CatchProj() && pk->in(0) == n ) {
1716 const CatchProjNode* p = pk->as_CatchProj();
1717 found_p = true;
1718 // add the corresponding handler bci & pco information
1719 if( p->_con != CatchProjNode::fall_through_index ) {
1720 // p leads to an exception handler (and is not fall through)
1721 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering");
1722 // no duplicates, please
1723 if( !handler_bcis.contains(p->handler_bci()) ) {
1724 uint block_num = s->non_connector()->_pre_order;
1725 handler_bcis.append(p->handler_bci());
1726 handler_pcos.append(blk_labels[block_num].loc_pos());
1727 }
1728 }
1729 }
1730 }
1731 assert(found_p, "no matching predecessor found");
1732 // Note: Due to empty block removal, one block may have
1733 // several CatchProj inputs, from the same Catch.
1734 }
1736 // Set the offset of the return from the call
1737 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1738 continue;
1739 }
1741 // Handle implicit null exception table updates
1742 if( n->is_MachNullCheck() ) {
1743 uint block_num = b->non_connector_successor(0)->_pre_order;
1744 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() );
1745 continue;
1746 }
1747 } // End of for all blocks fill in exception table entries
1748 }
1750 // Static Variables
1751 #ifndef PRODUCT
1752 uint Scheduling::_total_nop_size = 0;
1753 uint Scheduling::_total_method_size = 0;
1754 uint Scheduling::_total_branches = 0;
1755 uint Scheduling::_total_unconditional_delays = 0;
1756 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1757 #endif
1759 // Initializer for class Scheduling
1761 Scheduling::Scheduling(Arena *arena, Compile &compile)
1762 : _arena(arena),
1763 _cfg(compile.cfg()),
1764 _bbs(compile.cfg()->_bbs),
1765 _regalloc(compile.regalloc()),
1766 _reg_node(arena),
1767 _bundle_instr_count(0),
1768 _bundle_cycle_number(0),
1769 _scheduled(arena),
1770 _available(arena),
1771 _next_node(NULL),
1772 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1773 _pinch_free_list(arena)
1774 #ifndef PRODUCT
1775 , _branches(0)
1776 , _unconditional_delays(0)
1777 #endif
1778 {
1779 // Create a MachNopNode
1780 _nop = new (&compile) MachNopNode();
1782 // Now that the nops are in the array, save the count
1783 // (but allow entries for the nops)
1784 _node_bundling_limit = compile.unique();
1785 uint node_max = _regalloc->node_regs_max_index();
1787 compile.set_node_bundling_limit(_node_bundling_limit);
1789 // This one is persistent within the Compile class
1790 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1792 // Allocate space for fixed-size arrays
1793 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1794 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1795 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1797 // Clear the arrays
1798 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1799 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1800 memset(_uses, 0, node_max * sizeof(short));
1801 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1803 // Clear the bundling information
1804 memcpy(_bundle_use_elements,
1805 Pipeline_Use::elaborated_elements,
1806 sizeof(Pipeline_Use::elaborated_elements));
1808 // Get the last node
1809 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1];
1811 _next_node = bb->_nodes[bb->_nodes.size()-1];
1812 }
1814 #ifndef PRODUCT
1815 // Scheduling destructor
1816 Scheduling::~Scheduling() {
1817 _total_branches += _branches;
1818 _total_unconditional_delays += _unconditional_delays;
1819 }
1820 #endif
1822 // Step ahead "i" cycles
1823 void Scheduling::step(uint i) {
1825 Bundle *bundle = node_bundling(_next_node);
1826 bundle->set_starts_bundle();
1828 // Update the bundle record, but leave the flags information alone
1829 if (_bundle_instr_count > 0) {
1830 bundle->set_instr_count(_bundle_instr_count);
1831 bundle->set_resources_used(_bundle_use.resourcesUsed());
1832 }
1834 // Update the state information
1835 _bundle_instr_count = 0;
1836 _bundle_cycle_number += i;
1837 _bundle_use.step(i);
1838 }
1840 void Scheduling::step_and_clear() {
1841 Bundle *bundle = node_bundling(_next_node);
1842 bundle->set_starts_bundle();
1844 // Update the bundle record
1845 if (_bundle_instr_count > 0) {
1846 bundle->set_instr_count(_bundle_instr_count);
1847 bundle->set_resources_used(_bundle_use.resourcesUsed());
1849 _bundle_cycle_number += 1;
1850 }
1852 // Clear the bundling information
1853 _bundle_instr_count = 0;
1854 _bundle_use.reset();
1856 memcpy(_bundle_use_elements,
1857 Pipeline_Use::elaborated_elements,
1858 sizeof(Pipeline_Use::elaborated_elements));
1859 }
1861 //------------------------------ScheduleAndBundle------------------------------
1862 // Perform instruction scheduling and bundling over the sequence of
1863 // instructions in backwards order.
1864 void Compile::ScheduleAndBundle() {
1866 // Don't optimize this if it isn't a method
1867 if (!_method)
1868 return;
1870 // Don't optimize this if scheduling is disabled
1871 if (!do_scheduling())
1872 return;
1874 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1876 // Create a data structure for all the scheduling information
1877 Scheduling scheduling(Thread::current()->resource_area(), *this);
1879 // Walk backwards over each basic block, computing the needed alignment
1880 // Walk over all the basic blocks
1881 scheduling.DoScheduling();
1882 }
1884 //------------------------------ComputeLocalLatenciesForward-------------------
1885 // Compute the latency of all the instructions. This is fairly simple,
1886 // because we already have a legal ordering. Walk over the instructions
1887 // from first to last, and compute the latency of the instruction based
1888 // on the latency of the preceding instruction(s).
1889 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1890 #ifndef PRODUCT
1891 if (_cfg->C->trace_opto_output())
1892 tty->print("# -> ComputeLocalLatenciesForward\n");
1893 #endif
1895 // Walk over all the schedulable instructions
1896 for( uint j=_bb_start; j < _bb_end; j++ ) {
1898 // This is a kludge, forcing all latency calculations to start at 1.
1899 // Used to allow latency 0 to force an instruction to the beginning
1900 // of the bb
1901 uint latency = 1;
1902 Node *use = bb->_nodes[j];
1903 uint nlen = use->len();
1905 // Walk over all the inputs
1906 for ( uint k=0; k < nlen; k++ ) {
1907 Node *def = use->in(k);
1908 if (!def)
1909 continue;
1911 uint l = _node_latency[def->_idx] + use->latency(k);
1912 if (latency < l)
1913 latency = l;
1914 }
1916 _node_latency[use->_idx] = latency;
1918 #ifndef PRODUCT
1919 if (_cfg->C->trace_opto_output()) {
1920 tty->print("# latency %4d: ", latency);
1921 use->dump();
1922 }
1923 #endif
1924 }
1926 #ifndef PRODUCT
1927 if (_cfg->C->trace_opto_output())
1928 tty->print("# <- ComputeLocalLatenciesForward\n");
1929 #endif
1931 } // end ComputeLocalLatenciesForward
1933 // See if this node fits into the present instruction bundle
1934 bool Scheduling::NodeFitsInBundle(Node *n) {
1935 uint n_idx = n->_idx;
1937 // If this is the unconditional delay instruction, then it fits
1938 if (n == _unconditional_delay_slot) {
1939 #ifndef PRODUCT
1940 if (_cfg->C->trace_opto_output())
1941 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1942 #endif
1943 return (true);
1944 }
1946 // If the node cannot be scheduled this cycle, skip it
1947 if (_current_latency[n_idx] > _bundle_cycle_number) {
1948 #ifndef PRODUCT
1949 if (_cfg->C->trace_opto_output())
1950 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1951 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1952 #endif
1953 return (false);
1954 }
1956 const Pipeline *node_pipeline = n->pipeline();
1958 uint instruction_count = node_pipeline->instructionCount();
1959 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1960 instruction_count = 0;
1961 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1962 instruction_count++;
1964 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1965 #ifndef PRODUCT
1966 if (_cfg->C->trace_opto_output())
1967 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1968 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1969 #endif
1970 return (false);
1971 }
1973 // Don't allow non-machine nodes to be handled this way
1974 if (!n->is_Mach() && instruction_count == 0)
1975 return (false);
1977 // See if there is any overlap
1978 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1980 if (delay > 0) {
1981 #ifndef PRODUCT
1982 if (_cfg->C->trace_opto_output())
1983 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1984 #endif
1985 return false;
1986 }
1988 #ifndef PRODUCT
1989 if (_cfg->C->trace_opto_output())
1990 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1991 #endif
1993 return true;
1994 }
1996 Node * Scheduling::ChooseNodeToBundle() {
1997 uint siz = _available.size();
1999 if (siz == 0) {
2001 #ifndef PRODUCT
2002 if (_cfg->C->trace_opto_output())
2003 tty->print("# ChooseNodeToBundle: NULL\n");
2004 #endif
2005 return (NULL);
2006 }
2008 // Fast path, if only 1 instruction in the bundle
2009 if (siz == 1) {
2010 #ifndef PRODUCT
2011 if (_cfg->C->trace_opto_output()) {
2012 tty->print("# ChooseNodeToBundle (only 1): ");
2013 _available[0]->dump();
2014 }
2015 #endif
2016 return (_available[0]);
2017 }
2019 // Don't bother, if the bundle is already full
2020 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
2021 for ( uint i = 0; i < siz; i++ ) {
2022 Node *n = _available[i];
2024 // Skip projections, we'll handle them another way
2025 if (n->is_Proj())
2026 continue;
2028 // This presupposed that instructions are inserted into the
2029 // available list in a legality order; i.e. instructions that
2030 // must be inserted first are at the head of the list
2031 if (NodeFitsInBundle(n)) {
2032 #ifndef PRODUCT
2033 if (_cfg->C->trace_opto_output()) {
2034 tty->print("# ChooseNodeToBundle: ");
2035 n->dump();
2036 }
2037 #endif
2038 return (n);
2039 }
2040 }
2041 }
2043 // Nothing fits in this bundle, choose the highest priority
2044 #ifndef PRODUCT
2045 if (_cfg->C->trace_opto_output()) {
2046 tty->print("# ChooseNodeToBundle: ");
2047 _available[0]->dump();
2048 }
2049 #endif
2051 return _available[0];
2052 }
2054 //------------------------------AddNodeToAvailableList-------------------------
2055 void Scheduling::AddNodeToAvailableList(Node *n) {
2056 assert( !n->is_Proj(), "projections never directly made available" );
2057 #ifndef PRODUCT
2058 if (_cfg->C->trace_opto_output()) {
2059 tty->print("# AddNodeToAvailableList: ");
2060 n->dump();
2061 }
2062 #endif
2064 int latency = _current_latency[n->_idx];
2066 // Insert in latency order (insertion sort)
2067 uint i;
2068 for ( i=0; i < _available.size(); i++ )
2069 if (_current_latency[_available[i]->_idx] > latency)
2070 break;
2072 // Special Check for compares following branches
2073 if( n->is_Mach() && _scheduled.size() > 0 ) {
2074 int op = n->as_Mach()->ideal_Opcode();
2075 Node *last = _scheduled[0];
2076 if( last->is_MachIf() && last->in(1) == n &&
2077 ( op == Op_CmpI ||
2078 op == Op_CmpU ||
2079 op == Op_CmpP ||
2080 op == Op_CmpF ||
2081 op == Op_CmpD ||
2082 op == Op_CmpL ) ) {
2084 // Recalculate position, moving to front of same latency
2085 for ( i=0 ; i < _available.size(); i++ )
2086 if (_current_latency[_available[i]->_idx] >= latency)
2087 break;
2088 }
2089 }
2091 // Insert the node in the available list
2092 _available.insert(i, n);
2094 #ifndef PRODUCT
2095 if (_cfg->C->trace_opto_output())
2096 dump_available();
2097 #endif
2098 }
2100 //------------------------------DecrementUseCounts-----------------------------
2101 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2102 for ( uint i=0; i < n->len(); i++ ) {
2103 Node *def = n->in(i);
2104 if (!def) continue;
2105 if( def->is_Proj() ) // If this is a machine projection, then
2106 def = def->in(0); // propagate usage thru to the base instruction
2108 if( _bbs[def->_idx] != bb ) // Ignore if not block-local
2109 continue;
2111 // Compute the latency
2112 uint l = _bundle_cycle_number + n->latency(i);
2113 if (_current_latency[def->_idx] < l)
2114 _current_latency[def->_idx] = l;
2116 // If this does not have uses then schedule it
2117 if ((--_uses[def->_idx]) == 0)
2118 AddNodeToAvailableList(def);
2119 }
2120 }
2122 //------------------------------AddNodeToBundle--------------------------------
2123 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2124 #ifndef PRODUCT
2125 if (_cfg->C->trace_opto_output()) {
2126 tty->print("# AddNodeToBundle: ");
2127 n->dump();
2128 }
2129 #endif
2131 // Remove this from the available list
2132 uint i;
2133 for (i = 0; i < _available.size(); i++)
2134 if (_available[i] == n)
2135 break;
2136 assert(i < _available.size(), "entry in _available list not found");
2137 _available.remove(i);
2139 // See if this fits in the current bundle
2140 const Pipeline *node_pipeline = n->pipeline();
2141 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2143 // Check for instructions to be placed in the delay slot. We
2144 // do this before we actually schedule the current instruction,
2145 // because the delay slot follows the current instruction.
2146 if (Pipeline::_branch_has_delay_slot &&
2147 node_pipeline->hasBranchDelay() &&
2148 !_unconditional_delay_slot) {
2150 uint siz = _available.size();
2152 // Conditional branches can support an instruction that
2153 // is unconditionally executed and not dependent by the
2154 // branch, OR a conditionally executed instruction if
2155 // the branch is taken. In practice, this means that
2156 // the first instruction at the branch target is
2157 // copied to the delay slot, and the branch goes to
2158 // the instruction after that at the branch target
2159 if ( n->is_MachBranch() ) {
2161 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2162 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2164 #ifndef PRODUCT
2165 _branches++;
2166 #endif
2168 // At least 1 instruction is on the available list
2169 // that is not dependent on the branch
2170 for (uint i = 0; i < siz; i++) {
2171 Node *d = _available[i];
2172 const Pipeline *avail_pipeline = d->pipeline();
2174 // Don't allow safepoints in the branch shadow, that will
2175 // cause a number of difficulties
2176 if ( avail_pipeline->instructionCount() == 1 &&
2177 !avail_pipeline->hasMultipleBundles() &&
2178 !avail_pipeline->hasBranchDelay() &&
2179 Pipeline::instr_has_unit_size() &&
2180 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2181 NodeFitsInBundle(d) &&
2182 !node_bundling(d)->used_in_delay()) {
2184 if (d->is_Mach() && !d->is_MachSafePoint()) {
2185 // A node that fits in the delay slot was found, so we need to
2186 // set the appropriate bits in the bundle pipeline information so
2187 // that it correctly indicates resource usage. Later, when we
2188 // attempt to add this instruction to the bundle, we will skip
2189 // setting the resource usage.
2190 _unconditional_delay_slot = d;
2191 node_bundling(n)->set_use_unconditional_delay();
2192 node_bundling(d)->set_used_in_unconditional_delay();
2193 _bundle_use.add_usage(avail_pipeline->resourceUse());
2194 _current_latency[d->_idx] = _bundle_cycle_number;
2195 _next_node = d;
2196 ++_bundle_instr_count;
2197 #ifndef PRODUCT
2198 _unconditional_delays++;
2199 #endif
2200 break;
2201 }
2202 }
2203 }
2204 }
2206 // No delay slot, add a nop to the usage
2207 if (!_unconditional_delay_slot) {
2208 // See if adding an instruction in the delay slot will overflow
2209 // the bundle.
2210 if (!NodeFitsInBundle(_nop)) {
2211 #ifndef PRODUCT
2212 if (_cfg->C->trace_opto_output())
2213 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2214 #endif
2215 step(1);
2216 }
2218 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2219 _next_node = _nop;
2220 ++_bundle_instr_count;
2221 }
2223 // See if the instruction in the delay slot requires a
2224 // step of the bundles
2225 if (!NodeFitsInBundle(n)) {
2226 #ifndef PRODUCT
2227 if (_cfg->C->trace_opto_output())
2228 tty->print("# *** STEP(branch won't fit) ***\n");
2229 #endif
2230 // Update the state information
2231 _bundle_instr_count = 0;
2232 _bundle_cycle_number += 1;
2233 _bundle_use.step(1);
2234 }
2235 }
2237 // Get the number of instructions
2238 uint instruction_count = node_pipeline->instructionCount();
2239 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2240 instruction_count = 0;
2242 // Compute the latency information
2243 uint delay = 0;
2245 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2246 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2247 if (relative_latency < 0)
2248 relative_latency = 0;
2250 delay = _bundle_use.full_latency(relative_latency, node_usage);
2252 // Does not fit in this bundle, start a new one
2253 if (delay > 0) {
2254 step(delay);
2256 #ifndef PRODUCT
2257 if (_cfg->C->trace_opto_output())
2258 tty->print("# *** STEP(%d) ***\n", delay);
2259 #endif
2260 }
2261 }
2263 // If this was placed in the delay slot, ignore it
2264 if (n != _unconditional_delay_slot) {
2266 if (delay == 0) {
2267 if (node_pipeline->hasMultipleBundles()) {
2268 #ifndef PRODUCT
2269 if (_cfg->C->trace_opto_output())
2270 tty->print("# *** STEP(multiple instructions) ***\n");
2271 #endif
2272 step(1);
2273 }
2275 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2276 #ifndef PRODUCT
2277 if (_cfg->C->trace_opto_output())
2278 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2279 instruction_count + _bundle_instr_count,
2280 Pipeline::_max_instrs_per_cycle);
2281 #endif
2282 step(1);
2283 }
2284 }
2286 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2287 _bundle_instr_count++;
2289 // Set the node's latency
2290 _current_latency[n->_idx] = _bundle_cycle_number;
2292 // Now merge the functional unit information
2293 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2294 _bundle_use.add_usage(node_usage);
2296 // Increment the number of instructions in this bundle
2297 _bundle_instr_count += instruction_count;
2299 // Remember this node for later
2300 if (n->is_Mach())
2301 _next_node = n;
2302 }
2304 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2305 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2306 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2307 // into the block. All other scheduled nodes get put in the schedule here.
2308 int op = n->Opcode();
2309 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2310 (op != Op_Node && // Not an unused antidepedence node and
2311 // not an unallocated boxlock
2312 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2314 // Push any trailing projections
2315 if( bb->_nodes[bb->_nodes.size()-1] != n ) {
2316 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2317 Node *foi = n->fast_out(i);
2318 if( foi->is_Proj() )
2319 _scheduled.push(foi);
2320 }
2321 }
2323 // Put the instruction in the schedule list
2324 _scheduled.push(n);
2325 }
2327 #ifndef PRODUCT
2328 if (_cfg->C->trace_opto_output())
2329 dump_available();
2330 #endif
2332 // Walk all the definitions, decrementing use counts, and
2333 // if a definition has a 0 use count, place it in the available list.
2334 DecrementUseCounts(n,bb);
2335 }
2337 //------------------------------ComputeUseCount--------------------------------
2338 // This method sets the use count within a basic block. We will ignore all
2339 // uses outside the current basic block. As we are doing a backwards walk,
2340 // any node we reach that has a use count of 0 may be scheduled. This also
2341 // avoids the problem of cyclic references from phi nodes, as long as phi
2342 // nodes are at the front of the basic block. This method also initializes
2343 // the available list to the set of instructions that have no uses within this
2344 // basic block.
2345 void Scheduling::ComputeUseCount(const Block *bb) {
2346 #ifndef PRODUCT
2347 if (_cfg->C->trace_opto_output())
2348 tty->print("# -> ComputeUseCount\n");
2349 #endif
2351 // Clear the list of available and scheduled instructions, just in case
2352 _available.clear();
2353 _scheduled.clear();
2355 // No delay slot specified
2356 _unconditional_delay_slot = NULL;
2358 #ifdef ASSERT
2359 for( uint i=0; i < bb->_nodes.size(); i++ )
2360 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );
2361 #endif
2363 // Force the _uses count to never go to zero for unscheduable pieces
2364 // of the block
2365 for( uint k = 0; k < _bb_start; k++ )
2366 _uses[bb->_nodes[k]->_idx] = 1;
2367 for( uint l = _bb_end; l < bb->_nodes.size(); l++ )
2368 _uses[bb->_nodes[l]->_idx] = 1;
2370 // Iterate backwards over the instructions in the block. Don't count the
2371 // branch projections at end or the block header instructions.
2372 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2373 Node *n = bb->_nodes[j];
2374 if( n->is_Proj() ) continue; // Projections handled another way
2376 // Account for all uses
2377 for ( uint k = 0; k < n->len(); k++ ) {
2378 Node *inp = n->in(k);
2379 if (!inp) continue;
2380 assert(inp != n, "no cycles allowed" );
2381 if( _bbs[inp->_idx] == bb ) { // Block-local use?
2382 if( inp->is_Proj() ) // Skip through Proj's
2383 inp = inp->in(0);
2384 ++_uses[inp->_idx]; // Count 1 block-local use
2385 }
2386 }
2388 // If this instruction has a 0 use count, then it is available
2389 if (!_uses[n->_idx]) {
2390 _current_latency[n->_idx] = _bundle_cycle_number;
2391 AddNodeToAvailableList(n);
2392 }
2394 #ifndef PRODUCT
2395 if (_cfg->C->trace_opto_output()) {
2396 tty->print("# uses: %3d: ", _uses[n->_idx]);
2397 n->dump();
2398 }
2399 #endif
2400 }
2402 #ifndef PRODUCT
2403 if (_cfg->C->trace_opto_output())
2404 tty->print("# <- ComputeUseCount\n");
2405 #endif
2406 }
2408 // This routine performs scheduling on each basic block in reverse order,
2409 // using instruction latencies and taking into account function unit
2410 // availability.
2411 void Scheduling::DoScheduling() {
2412 #ifndef PRODUCT
2413 if (_cfg->C->trace_opto_output())
2414 tty->print("# -> DoScheduling\n");
2415 #endif
2417 Block *succ_bb = NULL;
2418 Block *bb;
2420 // Walk over all the basic blocks in reverse order
2421 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) {
2422 bb = _cfg->_blocks[i];
2424 #ifndef PRODUCT
2425 if (_cfg->C->trace_opto_output()) {
2426 tty->print("# Schedule BB#%03d (initial)\n", i);
2427 for (uint j = 0; j < bb->_nodes.size(); j++)
2428 bb->_nodes[j]->dump();
2429 }
2430 #endif
2432 // On the head node, skip processing
2433 if( bb == _cfg->_broot )
2434 continue;
2436 // Skip empty, connector blocks
2437 if (bb->is_connector())
2438 continue;
2440 // If the following block is not the sole successor of
2441 // this one, then reset the pipeline information
2442 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2443 #ifndef PRODUCT
2444 if (_cfg->C->trace_opto_output()) {
2445 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2446 _next_node->_idx, _bundle_instr_count);
2447 }
2448 #endif
2449 step_and_clear();
2450 }
2452 // Leave untouched the starting instruction, any Phis, a CreateEx node
2453 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction.
2454 _bb_end = bb->_nodes.size()-1;
2455 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2456 Node *n = bb->_nodes[_bb_start];
2457 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2458 // Also, MachIdealNodes do not get scheduled
2459 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2460 MachNode *mach = n->as_Mach();
2461 int iop = mach->ideal_Opcode();
2462 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2463 if( iop == Op_Con ) continue; // Do not schedule Top
2464 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2465 mach->pipeline() == MachNode::pipeline_class() &&
2466 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2467 continue;
2468 break; // Funny loop structure to be sure...
2469 }
2470 // Compute last "interesting" instruction in block - last instruction we
2471 // might schedule. _bb_end points just after last schedulable inst. We
2472 // normally schedule conditional branches (despite them being forced last
2473 // in the block), because they have delay slots we can fill. Calls all
2474 // have their delay slots filled in the template expansions, so we don't
2475 // bother scheduling them.
2476 Node *last = bb->_nodes[_bb_end];
2477 // Ignore trailing NOPs.
2478 while (_bb_end > 0 && last->is_Mach() &&
2479 last->as_Mach()->ideal_Opcode() == Op_Con) {
2480 last = bb->_nodes[--_bb_end];
2481 }
2482 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2483 if( last->is_Catch() ||
2484 // Exclude unreachable path case when Halt node is in a separate block.
2485 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2486 // There must be a prior call. Skip it.
2487 while( !bb->_nodes[--_bb_end]->is_MachCall() ) {
2488 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" );
2489 }
2490 } else if( last->is_MachNullCheck() ) {
2491 // Backup so the last null-checked memory instruction is
2492 // outside the schedulable range. Skip over the nullcheck,
2493 // projection, and the memory nodes.
2494 Node *mem = last->in(1);
2495 do {
2496 _bb_end--;
2497 } while (mem != bb->_nodes[_bb_end]);
2498 } else {
2499 // Set _bb_end to point after last schedulable inst.
2500 _bb_end++;
2501 }
2503 assert( _bb_start <= _bb_end, "inverted block ends" );
2505 // Compute the register antidependencies for the basic block
2506 ComputeRegisterAntidependencies(bb);
2507 if (_cfg->C->failing()) return; // too many D-U pinch points
2509 // Compute intra-bb latencies for the nodes
2510 ComputeLocalLatenciesForward(bb);
2512 // Compute the usage within the block, and set the list of all nodes
2513 // in the block that have no uses within the block.
2514 ComputeUseCount(bb);
2516 // Schedule the remaining instructions in the block
2517 while ( _available.size() > 0 ) {
2518 Node *n = ChooseNodeToBundle();
2519 AddNodeToBundle(n,bb);
2520 }
2522 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2523 #ifdef ASSERT
2524 for( uint l = _bb_start; l < _bb_end; l++ ) {
2525 Node *n = bb->_nodes[l];
2526 uint m;
2527 for( m = 0; m < _bb_end-_bb_start; m++ )
2528 if( _scheduled[m] == n )
2529 break;
2530 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2531 }
2532 #endif
2534 // Now copy the instructions (in reverse order) back to the block
2535 for ( uint k = _bb_start; k < _bb_end; k++ )
2536 bb->_nodes.map(k, _scheduled[_bb_end-k-1]);
2538 #ifndef PRODUCT
2539 if (_cfg->C->trace_opto_output()) {
2540 tty->print("# Schedule BB#%03d (final)\n", i);
2541 uint current = 0;
2542 for (uint j = 0; j < bb->_nodes.size(); j++) {
2543 Node *n = bb->_nodes[j];
2544 if( valid_bundle_info(n) ) {
2545 Bundle *bundle = node_bundling(n);
2546 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2547 tty->print("*** Bundle: ");
2548 bundle->dump();
2549 }
2550 n->dump();
2551 }
2552 }
2553 }
2554 #endif
2555 #ifdef ASSERT
2556 verify_good_schedule(bb,"after block local scheduling");
2557 #endif
2558 }
2560 #ifndef PRODUCT
2561 if (_cfg->C->trace_opto_output())
2562 tty->print("# <- DoScheduling\n");
2563 #endif
2565 // Record final node-bundling array location
2566 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2568 } // end DoScheduling
2570 //------------------------------verify_good_schedule---------------------------
2571 // Verify that no live-range used in the block is killed in the block by a
2572 // wrong DEF. This doesn't verify live-ranges that span blocks.
2574 // Check for edge existence. Used to avoid adding redundant precedence edges.
2575 static bool edge_from_to( Node *from, Node *to ) {
2576 for( uint i=0; i<from->len(); i++ )
2577 if( from->in(i) == to )
2578 return true;
2579 return false;
2580 }
2582 #ifdef ASSERT
2583 //------------------------------verify_do_def----------------------------------
2584 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2585 // Check for bad kills
2586 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2587 Node *prior_use = _reg_node[def];
2588 if( prior_use && !edge_from_to(prior_use,n) ) {
2589 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2590 n->dump();
2591 tty->print_cr("...");
2592 prior_use->dump();
2593 assert(edge_from_to(prior_use,n),msg);
2594 }
2595 _reg_node.map(def,NULL); // Kill live USEs
2596 }
2597 }
2599 //------------------------------verify_good_schedule---------------------------
2600 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2602 // Zap to something reasonable for the verify code
2603 _reg_node.clear();
2605 // Walk over the block backwards. Check to make sure each DEF doesn't
2606 // kill a live value (other than the one it's supposed to). Add each
2607 // USE to the live set.
2608 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {
2609 Node *n = b->_nodes[i];
2610 int n_op = n->Opcode();
2611 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2612 // Fat-proj kills a slew of registers
2613 RegMask rm = n->out_RegMask();// Make local copy
2614 while( rm.is_NotEmpty() ) {
2615 OptoReg::Name kill = rm.find_first_elem();
2616 rm.Remove(kill);
2617 verify_do_def( n, kill, msg );
2618 }
2619 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2620 // Get DEF'd registers the normal way
2621 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2622 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2623 }
2625 // Now make all USEs live
2626 for( uint i=1; i<n->req(); i++ ) {
2627 Node *def = n->in(i);
2628 assert(def != 0, "input edge required");
2629 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2630 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2631 if( OptoReg::is_valid(reg_lo) ) {
2632 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2633 _reg_node.map(reg_lo,n);
2634 }
2635 if( OptoReg::is_valid(reg_hi) ) {
2636 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2637 _reg_node.map(reg_hi,n);
2638 }
2639 }
2641 }
2643 // Zap to something reasonable for the Antidependence code
2644 _reg_node.clear();
2645 }
2646 #endif
2648 // Conditionally add precedence edges. Avoid putting edges on Projs.
2649 static void add_prec_edge_from_to( Node *from, Node *to ) {
2650 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2651 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2652 from = from->in(0);
2653 }
2654 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2655 !edge_from_to( from, to ) ) // Avoid duplicate edge
2656 from->add_prec(to);
2657 }
2659 //------------------------------anti_do_def------------------------------------
2660 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2661 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2662 return;
2664 Node *pinch = _reg_node[def_reg]; // Get pinch point
2665 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet?
2666 is_def ) { // Check for a true def (not a kill)
2667 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2668 return;
2669 }
2671 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2672 debug_only( def = (Node*)0xdeadbeef; )
2674 // After some number of kills there _may_ be a later def
2675 Node *later_def = NULL;
2677 // Finding a kill requires a real pinch-point.
2678 // Check for not already having a pinch-point.
2679 // Pinch points are Op_Node's.
2680 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2681 later_def = pinch; // Must be def/kill as optimistic pinch-point
2682 if ( _pinch_free_list.size() > 0) {
2683 pinch = _pinch_free_list.pop();
2684 } else {
2685 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be
2686 }
2687 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2688 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2689 return;
2690 }
2691 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init)
2692 _reg_node.map(def_reg,pinch); // Record pinch-point
2693 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2694 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2695 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2696 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2697 later_def = NULL; // and no later def
2698 }
2699 pinch->set_req(0,later_def); // Hook later def so we can find it
2700 } else { // Else have valid pinch point
2701 if( pinch->in(0) ) // If there is a later-def
2702 later_def = pinch->in(0); // Get it
2703 }
2705 // Add output-dependence edge from later def to kill
2706 if( later_def ) // If there is some original def
2707 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2709 // See if current kill is also a use, and so is forced to be the pinch-point.
2710 if( pinch->Opcode() == Op_Node ) {
2711 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2712 for( uint i=1; i<uses->req(); i++ ) {
2713 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2714 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2715 // Yes, found a use/kill pinch-point
2716 pinch->set_req(0,NULL); //
2717 pinch->replace_by(kill); // Move anti-dep edges up
2718 pinch = kill;
2719 _reg_node.map(def_reg,pinch);
2720 return;
2721 }
2722 }
2723 }
2725 // Add edge from kill to pinch-point
2726 add_prec_edge_from_to(kill,pinch);
2727 }
2729 //------------------------------anti_do_use------------------------------------
2730 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2731 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2732 return;
2733 Node *pinch = _reg_node[use_reg]; // Get pinch point
2734 // Check for no later def_reg/kill in block
2735 if( pinch && _bbs[pinch->_idx] == b &&
2736 // Use has to be block-local as well
2737 _bbs[use->_idx] == b ) {
2738 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2739 pinch->req() == 1 ) { // pinch not yet in block?
2740 pinch->del_req(0); // yank pointer to later-def, also set flag
2741 // Insert the pinch-point in the block just after the last use
2742 b->_nodes.insert(b->find_node(use)+1,pinch);
2743 _bb_end++; // Increase size scheduled region in block
2744 }
2746 add_prec_edge_from_to(pinch,use);
2747 }
2748 }
2750 //------------------------------ComputeRegisterAntidependences-----------------
2751 // We insert antidependences between the reads and following write of
2752 // allocated registers to prevent illegal code motion. Hopefully, the
2753 // number of added references should be fairly small, especially as we
2754 // are only adding references within the current basic block.
2755 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2757 #ifdef ASSERT
2758 verify_good_schedule(b,"before block local scheduling");
2759 #endif
2761 // A valid schedule, for each register independently, is an endless cycle
2762 // of: a def, then some uses (connected to the def by true dependencies),
2763 // then some kills (defs with no uses), finally the cycle repeats with a new
2764 // def. The uses are allowed to float relative to each other, as are the
2765 // kills. No use is allowed to slide past a kill (or def). This requires
2766 // antidependencies between all uses of a single def and all kills that
2767 // follow, up to the next def. More edges are redundant, because later defs
2768 // & kills are already serialized with true or antidependencies. To keep
2769 // the edge count down, we add a 'pinch point' node if there's more than
2770 // one use or more than one kill/def.
2772 // We add dependencies in one bottom-up pass.
2774 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2776 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2777 // register. If not, we record the DEF/KILL in _reg_node, the
2778 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2779 // "pinch point", a new Node that's in the graph but not in the block.
2780 // We put edges from the prior and current DEF/KILLs to the pinch point.
2781 // We put the pinch point in _reg_node. If there's already a pinch point
2782 // we merely add an edge from the current DEF/KILL to the pinch point.
2784 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2785 // put an edge from the pinch point to the USE.
2787 // To be expedient, the _reg_node array is pre-allocated for the whole
2788 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2789 // or a valid def/kill/pinch-point, or a leftover node from some prior
2790 // block. Leftover node from some prior block is treated like a NULL (no
2791 // prior def, so no anti-dependence needed). Valid def is distinguished by
2792 // it being in the current block.
2793 bool fat_proj_seen = false;
2794 uint last_safept = _bb_end-1;
2795 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL;
2796 Node* last_safept_node = end_node;
2797 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2798 Node *n = b->_nodes[i];
2799 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2800 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2801 // Fat-proj kills a slew of registers
2802 // This can add edges to 'n' and obscure whether or not it was a def,
2803 // hence the is_def flag.
2804 fat_proj_seen = true;
2805 RegMask rm = n->out_RegMask();// Make local copy
2806 while( rm.is_NotEmpty() ) {
2807 OptoReg::Name kill = rm.find_first_elem();
2808 rm.Remove(kill);
2809 anti_do_def( b, n, kill, is_def );
2810 }
2811 } else {
2812 // Get DEF'd registers the normal way
2813 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2814 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2815 }
2817 // Kill projections on a branch should appear to occur on the
2818 // branch, not afterwards, so grab the masks from the projections
2819 // and process them.
2820 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
2821 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2822 Node* use = n->fast_out(i);
2823 if (use->is_Proj()) {
2824 RegMask rm = use->out_RegMask();// Make local copy
2825 while( rm.is_NotEmpty() ) {
2826 OptoReg::Name kill = rm.find_first_elem();
2827 rm.Remove(kill);
2828 anti_do_def( b, n, kill, false );
2829 }
2830 }
2831 }
2832 }
2834 // Check each register used by this instruction for a following DEF/KILL
2835 // that must occur afterward and requires an anti-dependence edge.
2836 for( uint j=0; j<n->req(); j++ ) {
2837 Node *def = n->in(j);
2838 if( def ) {
2839 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2840 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2841 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2842 }
2843 }
2844 // Do not allow defs of new derived values to float above GC
2845 // points unless the base is definitely available at the GC point.
2847 Node *m = b->_nodes[i];
2849 // Add precedence edge from following safepoint to use of derived pointer
2850 if( last_safept_node != end_node &&
2851 m != last_safept_node) {
2852 for (uint k = 1; k < m->req(); k++) {
2853 const Type *t = m->in(k)->bottom_type();
2854 if( t->isa_oop_ptr() &&
2855 t->is_ptr()->offset() != 0 ) {
2856 last_safept_node->add_prec( m );
2857 break;
2858 }
2859 }
2860 }
2862 if( n->jvms() ) { // Precedence edge from derived to safept
2863 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2864 if( b->_nodes[last_safept] != last_safept_node ) {
2865 last_safept = b->find_node(last_safept_node);
2866 }
2867 for( uint j=last_safept; j > i; j-- ) {
2868 Node *mach = b->_nodes[j];
2869 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2870 mach->add_prec( n );
2871 }
2872 last_safept = i;
2873 last_safept_node = m;
2874 }
2875 }
2877 if (fat_proj_seen) {
2878 // Garbage collect pinch nodes that were not consumed.
2879 // They are usually created by a fat kill MachProj for a call.
2880 garbage_collect_pinch_nodes();
2881 }
2882 }
2884 //------------------------------garbage_collect_pinch_nodes-------------------------------
2886 // Garbage collect pinch nodes for reuse by other blocks.
2887 //
2888 // The block scheduler's insertion of anti-dependence
2889 // edges creates many pinch nodes when the block contains
2890 // 2 or more Calls. A pinch node is used to prevent a
2891 // combinatorial explosion of edges. If a set of kills for a
2892 // register is anti-dependent on a set of uses (or defs), rather
2893 // than adding an edge in the graph between each pair of kill
2894 // and use (or def), a pinch is inserted between them:
2895 //
2896 // use1 use2 use3
2897 // \ | /
2898 // \ | /
2899 // pinch
2900 // / | \
2901 // / | \
2902 // kill1 kill2 kill3
2903 //
2904 // One pinch node is created per register killed when
2905 // the second call is encountered during a backwards pass
2906 // over the block. Most of these pinch nodes are never
2907 // wired into the graph because the register is never
2908 // used or def'ed in the block.
2909 //
2910 void Scheduling::garbage_collect_pinch_nodes() {
2911 #ifndef PRODUCT
2912 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2913 #endif
2914 int trace_cnt = 0;
2915 for (uint k = 0; k < _reg_node.Size(); k++) {
2916 Node* pinch = _reg_node[k];
2917 if (pinch != NULL && pinch->Opcode() == Op_Node &&
2918 // no predecence input edges
2919 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2920 cleanup_pinch(pinch);
2921 _pinch_free_list.push(pinch);
2922 _reg_node.map(k, NULL);
2923 #ifndef PRODUCT
2924 if (_cfg->C->trace_opto_output()) {
2925 trace_cnt++;
2926 if (trace_cnt > 40) {
2927 tty->print("\n");
2928 trace_cnt = 0;
2929 }
2930 tty->print(" %d", pinch->_idx);
2931 }
2932 #endif
2933 }
2934 }
2935 #ifndef PRODUCT
2936 if (_cfg->C->trace_opto_output()) tty->print("\n");
2937 #endif
2938 }
2940 // Clean up a pinch node for reuse.
2941 void Scheduling::cleanup_pinch( Node *pinch ) {
2942 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2944 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2945 Node* use = pinch->last_out(i);
2946 uint uses_found = 0;
2947 for (uint j = use->req(); j < use->len(); j++) {
2948 if (use->in(j) == pinch) {
2949 use->rm_prec(j);
2950 uses_found++;
2951 }
2952 }
2953 assert(uses_found > 0, "must be a precedence edge");
2954 i -= uses_found; // we deleted 1 or more copies of this edge
2955 }
2956 // May have a later_def entry
2957 pinch->set_req(0, NULL);
2958 }
2960 //------------------------------print_statistics-------------------------------
2961 #ifndef PRODUCT
2963 void Scheduling::dump_available() const {
2964 tty->print("#Availist ");
2965 for (uint i = 0; i < _available.size(); i++)
2966 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2967 tty->cr();
2968 }
2970 // Print Scheduling Statistics
2971 void Scheduling::print_statistics() {
2972 // Print the size added by nops for bundling
2973 tty->print("Nops added %d bytes to total of %d bytes",
2974 _total_nop_size, _total_method_size);
2975 if (_total_method_size > 0)
2976 tty->print(", for %.2f%%",
2977 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2978 tty->print("\n");
2980 // Print the number of branch shadows filled
2981 if (Pipeline::_branch_has_delay_slot) {
2982 tty->print("Of %d branches, %d had unconditional delay slots filled",
2983 _total_branches, _total_unconditional_delays);
2984 if (_total_branches > 0)
2985 tty->print(", for %.2f%%",
2986 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2987 tty->print("\n");
2988 }
2990 uint total_instructions = 0, total_bundles = 0;
2992 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2993 uint bundle_count = _total_instructions_per_bundle[i];
2994 total_instructions += bundle_count * i;
2995 total_bundles += bundle_count;
2996 }
2998 if (total_bundles > 0)
2999 tty->print("Average ILP (excluding nops) is %.2f\n",
3000 ((double)total_instructions) / ((double)total_bundles));
3001 }
3002 #endif