Mon, 28 Apr 2008 08:08:12 -0700
Merge
1 /*
2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_compile.cpp.incl"
28 /// Support for intrinsics.
30 // Return the index at which m must be inserted (or already exists).
31 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
33 #ifdef ASSERT
34 for (int i = 1; i < _intrinsics->length(); i++) {
35 CallGenerator* cg1 = _intrinsics->at(i-1);
36 CallGenerator* cg2 = _intrinsics->at(i);
37 assert(cg1->method() != cg2->method()
38 ? cg1->method() < cg2->method()
39 : cg1->is_virtual() < cg2->is_virtual(),
40 "compiler intrinsics list must stay sorted");
41 }
42 #endif
43 // Binary search sorted list, in decreasing intervals [lo, hi].
44 int lo = 0, hi = _intrinsics->length()-1;
45 while (lo <= hi) {
46 int mid = (uint)(hi + lo) / 2;
47 ciMethod* mid_m = _intrinsics->at(mid)->method();
48 if (m < mid_m) {
49 hi = mid-1;
50 } else if (m > mid_m) {
51 lo = mid+1;
52 } else {
53 // look at minor sort key
54 bool mid_virt = _intrinsics->at(mid)->is_virtual();
55 if (is_virtual < mid_virt) {
56 hi = mid-1;
57 } else if (is_virtual > mid_virt) {
58 lo = mid+1;
59 } else {
60 return mid; // exact match
61 }
62 }
63 }
64 return lo; // inexact match
65 }
67 void Compile::register_intrinsic(CallGenerator* cg) {
68 if (_intrinsics == NULL) {
69 _intrinsics = new GrowableArray<CallGenerator*>(60);
70 }
71 // This code is stolen from ciObjectFactory::insert.
72 // Really, GrowableArray should have methods for
73 // insert_at, remove_at, and binary_search.
74 int len = _intrinsics->length();
75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
76 if (index == len) {
77 _intrinsics->append(cg);
78 } else {
79 #ifdef ASSERT
80 CallGenerator* oldcg = _intrinsics->at(index);
81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
82 #endif
83 _intrinsics->append(_intrinsics->at(len-1));
84 int pos;
85 for (pos = len-2; pos >= index; pos--) {
86 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
87 }
88 _intrinsics->at_put(index, cg);
89 }
90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
91 }
93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
94 assert(m->is_loaded(), "don't try this on unloaded methods");
95 if (_intrinsics != NULL) {
96 int index = intrinsic_insertion_index(m, is_virtual);
97 if (index < _intrinsics->length()
98 && _intrinsics->at(index)->method() == m
99 && _intrinsics->at(index)->is_virtual() == is_virtual) {
100 return _intrinsics->at(index);
101 }
102 }
103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
104 if (m->intrinsic_id() != vmIntrinsics::_none) {
105 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
106 if (cg != NULL) {
107 // Save it for next time:
108 register_intrinsic(cg);
109 return cg;
110 } else {
111 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
112 }
113 }
114 return NULL;
115 }
117 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
118 // in library_call.cpp.
121 #ifndef PRODUCT
122 // statistics gathering...
124 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
125 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
127 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
128 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
129 int oflags = _intrinsic_hist_flags[id];
130 assert(flags != 0, "what happened?");
131 if (is_virtual) {
132 flags |= _intrinsic_virtual;
133 }
134 bool changed = (flags != oflags);
135 if ((flags & _intrinsic_worked) != 0) {
136 juint count = (_intrinsic_hist_count[id] += 1);
137 if (count == 1) {
138 changed = true; // first time
139 }
140 // increment the overall count also:
141 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
142 }
143 if (changed) {
144 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
145 // Something changed about the intrinsic's virtuality.
146 if ((flags & _intrinsic_virtual) != 0) {
147 // This is the first use of this intrinsic as a virtual call.
148 if (oflags != 0) {
149 // We already saw it as a non-virtual, so note both cases.
150 flags |= _intrinsic_both;
151 }
152 } else if ((oflags & _intrinsic_both) == 0) {
153 // This is the first use of this intrinsic as a non-virtual
154 flags |= _intrinsic_both;
155 }
156 }
157 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
158 }
159 // update the overall flags also:
160 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
161 return changed;
162 }
164 static char* format_flags(int flags, char* buf) {
165 buf[0] = 0;
166 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
167 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
168 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
169 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
170 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
171 if (buf[0] == 0) strcat(buf, ",");
172 assert(buf[0] == ',', "must be");
173 return &buf[1];
174 }
176 void Compile::print_intrinsic_statistics() {
177 char flagsbuf[100];
178 ttyLocker ttyl;
179 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
180 tty->print_cr("Compiler intrinsic usage:");
181 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
182 if (total == 0) total = 1; // avoid div0 in case of no successes
183 #define PRINT_STAT_LINE(name, c, f) \
184 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
185 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
186 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
187 int flags = _intrinsic_hist_flags[id];
188 juint count = _intrinsic_hist_count[id];
189 if ((flags | count) != 0) {
190 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
191 }
192 }
193 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
194 if (xtty != NULL) xtty->tail("statistics");
195 }
197 void Compile::print_statistics() {
198 { ttyLocker ttyl;
199 if (xtty != NULL) xtty->head("statistics type='opto'");
200 Parse::print_statistics();
201 PhaseCCP::print_statistics();
202 PhaseRegAlloc::print_statistics();
203 Scheduling::print_statistics();
204 PhasePeephole::print_statistics();
205 PhaseIdealLoop::print_statistics();
206 if (xtty != NULL) xtty->tail("statistics");
207 }
208 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
209 // put this under its own <statistics> element.
210 print_intrinsic_statistics();
211 }
212 }
213 #endif //PRODUCT
215 // Support for bundling info
216 Bundle* Compile::node_bundling(const Node *n) {
217 assert(valid_bundle_info(n), "oob");
218 return &_node_bundling_base[n->_idx];
219 }
221 bool Compile::valid_bundle_info(const Node *n) {
222 return (_node_bundling_limit > n->_idx);
223 }
226 // Identify all nodes that are reachable from below, useful.
227 // Use breadth-first pass that records state in a Unique_Node_List,
228 // recursive traversal is slower.
229 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
230 int estimated_worklist_size = unique();
231 useful.map( estimated_worklist_size, NULL ); // preallocate space
233 // Initialize worklist
234 if (root() != NULL) { useful.push(root()); }
235 // If 'top' is cached, declare it useful to preserve cached node
236 if( cached_top_node() ) { useful.push(cached_top_node()); }
238 // Push all useful nodes onto the list, breadthfirst
239 for( uint next = 0; next < useful.size(); ++next ) {
240 assert( next < unique(), "Unique useful nodes < total nodes");
241 Node *n = useful.at(next);
242 uint max = n->len();
243 for( uint i = 0; i < max; ++i ) {
244 Node *m = n->in(i);
245 if( m == NULL ) continue;
246 useful.push(m);
247 }
248 }
249 }
251 // Disconnect all useless nodes by disconnecting those at the boundary.
252 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
253 uint next = 0;
254 while( next < useful.size() ) {
255 Node *n = useful.at(next++);
256 // Use raw traversal of out edges since this code removes out edges
257 int max = n->outcnt();
258 for (int j = 0; j < max; ++j ) {
259 Node* child = n->raw_out(j);
260 if( ! useful.member(child) ) {
261 assert( !child->is_top() || child != top(),
262 "If top is cached in Compile object it is in useful list");
263 // Only need to remove this out-edge to the useless node
264 n->raw_del_out(j);
265 --j;
266 --max;
267 }
268 }
269 if (n->outcnt() == 1 && n->has_special_unique_user()) {
270 record_for_igvn( n->unique_out() );
271 }
272 }
273 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
274 }
276 //------------------------------frame_size_in_words-----------------------------
277 // frame_slots in units of words
278 int Compile::frame_size_in_words() const {
279 // shift is 0 in LP32 and 1 in LP64
280 const int shift = (LogBytesPerWord - LogBytesPerInt);
281 int words = _frame_slots >> shift;
282 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
283 return words;
284 }
286 // ============================================================================
287 //------------------------------CompileWrapper---------------------------------
288 class CompileWrapper : public StackObj {
289 Compile *const _compile;
290 public:
291 CompileWrapper(Compile* compile);
293 ~CompileWrapper();
294 };
296 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
297 // the Compile* pointer is stored in the current ciEnv:
298 ciEnv* env = compile->env();
299 assert(env == ciEnv::current(), "must already be a ciEnv active");
300 assert(env->compiler_data() == NULL, "compile already active?");
301 env->set_compiler_data(compile);
302 assert(compile == Compile::current(), "sanity");
304 compile->set_type_dict(NULL);
305 compile->set_type_hwm(NULL);
306 compile->set_type_last_size(0);
307 compile->set_last_tf(NULL, NULL);
308 compile->set_indexSet_arena(NULL);
309 compile->set_indexSet_free_block_list(NULL);
310 compile->init_type_arena();
311 Type::Initialize(compile);
312 _compile->set_scratch_buffer_blob(NULL);
313 _compile->begin_method();
314 }
315 CompileWrapper::~CompileWrapper() {
316 if (_compile->failing()) {
317 _compile->print_method("Failed");
318 }
319 _compile->end_method();
320 if (_compile->scratch_buffer_blob() != NULL)
321 BufferBlob::free(_compile->scratch_buffer_blob());
322 _compile->env()->set_compiler_data(NULL);
323 }
326 //----------------------------print_compile_messages---------------------------
327 void Compile::print_compile_messages() {
328 #ifndef PRODUCT
329 // Check if recompiling
330 if (_subsume_loads == false && PrintOpto) {
331 // Recompiling without allowing machine instructions to subsume loads
332 tty->print_cr("*********************************************************");
333 tty->print_cr("** Bailout: Recompile without subsuming loads **");
334 tty->print_cr("*********************************************************");
335 }
336 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
337 // Recompiling without escape analysis
338 tty->print_cr("*********************************************************");
339 tty->print_cr("** Bailout: Recompile without escape analysis **");
340 tty->print_cr("*********************************************************");
341 }
342 if (env()->break_at_compile()) {
343 // Open the debugger when compiing this method.
344 tty->print("### Breaking when compiling: ");
345 method()->print_short_name();
346 tty->cr();
347 BREAKPOINT;
348 }
350 if( PrintOpto ) {
351 if (is_osr_compilation()) {
352 tty->print("[OSR]%3d", _compile_id);
353 } else {
354 tty->print("%3d", _compile_id);
355 }
356 }
357 #endif
358 }
361 void Compile::init_scratch_buffer_blob() {
362 if( scratch_buffer_blob() != NULL ) return;
364 // Construct a temporary CodeBuffer to have it construct a BufferBlob
365 // Cache this BufferBlob for this compile.
366 ResourceMark rm;
367 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size);
368 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size);
369 // Record the buffer blob for next time.
370 set_scratch_buffer_blob(blob);
371 guarantee(scratch_buffer_blob() != NULL, "Need BufferBlob for code generation");
373 // Initialize the relocation buffers
374 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size;
375 set_scratch_locs_memory(locs_buf);
376 }
379 //-----------------------scratch_emit_size-------------------------------------
380 // Helper function that computes size by emitting code
381 uint Compile::scratch_emit_size(const Node* n) {
382 // Emit into a trash buffer and count bytes emitted.
383 // This is a pretty expensive way to compute a size,
384 // but it works well enough if seldom used.
385 // All common fixed-size instructions are given a size
386 // method by the AD file.
387 // Note that the scratch buffer blob and locs memory are
388 // allocated at the beginning of the compile task, and
389 // may be shared by several calls to scratch_emit_size.
390 // The allocation of the scratch buffer blob is particularly
391 // expensive, since it has to grab the code cache lock.
392 BufferBlob* blob = this->scratch_buffer_blob();
393 assert(blob != NULL, "Initialize BufferBlob at start");
394 assert(blob->size() > MAX_inst_size, "sanity");
395 relocInfo* locs_buf = scratch_locs_memory();
396 address blob_begin = blob->instructions_begin();
397 address blob_end = (address)locs_buf;
398 assert(blob->instructions_contains(blob_end), "sanity");
399 CodeBuffer buf(blob_begin, blob_end - blob_begin);
400 buf.initialize_consts_size(MAX_const_size);
401 buf.initialize_stubs_size(MAX_stubs_size);
402 assert(locs_buf != NULL, "sanity");
403 int lsize = MAX_locs_size / 2;
404 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize);
405 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize);
406 n->emit(buf, this->regalloc());
407 return buf.code_size();
408 }
411 // ============================================================================
412 //------------------------------Compile standard-------------------------------
413 debug_only( int Compile::_debug_idx = 100000; )
415 // Compile a method. entry_bci is -1 for normal compilations and indicates
416 // the continuation bci for on stack replacement.
419 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
420 : Phase(Compiler),
421 _env(ci_env),
422 _log(ci_env->log()),
423 _compile_id(ci_env->compile_id()),
424 _save_argument_registers(false),
425 _stub_name(NULL),
426 _stub_function(NULL),
427 _stub_entry_point(NULL),
428 _method(target),
429 _entry_bci(osr_bci),
430 _initial_gvn(NULL),
431 _for_igvn(NULL),
432 _warm_calls(NULL),
433 _subsume_loads(subsume_loads),
434 _do_escape_analysis(do_escape_analysis),
435 _failure_reason(NULL),
436 _code_buffer("Compile::Fill_buffer"),
437 _orig_pc_slot(0),
438 _orig_pc_slot_offset_in_bytes(0),
439 _node_bundling_limit(0),
440 _node_bundling_base(NULL),
441 #ifndef PRODUCT
442 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
443 _printer(IdealGraphPrinter::printer()),
444 #endif
445 _congraph(NULL) {
446 C = this;
448 CompileWrapper cw(this);
449 #ifndef PRODUCT
450 if (TimeCompiler2) {
451 tty->print(" ");
452 target->holder()->name()->print();
453 tty->print(".");
454 target->print_short_name();
455 tty->print(" ");
456 }
457 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
458 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
459 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
460 if (!print_opto_assembly) {
461 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
462 if (print_assembly && !Disassembler::can_decode()) {
463 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
464 print_opto_assembly = true;
465 }
466 }
467 set_print_assembly(print_opto_assembly);
468 #endif
470 if (ProfileTraps) {
471 // Make sure the method being compiled gets its own MDO,
472 // so we can at least track the decompile_count().
473 method()->build_method_data();
474 }
476 Init(::AliasLevel);
479 print_compile_messages();
481 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
482 _ilt = InlineTree::build_inline_tree_root();
483 else
484 _ilt = NULL;
486 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
487 assert(num_alias_types() >= AliasIdxRaw, "");
489 #define MINIMUM_NODE_HASH 1023
490 // Node list that Iterative GVN will start with
491 Unique_Node_List for_igvn(comp_arena());
492 set_for_igvn(&for_igvn);
494 // GVN that will be run immediately on new nodes
495 uint estimated_size = method()->code_size()*4+64;
496 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
497 PhaseGVN gvn(node_arena(), estimated_size);
498 set_initial_gvn(&gvn);
500 { // Scope for timing the parser
501 TracePhase t3("parse", &_t_parser, true);
503 // Put top into the hash table ASAP.
504 initial_gvn()->transform_no_reclaim(top());
506 // Set up tf(), start(), and find a CallGenerator.
507 CallGenerator* cg;
508 if (is_osr_compilation()) {
509 const TypeTuple *domain = StartOSRNode::osr_domain();
510 const TypeTuple *range = TypeTuple::make_range(method()->signature());
511 init_tf(TypeFunc::make(domain, range));
512 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
513 initial_gvn()->set_type_bottom(s);
514 init_start(s);
515 cg = CallGenerator::for_osr(method(), entry_bci());
516 } else {
517 // Normal case.
518 init_tf(TypeFunc::make(method()));
519 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
520 initial_gvn()->set_type_bottom(s);
521 init_start(s);
522 float past_uses = method()->interpreter_invocation_count();
523 float expected_uses = past_uses;
524 cg = CallGenerator::for_inline(method(), expected_uses);
525 }
526 if (failing()) return;
527 if (cg == NULL) {
528 record_method_not_compilable_all_tiers("cannot parse method");
529 return;
530 }
531 JVMState* jvms = build_start_state(start(), tf());
532 if ((jvms = cg->generate(jvms)) == NULL) {
533 record_method_not_compilable("method parse failed");
534 return;
535 }
536 GraphKit kit(jvms);
538 if (!kit.stopped()) {
539 // Accept return values, and transfer control we know not where.
540 // This is done by a special, unique ReturnNode bound to root.
541 return_values(kit.jvms());
542 }
544 if (kit.has_exceptions()) {
545 // Any exceptions that escape from this call must be rethrown
546 // to whatever caller is dynamically above us on the stack.
547 // This is done by a special, unique RethrowNode bound to root.
548 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
549 }
551 // Remove clutter produced by parsing.
552 if (!failing()) {
553 ResourceMark rm;
554 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
555 }
556 }
558 // Note: Large methods are capped off in do_one_bytecode().
559 if (failing()) return;
561 // After parsing, node notes are no longer automagic.
562 // They must be propagated by register_new_node_with_optimizer(),
563 // clone(), or the like.
564 set_default_node_notes(NULL);
566 for (;;) {
567 int successes = Inline_Warm();
568 if (failing()) return;
569 if (successes == 0) break;
570 }
572 // Drain the list.
573 Finish_Warm();
574 #ifndef PRODUCT
575 if (_printer) {
576 _printer->print_inlining(this);
577 }
578 #endif
580 if (failing()) return;
581 NOT_PRODUCT( verify_graph_edges(); )
583 // Perform escape analysis
584 if (_do_escape_analysis)
585 _congraph = new ConnectionGraph(this);
586 if (_congraph != NULL) {
587 NOT_PRODUCT( TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, TimeCompiler); )
588 _congraph->compute_escape();
589 if (failing()) return;
591 #ifndef PRODUCT
592 if (PrintEscapeAnalysis) {
593 _congraph->dump();
594 }
595 #endif
596 }
597 // Now optimize
598 Optimize();
599 if (failing()) return;
600 NOT_PRODUCT( verify_graph_edges(); )
602 #ifndef PRODUCT
603 if (PrintIdeal) {
604 ttyLocker ttyl; // keep the following output all in one block
605 // This output goes directly to the tty, not the compiler log.
606 // To enable tools to match it up with the compilation activity,
607 // be sure to tag this tty output with the compile ID.
608 if (xtty != NULL) {
609 xtty->head("ideal compile_id='%d'%s", compile_id(),
610 is_osr_compilation() ? " compile_kind='osr'" :
611 "");
612 }
613 root()->dump(9999);
614 if (xtty != NULL) {
615 xtty->tail("ideal");
616 }
617 }
618 #endif
620 // Now that we know the size of all the monitors we can add a fixed slot
621 // for the original deopt pc.
623 _orig_pc_slot = fixed_slots();
624 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
625 set_fixed_slots(next_slot);
627 // Now generate code
628 Code_Gen();
629 if (failing()) return;
631 // Check if we want to skip execution of all compiled code.
632 {
633 #ifndef PRODUCT
634 if (OptoNoExecute) {
635 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
636 return;
637 }
638 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
639 #endif
641 if (is_osr_compilation()) {
642 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
643 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
644 } else {
645 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
646 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
647 }
649 env()->register_method(_method, _entry_bci,
650 &_code_offsets,
651 _orig_pc_slot_offset_in_bytes,
652 code_buffer(),
653 frame_size_in_words(), _oop_map_set,
654 &_handler_table, &_inc_table,
655 compiler,
656 env()->comp_level(),
657 true, /*has_debug_info*/
658 has_unsafe_access()
659 );
660 }
661 }
663 //------------------------------Compile----------------------------------------
664 // Compile a runtime stub
665 Compile::Compile( ciEnv* ci_env,
666 TypeFunc_generator generator,
667 address stub_function,
668 const char *stub_name,
669 int is_fancy_jump,
670 bool pass_tls,
671 bool save_arg_registers,
672 bool return_pc )
673 : Phase(Compiler),
674 _env(ci_env),
675 _log(ci_env->log()),
676 _compile_id(-1),
677 _save_argument_registers(save_arg_registers),
678 _method(NULL),
679 _stub_name(stub_name),
680 _stub_function(stub_function),
681 _stub_entry_point(NULL),
682 _entry_bci(InvocationEntryBci),
683 _initial_gvn(NULL),
684 _for_igvn(NULL),
685 _warm_calls(NULL),
686 _orig_pc_slot(0),
687 _orig_pc_slot_offset_in_bytes(0),
688 _subsume_loads(true),
689 _do_escape_analysis(false),
690 _failure_reason(NULL),
691 _code_buffer("Compile::Fill_buffer"),
692 _node_bundling_limit(0),
693 _node_bundling_base(NULL),
694 #ifndef PRODUCT
695 _trace_opto_output(TraceOptoOutput),
696 _printer(NULL),
697 #endif
698 _congraph(NULL) {
699 C = this;
701 #ifndef PRODUCT
702 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
703 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
704 set_print_assembly(PrintFrameConverterAssembly);
705 #endif
706 CompileWrapper cw(this);
707 Init(/*AliasLevel=*/ 0);
708 init_tf((*generator)());
710 {
711 // The following is a dummy for the sake of GraphKit::gen_stub
712 Unique_Node_List for_igvn(comp_arena());
713 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
714 PhaseGVN gvn(Thread::current()->resource_area(),255);
715 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
716 gvn.transform_no_reclaim(top());
718 GraphKit kit;
719 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
720 }
722 NOT_PRODUCT( verify_graph_edges(); )
723 Code_Gen();
724 if (failing()) return;
727 // Entry point will be accessed using compile->stub_entry_point();
728 if (code_buffer() == NULL) {
729 Matcher::soft_match_failure();
730 } else {
731 if (PrintAssembly && (WizardMode || Verbose))
732 tty->print_cr("### Stub::%s", stub_name);
734 if (!failing()) {
735 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
737 // Make the NMethod
738 // For now we mark the frame as never safe for profile stackwalking
739 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
740 code_buffer(),
741 CodeOffsets::frame_never_safe,
742 // _code_offsets.value(CodeOffsets::Frame_Complete),
743 frame_size_in_words(),
744 _oop_map_set,
745 save_arg_registers);
746 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
748 _stub_entry_point = rs->entry_point();
749 }
750 }
751 }
753 #ifndef PRODUCT
754 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
755 if(PrintOpto && Verbose) {
756 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
757 }
758 }
759 #endif
761 void Compile::print_codes() {
762 }
764 //------------------------------Init-------------------------------------------
765 // Prepare for a single compilation
766 void Compile::Init(int aliaslevel) {
767 _unique = 0;
768 _regalloc = NULL;
770 _tf = NULL; // filled in later
771 _top = NULL; // cached later
772 _matcher = NULL; // filled in later
773 _cfg = NULL; // filled in later
775 set_24_bit_selection_and_mode(Use24BitFP, false);
777 _node_note_array = NULL;
778 _default_node_notes = NULL;
780 _immutable_memory = NULL; // filled in at first inquiry
782 // Globally visible Nodes
783 // First set TOP to NULL to give safe behavior during creation of RootNode
784 set_cached_top_node(NULL);
785 set_root(new (this, 3) RootNode());
786 // Now that you have a Root to point to, create the real TOP
787 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
788 set_recent_alloc(NULL, NULL);
790 // Create Debug Information Recorder to record scopes, oopmaps, etc.
791 env()->set_oop_recorder(new OopRecorder(comp_arena()));
792 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
793 env()->set_dependencies(new Dependencies(env()));
795 _fixed_slots = 0;
796 set_has_split_ifs(false);
797 set_has_loops(has_method() && method()->has_loops()); // first approximation
798 _deopt_happens = true; // start out assuming the worst
799 _trap_can_recompile = false; // no traps emitted yet
800 _major_progress = true; // start out assuming good things will happen
801 set_has_unsafe_access(false);
802 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
803 set_decompile_count(0);
805 // Compilation level related initialization
806 if (env()->comp_level() == CompLevel_fast_compile) {
807 set_num_loop_opts(Tier1LoopOptsCount);
808 set_do_inlining(Tier1Inline != 0);
809 set_max_inline_size(Tier1MaxInlineSize);
810 set_freq_inline_size(Tier1FreqInlineSize);
811 set_do_scheduling(false);
812 set_do_count_invocations(Tier1CountInvocations);
813 set_do_method_data_update(Tier1UpdateMethodData);
814 } else {
815 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
816 set_num_loop_opts(LoopOptsCount);
817 set_do_inlining(Inline);
818 set_max_inline_size(MaxInlineSize);
819 set_freq_inline_size(FreqInlineSize);
820 set_do_scheduling(OptoScheduling);
821 set_do_count_invocations(false);
822 set_do_method_data_update(false);
823 }
825 if (debug_info()->recording_non_safepoints()) {
826 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
827 (comp_arena(), 8, 0, NULL));
828 set_default_node_notes(Node_Notes::make(this));
829 }
831 // // -- Initialize types before each compile --
832 // // Update cached type information
833 // if( _method && _method->constants() )
834 // Type::update_loaded_types(_method, _method->constants());
836 // Init alias_type map.
837 if (!_do_escape_analysis && aliaslevel == 3)
838 aliaslevel = 2; // No unique types without escape analysis
839 _AliasLevel = aliaslevel;
840 const int grow_ats = 16;
841 _max_alias_types = grow_ats;
842 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
843 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
844 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
845 {
846 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
847 }
848 // Initialize the first few types.
849 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
850 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
851 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
852 _num_alias_types = AliasIdxRaw+1;
853 // Zero out the alias type cache.
854 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
855 // A NULL adr_type hits in the cache right away. Preload the right answer.
856 probe_alias_cache(NULL)->_index = AliasIdxTop;
858 _intrinsics = NULL;
859 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
860 register_library_intrinsics();
861 }
863 //---------------------------init_start----------------------------------------
864 // Install the StartNode on this compile object.
865 void Compile::init_start(StartNode* s) {
866 if (failing())
867 return; // already failing
868 assert(s == start(), "");
869 }
871 StartNode* Compile::start() const {
872 assert(!failing(), "");
873 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
874 Node* start = root()->fast_out(i);
875 if( start->is_Start() )
876 return start->as_Start();
877 }
878 ShouldNotReachHere();
879 return NULL;
880 }
882 //-------------------------------immutable_memory-------------------------------------
883 // Access immutable memory
884 Node* Compile::immutable_memory() {
885 if (_immutable_memory != NULL) {
886 return _immutable_memory;
887 }
888 StartNode* s = start();
889 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
890 Node *p = s->fast_out(i);
891 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
892 _immutable_memory = p;
893 return _immutable_memory;
894 }
895 }
896 ShouldNotReachHere();
897 return NULL;
898 }
900 //----------------------set_cached_top_node------------------------------------
901 // Install the cached top node, and make sure Node::is_top works correctly.
902 void Compile::set_cached_top_node(Node* tn) {
903 if (tn != NULL) verify_top(tn);
904 Node* old_top = _top;
905 _top = tn;
906 // Calling Node::setup_is_top allows the nodes the chance to adjust
907 // their _out arrays.
908 if (_top != NULL) _top->setup_is_top();
909 if (old_top != NULL) old_top->setup_is_top();
910 assert(_top == NULL || top()->is_top(), "");
911 }
913 #ifndef PRODUCT
914 void Compile::verify_top(Node* tn) const {
915 if (tn != NULL) {
916 assert(tn->is_Con(), "top node must be a constant");
917 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
918 assert(tn->in(0) != NULL, "must have live top node");
919 }
920 }
921 #endif
924 ///-------------------Managing Per-Node Debug & Profile Info-------------------
926 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
927 guarantee(arr != NULL, "");
928 int num_blocks = arr->length();
929 if (grow_by < num_blocks) grow_by = num_blocks;
930 int num_notes = grow_by * _node_notes_block_size;
931 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
932 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
933 while (num_notes > 0) {
934 arr->append(notes);
935 notes += _node_notes_block_size;
936 num_notes -= _node_notes_block_size;
937 }
938 assert(num_notes == 0, "exact multiple, please");
939 }
941 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
942 if (source == NULL || dest == NULL) return false;
944 if (dest->is_Con())
945 return false; // Do not push debug info onto constants.
947 #ifdef ASSERT
948 // Leave a bread crumb trail pointing to the original node:
949 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
950 dest->set_debug_orig(source);
951 }
952 #endif
954 if (node_note_array() == NULL)
955 return false; // Not collecting any notes now.
957 // This is a copy onto a pre-existing node, which may already have notes.
958 // If both nodes have notes, do not overwrite any pre-existing notes.
959 Node_Notes* source_notes = node_notes_at(source->_idx);
960 if (source_notes == NULL || source_notes->is_clear()) return false;
961 Node_Notes* dest_notes = node_notes_at(dest->_idx);
962 if (dest_notes == NULL || dest_notes->is_clear()) {
963 return set_node_notes_at(dest->_idx, source_notes);
964 }
966 Node_Notes merged_notes = (*source_notes);
967 // The order of operations here ensures that dest notes will win...
968 merged_notes.update_from(dest_notes);
969 return set_node_notes_at(dest->_idx, &merged_notes);
970 }
973 //--------------------------allow_range_check_smearing-------------------------
974 // Gating condition for coalescing similar range checks.
975 // Sometimes we try 'speculatively' replacing a series of a range checks by a
976 // single covering check that is at least as strong as any of them.
977 // If the optimization succeeds, the simplified (strengthened) range check
978 // will always succeed. If it fails, we will deopt, and then give up
979 // on the optimization.
980 bool Compile::allow_range_check_smearing() const {
981 // If this method has already thrown a range-check,
982 // assume it was because we already tried range smearing
983 // and it failed.
984 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
985 return !already_trapped;
986 }
989 //------------------------------flatten_alias_type-----------------------------
990 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
991 int offset = tj->offset();
992 TypePtr::PTR ptr = tj->ptr();
994 // Process weird unsafe references.
995 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
996 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
997 tj = TypeOopPtr::BOTTOM;
998 ptr = tj->ptr();
999 offset = tj->offset();
1000 }
1002 // Array pointers need some flattening
1003 const TypeAryPtr *ta = tj->isa_aryptr();
1004 if( ta && _AliasLevel >= 2 ) {
1005 // For arrays indexed by constant indices, we flatten the alias
1006 // space to include all of the array body. Only the header, klass
1007 // and array length can be accessed un-aliased.
1008 if( offset != Type::OffsetBot ) {
1009 if( ta->const_oop() ) { // methodDataOop or methodOop
1010 offset = Type::OffsetBot; // Flatten constant access into array body
1011 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id());
1012 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1013 // range is OK as-is.
1014 tj = ta = TypeAryPtr::RANGE;
1015 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1016 tj = TypeInstPtr::KLASS; // all klass loads look alike
1017 ta = TypeAryPtr::RANGE; // generic ignored junk
1018 ptr = TypePtr::BotPTR;
1019 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1020 tj = TypeInstPtr::MARK;
1021 ta = TypeAryPtr::RANGE; // generic ignored junk
1022 ptr = TypePtr::BotPTR;
1023 } else { // Random constant offset into array body
1024 offset = Type::OffsetBot; // Flatten constant access into array body
1025 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id());
1026 }
1027 }
1028 // Arrays of fixed size alias with arrays of unknown size.
1029 if (ta->size() != TypeInt::POS) {
1030 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1031 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset, ta->instance_id());
1032 }
1033 // Arrays of known objects become arrays of unknown objects.
1034 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1035 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1036 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id());
1037 }
1038 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1039 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1040 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id());
1041 }
1042 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1043 // cannot be distinguished by bytecode alone.
1044 if (ta->elem() == TypeInt::BOOL) {
1045 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1046 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1047 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset, ta->instance_id());
1048 }
1049 // During the 2nd round of IterGVN, NotNull castings are removed.
1050 // Make sure the Bottom and NotNull variants alias the same.
1051 // Also, make sure exact and non-exact variants alias the same.
1052 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1053 if (ta->const_oop()) {
1054 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1055 } else {
1056 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1057 }
1058 }
1059 }
1061 // Oop pointers need some flattening
1062 const TypeInstPtr *to = tj->isa_instptr();
1063 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1064 if( ptr == TypePtr::Constant ) {
1065 // No constant oop pointers (such as Strings); they alias with
1066 // unknown strings.
1067 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1068 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1069 // During the 2nd round of IterGVN, NotNull castings are removed.
1070 // Make sure the Bottom and NotNull variants alias the same.
1071 // Also, make sure exact and non-exact variants alias the same.
1072 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset, to->instance_id());
1073 }
1074 // Canonicalize the holder of this field
1075 ciInstanceKlass *k = to->klass()->as_instance_klass();
1076 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1077 // First handle header references such as a LoadKlassNode, even if the
1078 // object's klass is unloaded at compile time (4965979).
1079 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset, to->instance_id());
1080 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1081 to = NULL;
1082 tj = TypeOopPtr::BOTTOM;
1083 offset = tj->offset();
1084 } else {
1085 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1086 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1087 tj = to = TypeInstPtr::make(TypePtr::BotPTR, canonical_holder, false, NULL, offset, to->instance_id());
1088 }
1089 }
1090 }
1092 // Klass pointers to object array klasses need some flattening
1093 const TypeKlassPtr *tk = tj->isa_klassptr();
1094 if( tk ) {
1095 // If we are referencing a field within a Klass, we need
1096 // to assume the worst case of an Object. Both exact and
1097 // inexact types must flatten to the same alias class.
1098 // Since the flattened result for a klass is defined to be
1099 // precisely java.lang.Object, use a constant ptr.
1100 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1102 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1103 TypeKlassPtr::OBJECT->klass(),
1104 offset);
1105 }
1107 ciKlass* klass = tk->klass();
1108 if( klass->is_obj_array_klass() ) {
1109 ciKlass* k = TypeAryPtr::OOPS->klass();
1110 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1111 k = TypeInstPtr::BOTTOM->klass();
1112 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1113 }
1115 // Check for precise loads from the primary supertype array and force them
1116 // to the supertype cache alias index. Check for generic array loads from
1117 // the primary supertype array and also force them to the supertype cache
1118 // alias index. Since the same load can reach both, we need to merge
1119 // these 2 disparate memories into the same alias class. Since the
1120 // primary supertype array is read-only, there's no chance of confusion
1121 // where we bypass an array load and an array store.
1122 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1123 if( offset == Type::OffsetBot ||
1124 off2 < Klass::primary_super_limit()*wordSize ) {
1125 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1126 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1127 }
1128 }
1130 // Flatten all Raw pointers together.
1131 if (tj->base() == Type::RawPtr)
1132 tj = TypeRawPtr::BOTTOM;
1134 if (tj->base() == Type::AnyPtr)
1135 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1137 // Flatten all to bottom for now
1138 switch( _AliasLevel ) {
1139 case 0:
1140 tj = TypePtr::BOTTOM;
1141 break;
1142 case 1: // Flatten to: oop, static, field or array
1143 switch (tj->base()) {
1144 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1145 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1146 case Type::AryPtr: // do not distinguish arrays at all
1147 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1148 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1149 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1150 default: ShouldNotReachHere();
1151 }
1152 break;
1153 case 2: // No collasping at level 2; keep all splits
1154 case 3: // No collasping at level 3; keep all splits
1155 break;
1156 default:
1157 Unimplemented();
1158 }
1160 offset = tj->offset();
1161 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1163 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1164 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1165 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1166 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1167 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1168 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1169 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1170 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1171 assert( tj->ptr() != TypePtr::TopPTR &&
1172 tj->ptr() != TypePtr::AnyNull &&
1173 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1174 // assert( tj->ptr() != TypePtr::Constant ||
1175 // tj->base() == Type::RawPtr ||
1176 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1178 return tj;
1179 }
1181 void Compile::AliasType::Init(int i, const TypePtr* at) {
1182 _index = i;
1183 _adr_type = at;
1184 _field = NULL;
1185 _is_rewritable = true; // default
1186 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1187 if (atoop != NULL && atoop->is_instance()) {
1188 const TypeOopPtr *gt = atoop->cast_to_instance(TypeOopPtr::UNKNOWN_INSTANCE);
1189 _general_index = Compile::current()->get_alias_index(gt);
1190 } else {
1191 _general_index = 0;
1192 }
1193 }
1195 //---------------------------------print_on------------------------------------
1196 #ifndef PRODUCT
1197 void Compile::AliasType::print_on(outputStream* st) {
1198 if (index() < 10)
1199 st->print("@ <%d> ", index());
1200 else st->print("@ <%d>", index());
1201 st->print(is_rewritable() ? " " : " RO");
1202 int offset = adr_type()->offset();
1203 if (offset == Type::OffsetBot)
1204 st->print(" +any");
1205 else st->print(" +%-3d", offset);
1206 st->print(" in ");
1207 adr_type()->dump_on(st);
1208 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1209 if (field() != NULL && tjp) {
1210 if (tjp->klass() != field()->holder() ||
1211 tjp->offset() != field()->offset_in_bytes()) {
1212 st->print(" != ");
1213 field()->print();
1214 st->print(" ***");
1215 }
1216 }
1217 }
1219 void print_alias_types() {
1220 Compile* C = Compile::current();
1221 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1222 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1223 C->alias_type(idx)->print_on(tty);
1224 tty->cr();
1225 }
1226 }
1227 #endif
1230 //----------------------------probe_alias_cache--------------------------------
1231 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1232 intptr_t key = (intptr_t) adr_type;
1233 key ^= key >> logAliasCacheSize;
1234 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1235 }
1238 //-----------------------------grow_alias_types--------------------------------
1239 void Compile::grow_alias_types() {
1240 const int old_ats = _max_alias_types; // how many before?
1241 const int new_ats = old_ats; // how many more?
1242 const int grow_ats = old_ats+new_ats; // how many now?
1243 _max_alias_types = grow_ats;
1244 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1245 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1246 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1247 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1248 }
1251 //--------------------------------find_alias_type------------------------------
1252 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1253 if (_AliasLevel == 0)
1254 return alias_type(AliasIdxBot);
1256 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1257 if (ace->_adr_type == adr_type) {
1258 return alias_type(ace->_index);
1259 }
1261 // Handle special cases.
1262 if (adr_type == NULL) return alias_type(AliasIdxTop);
1263 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1265 // Do it the slow way.
1266 const TypePtr* flat = flatten_alias_type(adr_type);
1268 #ifdef ASSERT
1269 assert(flat == flatten_alias_type(flat), "idempotent");
1270 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1271 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1272 const TypeOopPtr* foop = flat->is_oopptr();
1273 const TypePtr* xoop = foop->cast_to_exactness(!foop->klass_is_exact())->is_ptr();
1274 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1275 }
1276 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1277 #endif
1279 int idx = AliasIdxTop;
1280 for (int i = 0; i < num_alias_types(); i++) {
1281 if (alias_type(i)->adr_type() == flat) {
1282 idx = i;
1283 break;
1284 }
1285 }
1287 if (idx == AliasIdxTop) {
1288 if (no_create) return NULL;
1289 // Grow the array if necessary.
1290 if (_num_alias_types == _max_alias_types) grow_alias_types();
1291 // Add a new alias type.
1292 idx = _num_alias_types++;
1293 _alias_types[idx]->Init(idx, flat);
1294 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1295 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1296 if (flat->isa_instptr()) {
1297 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1298 && flat->is_instptr()->klass() == env()->Class_klass())
1299 alias_type(idx)->set_rewritable(false);
1300 }
1301 if (flat->isa_klassptr()) {
1302 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1303 alias_type(idx)->set_rewritable(false);
1304 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1305 alias_type(idx)->set_rewritable(false);
1306 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1307 alias_type(idx)->set_rewritable(false);
1308 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1309 alias_type(idx)->set_rewritable(false);
1310 }
1311 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1312 // but the base pointer type is not distinctive enough to identify
1313 // references into JavaThread.)
1315 // Check for final instance fields.
1316 const TypeInstPtr* tinst = flat->isa_instptr();
1317 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1318 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1319 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1320 // Set field() and is_rewritable() attributes.
1321 if (field != NULL) alias_type(idx)->set_field(field);
1322 }
1323 const TypeKlassPtr* tklass = flat->isa_klassptr();
1324 // Check for final static fields.
1325 if (tklass && tklass->klass()->is_instance_klass()) {
1326 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1327 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1328 // Set field() and is_rewritable() attributes.
1329 if (field != NULL) alias_type(idx)->set_field(field);
1330 }
1331 }
1333 // Fill the cache for next time.
1334 ace->_adr_type = adr_type;
1335 ace->_index = idx;
1336 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1338 // Might as well try to fill the cache for the flattened version, too.
1339 AliasCacheEntry* face = probe_alias_cache(flat);
1340 if (face->_adr_type == NULL) {
1341 face->_adr_type = flat;
1342 face->_index = idx;
1343 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1344 }
1346 return alias_type(idx);
1347 }
1350 Compile::AliasType* Compile::alias_type(ciField* field) {
1351 const TypeOopPtr* t;
1352 if (field->is_static())
1353 t = TypeKlassPtr::make(field->holder());
1354 else
1355 t = TypeOopPtr::make_from_klass_raw(field->holder());
1356 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1357 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1358 return atp;
1359 }
1362 //------------------------------have_alias_type--------------------------------
1363 bool Compile::have_alias_type(const TypePtr* adr_type) {
1364 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1365 if (ace->_adr_type == adr_type) {
1366 return true;
1367 }
1369 // Handle special cases.
1370 if (adr_type == NULL) return true;
1371 if (adr_type == TypePtr::BOTTOM) return true;
1373 return find_alias_type(adr_type, true) != NULL;
1374 }
1376 //-----------------------------must_alias--------------------------------------
1377 // True if all values of the given address type are in the given alias category.
1378 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1379 if (alias_idx == AliasIdxBot) return true; // the universal category
1380 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1381 if (alias_idx == AliasIdxTop) return false; // the empty category
1382 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1384 // the only remaining possible overlap is identity
1385 int adr_idx = get_alias_index(adr_type);
1386 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1387 assert(adr_idx == alias_idx ||
1388 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1389 && adr_type != TypeOopPtr::BOTTOM),
1390 "should not be testing for overlap with an unsafe pointer");
1391 return adr_idx == alias_idx;
1392 }
1394 //------------------------------can_alias--------------------------------------
1395 // True if any values of the given address type are in the given alias category.
1396 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1397 if (alias_idx == AliasIdxTop) return false; // the empty category
1398 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1399 if (alias_idx == AliasIdxBot) return true; // the universal category
1400 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1402 // the only remaining possible overlap is identity
1403 int adr_idx = get_alias_index(adr_type);
1404 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1405 return adr_idx == alias_idx;
1406 }
1410 //---------------------------pop_warm_call-------------------------------------
1411 WarmCallInfo* Compile::pop_warm_call() {
1412 WarmCallInfo* wci = _warm_calls;
1413 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1414 return wci;
1415 }
1417 //----------------------------Inline_Warm--------------------------------------
1418 int Compile::Inline_Warm() {
1419 // If there is room, try to inline some more warm call sites.
1420 // %%% Do a graph index compaction pass when we think we're out of space?
1421 if (!InlineWarmCalls) return 0;
1423 int calls_made_hot = 0;
1424 int room_to_grow = NodeCountInliningCutoff - unique();
1425 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1426 int amount_grown = 0;
1427 WarmCallInfo* call;
1428 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1429 int est_size = (int)call->size();
1430 if (est_size > (room_to_grow - amount_grown)) {
1431 // This one won't fit anyway. Get rid of it.
1432 call->make_cold();
1433 continue;
1434 }
1435 call->make_hot();
1436 calls_made_hot++;
1437 amount_grown += est_size;
1438 amount_to_grow -= est_size;
1439 }
1441 if (calls_made_hot > 0) set_major_progress();
1442 return calls_made_hot;
1443 }
1446 //----------------------------Finish_Warm--------------------------------------
1447 void Compile::Finish_Warm() {
1448 if (!InlineWarmCalls) return;
1449 if (failing()) return;
1450 if (warm_calls() == NULL) return;
1452 // Clean up loose ends, if we are out of space for inlining.
1453 WarmCallInfo* call;
1454 while ((call = pop_warm_call()) != NULL) {
1455 call->make_cold();
1456 }
1457 }
1460 //------------------------------Optimize---------------------------------------
1461 // Given a graph, optimize it.
1462 void Compile::Optimize() {
1463 TracePhase t1("optimizer", &_t_optimizer, true);
1465 #ifndef PRODUCT
1466 if (env()->break_at_compile()) {
1467 BREAKPOINT;
1468 }
1470 #endif
1472 ResourceMark rm;
1473 int loop_opts_cnt;
1475 NOT_PRODUCT( verify_graph_edges(); )
1477 print_method("Start");
1479 {
1480 // Iterative Global Value Numbering, including ideal transforms
1481 // Initialize IterGVN with types and values from parse-time GVN
1482 PhaseIterGVN igvn(initial_gvn());
1483 {
1484 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1485 igvn.optimize();
1486 }
1488 print_method("Iter GVN 1", 2);
1490 if (failing()) return;
1492 // get rid of the connection graph since it's information is not
1493 // updated by optimizations
1494 _congraph = NULL;
1497 // Loop transforms on the ideal graph. Range Check Elimination,
1498 // peeling, unrolling, etc.
1500 // Set loop opts counter
1501 loop_opts_cnt = num_loop_opts();
1502 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1503 {
1504 TracePhase t2("idealLoop", &_t_idealLoop, true);
1505 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1506 loop_opts_cnt--;
1507 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1508 if (failing()) return;
1509 }
1510 // Loop opts pass if partial peeling occurred in previous pass
1511 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1512 TracePhase t3("idealLoop", &_t_idealLoop, true);
1513 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1514 loop_opts_cnt--;
1515 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1516 if (failing()) return;
1517 }
1518 // Loop opts pass for loop-unrolling before CCP
1519 if(major_progress() && (loop_opts_cnt > 0)) {
1520 TracePhase t4("idealLoop", &_t_idealLoop, true);
1521 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1522 loop_opts_cnt--;
1523 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1524 }
1525 }
1526 if (failing()) return;
1528 // Conditional Constant Propagation;
1529 PhaseCCP ccp( &igvn );
1530 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1531 {
1532 TracePhase t2("ccp", &_t_ccp, true);
1533 ccp.do_transform();
1534 }
1535 print_method("PhaseCPP 1", 2);
1537 assert( true, "Break here to ccp.dump_old2new_map()");
1539 // Iterative Global Value Numbering, including ideal transforms
1540 {
1541 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1542 igvn = ccp;
1543 igvn.optimize();
1544 }
1546 print_method("Iter GVN 2", 2);
1548 if (failing()) return;
1550 // Loop transforms on the ideal graph. Range Check Elimination,
1551 // peeling, unrolling, etc.
1552 if(loop_opts_cnt > 0) {
1553 debug_only( int cnt = 0; );
1554 while(major_progress() && (loop_opts_cnt > 0)) {
1555 TracePhase t2("idealLoop", &_t_idealLoop, true);
1556 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1557 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1558 loop_opts_cnt--;
1559 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1560 if (failing()) return;
1561 }
1562 }
1563 {
1564 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1565 PhaseMacroExpand mex(igvn);
1566 if (mex.expand_macro_nodes()) {
1567 assert(failing(), "must bail out w/ explicit message");
1568 return;
1569 }
1570 }
1572 } // (End scope of igvn; run destructor if necessary for asserts.)
1574 // A method with only infinite loops has no edges entering loops from root
1575 {
1576 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1577 if (final_graph_reshaping()) {
1578 assert(failing(), "must bail out w/ explicit message");
1579 return;
1580 }
1581 }
1583 print_method("Optimize finished", 2);
1584 }
1587 //------------------------------Code_Gen---------------------------------------
1588 // Given a graph, generate code for it
1589 void Compile::Code_Gen() {
1590 if (failing()) return;
1592 // Perform instruction selection. You might think we could reclaim Matcher
1593 // memory PDQ, but actually the Matcher is used in generating spill code.
1594 // Internals of the Matcher (including some VectorSets) must remain live
1595 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1596 // set a bit in reclaimed memory.
1598 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1599 // nodes. Mapping is only valid at the root of each matched subtree.
1600 NOT_PRODUCT( verify_graph_edges(); )
1602 Node_List proj_list;
1603 Matcher m(proj_list);
1604 _matcher = &m;
1605 {
1606 TracePhase t2("matcher", &_t_matcher, true);
1607 m.match();
1608 }
1609 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1610 // nodes. Mapping is only valid at the root of each matched subtree.
1611 NOT_PRODUCT( verify_graph_edges(); )
1613 // If you have too many nodes, or if matching has failed, bail out
1614 check_node_count(0, "out of nodes matching instructions");
1615 if (failing()) return;
1617 // Build a proper-looking CFG
1618 PhaseCFG cfg(node_arena(), root(), m);
1619 _cfg = &cfg;
1620 {
1621 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1622 cfg.Dominators();
1623 if (failing()) return;
1625 NOT_PRODUCT( verify_graph_edges(); )
1627 cfg.Estimate_Block_Frequency();
1628 cfg.GlobalCodeMotion(m,unique(),proj_list);
1630 print_method("Global code motion", 2);
1632 if (failing()) return;
1633 NOT_PRODUCT( verify_graph_edges(); )
1635 debug_only( cfg.verify(); )
1636 }
1637 NOT_PRODUCT( verify_graph_edges(); )
1639 PhaseChaitin regalloc(unique(),cfg,m);
1640 _regalloc = ®alloc;
1641 {
1642 TracePhase t2("regalloc", &_t_registerAllocation, true);
1643 // Perform any platform dependent preallocation actions. This is used,
1644 // for example, to avoid taking an implicit null pointer exception
1645 // using the frame pointer on win95.
1646 _regalloc->pd_preallocate_hook();
1648 // Perform register allocation. After Chaitin, use-def chains are
1649 // no longer accurate (at spill code) and so must be ignored.
1650 // Node->LRG->reg mappings are still accurate.
1651 _regalloc->Register_Allocate();
1653 // Bail out if the allocator builds too many nodes
1654 if (failing()) return;
1655 }
1657 // Prior to register allocation we kept empty basic blocks in case the
1658 // the allocator needed a place to spill. After register allocation we
1659 // are not adding any new instructions. If any basic block is empty, we
1660 // can now safely remove it.
1661 {
1662 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); )
1663 cfg.RemoveEmpty();
1664 }
1666 // Perform any platform dependent postallocation verifications.
1667 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1669 // Apply peephole optimizations
1670 if( OptoPeephole ) {
1671 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1672 PhasePeephole peep( _regalloc, cfg);
1673 peep.do_transform();
1674 }
1676 // Convert Nodes to instruction bits in a buffer
1677 {
1678 // %%%% workspace merge brought two timers together for one job
1679 TracePhase t2a("output", &_t_output, true);
1680 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1681 Output();
1682 }
1684 print_method("End");
1686 // He's dead, Jim.
1687 _cfg = (PhaseCFG*)0xdeadbeef;
1688 _regalloc = (PhaseChaitin*)0xdeadbeef;
1689 }
1692 //------------------------------dump_asm---------------------------------------
1693 // Dump formatted assembly
1694 #ifndef PRODUCT
1695 void Compile::dump_asm(int *pcs, uint pc_limit) {
1696 bool cut_short = false;
1697 tty->print_cr("#");
1698 tty->print("# "); _tf->dump(); tty->cr();
1699 tty->print_cr("#");
1701 // For all blocks
1702 int pc = 0x0; // Program counter
1703 char starts_bundle = ' ';
1704 _regalloc->dump_frame();
1706 Node *n = NULL;
1707 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1708 if (VMThread::should_terminate()) { cut_short = true; break; }
1709 Block *b = _cfg->_blocks[i];
1710 if (b->is_connector() && !Verbose) continue;
1711 n = b->_nodes[0];
1712 if (pcs && n->_idx < pc_limit)
1713 tty->print("%3.3x ", pcs[n->_idx]);
1714 else
1715 tty->print(" ");
1716 b->dump_head( &_cfg->_bbs );
1717 if (b->is_connector()) {
1718 tty->print_cr(" # Empty connector block");
1719 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1720 tty->print_cr(" # Block is sole successor of call");
1721 }
1723 // For all instructions
1724 Node *delay = NULL;
1725 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1726 if (VMThread::should_terminate()) { cut_short = true; break; }
1727 n = b->_nodes[j];
1728 if (valid_bundle_info(n)) {
1729 Bundle *bundle = node_bundling(n);
1730 if (bundle->used_in_unconditional_delay()) {
1731 delay = n;
1732 continue;
1733 }
1734 if (bundle->starts_bundle())
1735 starts_bundle = '+';
1736 }
1738 if (WizardMode) n->dump();
1740 if( !n->is_Region() && // Dont print in the Assembly
1741 !n->is_Phi() && // a few noisely useless nodes
1742 !n->is_Proj() &&
1743 !n->is_MachTemp() &&
1744 !n->is_Catch() && // Would be nice to print exception table targets
1745 !n->is_MergeMem() && // Not very interesting
1746 !n->is_top() && // Debug info table constants
1747 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1748 ) {
1749 if (pcs && n->_idx < pc_limit)
1750 tty->print("%3.3x", pcs[n->_idx]);
1751 else
1752 tty->print(" ");
1753 tty->print(" %c ", starts_bundle);
1754 starts_bundle = ' ';
1755 tty->print("\t");
1756 n->format(_regalloc, tty);
1757 tty->cr();
1758 }
1760 // If we have an instruction with a delay slot, and have seen a delay,
1761 // then back up and print it
1762 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1763 assert(delay != NULL, "no unconditional delay instruction");
1764 if (WizardMode) delay->dump();
1766 if (node_bundling(delay)->starts_bundle())
1767 starts_bundle = '+';
1768 if (pcs && n->_idx < pc_limit)
1769 tty->print("%3.3x", pcs[n->_idx]);
1770 else
1771 tty->print(" ");
1772 tty->print(" %c ", starts_bundle);
1773 starts_bundle = ' ';
1774 tty->print("\t");
1775 delay->format(_regalloc, tty);
1776 tty->print_cr("");
1777 delay = NULL;
1778 }
1780 // Dump the exception table as well
1781 if( n->is_Catch() && (Verbose || WizardMode) ) {
1782 // Print the exception table for this offset
1783 _handler_table.print_subtable_for(pc);
1784 }
1785 }
1787 if (pcs && n->_idx < pc_limit)
1788 tty->print_cr("%3.3x", pcs[n->_idx]);
1789 else
1790 tty->print_cr("");
1792 assert(cut_short || delay == NULL, "no unconditional delay branch");
1794 } // End of per-block dump
1795 tty->print_cr("");
1797 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1798 }
1799 #endif
1801 //------------------------------Final_Reshape_Counts---------------------------
1802 // This class defines counters to help identify when a method
1803 // may/must be executed using hardware with only 24-bit precision.
1804 struct Final_Reshape_Counts : public StackObj {
1805 int _call_count; // count non-inlined 'common' calls
1806 int _float_count; // count float ops requiring 24-bit precision
1807 int _double_count; // count double ops requiring more precision
1808 int _java_call_count; // count non-inlined 'java' calls
1809 VectorSet _visited; // Visitation flags
1810 Node_List _tests; // Set of IfNodes & PCTableNodes
1812 Final_Reshape_Counts() :
1813 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0),
1814 _visited( Thread::current()->resource_area() ) { }
1816 void inc_call_count () { _call_count ++; }
1817 void inc_float_count () { _float_count ++; }
1818 void inc_double_count() { _double_count++; }
1819 void inc_java_call_count() { _java_call_count++; }
1821 int get_call_count () const { return _call_count ; }
1822 int get_float_count () const { return _float_count ; }
1823 int get_double_count() const { return _double_count; }
1824 int get_java_call_count() const { return _java_call_count; }
1825 };
1827 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1828 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1829 // Make sure the offset goes inside the instance layout.
1830 return k->contains_field_offset(tp->offset());
1831 // Note that OffsetBot and OffsetTop are very negative.
1832 }
1834 //------------------------------final_graph_reshaping_impl----------------------
1835 // Implement items 1-5 from final_graph_reshaping below.
1836 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
1838 uint nop = n->Opcode();
1840 // Check for 2-input instruction with "last use" on right input.
1841 // Swap to left input. Implements item (2).
1842 if( n->req() == 3 && // two-input instruction
1843 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1844 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1845 n->in(2)->outcnt() == 1 &&// right use IS a last use
1846 !n->in(2)->is_Con() ) { // right use is not a constant
1847 // Check for commutative opcode
1848 switch( nop ) {
1849 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1850 case Op_MaxI: case Op_MinI:
1851 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1852 case Op_AndL: case Op_XorL: case Op_OrL:
1853 case Op_AndI: case Op_XorI: case Op_OrI: {
1854 // Move "last use" input to left by swapping inputs
1855 n->swap_edges(1, 2);
1856 break;
1857 }
1858 default:
1859 break;
1860 }
1861 }
1863 // Count FPU ops and common calls, implements item (3)
1864 switch( nop ) {
1865 // Count all float operations that may use FPU
1866 case Op_AddF:
1867 case Op_SubF:
1868 case Op_MulF:
1869 case Op_DivF:
1870 case Op_NegF:
1871 case Op_ModF:
1872 case Op_ConvI2F:
1873 case Op_ConF:
1874 case Op_CmpF:
1875 case Op_CmpF3:
1876 // case Op_ConvL2F: // longs are split into 32-bit halves
1877 fpu.inc_float_count();
1878 break;
1880 case Op_ConvF2D:
1881 case Op_ConvD2F:
1882 fpu.inc_float_count();
1883 fpu.inc_double_count();
1884 break;
1886 // Count all double operations that may use FPU
1887 case Op_AddD:
1888 case Op_SubD:
1889 case Op_MulD:
1890 case Op_DivD:
1891 case Op_NegD:
1892 case Op_ModD:
1893 case Op_ConvI2D:
1894 case Op_ConvD2I:
1895 // case Op_ConvL2D: // handled by leaf call
1896 // case Op_ConvD2L: // handled by leaf call
1897 case Op_ConD:
1898 case Op_CmpD:
1899 case Op_CmpD3:
1900 fpu.inc_double_count();
1901 break;
1902 case Op_Opaque1: // Remove Opaque Nodes before matching
1903 case Op_Opaque2: // Remove Opaque Nodes before matching
1904 n->replace_by(n->in(1));
1905 break;
1906 case Op_CallStaticJava:
1907 case Op_CallJava:
1908 case Op_CallDynamicJava:
1909 fpu.inc_java_call_count(); // Count java call site;
1910 case Op_CallRuntime:
1911 case Op_CallLeaf:
1912 case Op_CallLeafNoFP: {
1913 assert( n->is_Call(), "" );
1914 CallNode *call = n->as_Call();
1915 // Count call sites where the FP mode bit would have to be flipped.
1916 // Do not count uncommon runtime calls:
1917 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
1918 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
1919 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
1920 fpu.inc_call_count(); // Count the call site
1921 } else { // See if uncommon argument is shared
1922 Node *n = call->in(TypeFunc::Parms);
1923 int nop = n->Opcode();
1924 // Clone shared simple arguments to uncommon calls, item (1).
1925 if( n->outcnt() > 1 &&
1926 !n->is_Proj() &&
1927 nop != Op_CreateEx &&
1928 nop != Op_CheckCastPP &&
1929 !n->is_Mem() ) {
1930 Node *x = n->clone();
1931 call->set_req( TypeFunc::Parms, x );
1932 }
1933 }
1934 break;
1935 }
1937 case Op_StoreD:
1938 case Op_LoadD:
1939 case Op_LoadD_unaligned:
1940 fpu.inc_double_count();
1941 goto handle_mem;
1942 case Op_StoreF:
1943 case Op_LoadF:
1944 fpu.inc_float_count();
1945 goto handle_mem;
1947 case Op_StoreB:
1948 case Op_StoreC:
1949 case Op_StoreCM:
1950 case Op_StorePConditional:
1951 case Op_StoreI:
1952 case Op_StoreL:
1953 case Op_StoreLConditional:
1954 case Op_CompareAndSwapI:
1955 case Op_CompareAndSwapL:
1956 case Op_CompareAndSwapP:
1957 case Op_CompareAndSwapN:
1958 case Op_StoreP:
1959 case Op_StoreN:
1960 case Op_LoadB:
1961 case Op_LoadC:
1962 case Op_LoadI:
1963 case Op_LoadKlass:
1964 case Op_LoadL:
1965 case Op_LoadL_unaligned:
1966 case Op_LoadPLocked:
1967 case Op_LoadLLocked:
1968 case Op_LoadP:
1969 case Op_LoadN:
1970 case Op_LoadRange:
1971 case Op_LoadS: {
1972 handle_mem:
1973 #ifdef ASSERT
1974 if( VerifyOptoOopOffsets ) {
1975 assert( n->is_Mem(), "" );
1976 MemNode *mem = (MemNode*)n;
1977 // Check to see if address types have grounded out somehow.
1978 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1979 assert( !tp || oop_offset_is_sane(tp), "" );
1980 }
1981 #endif
1982 break;
1983 }
1985 case Op_AddP: { // Assert sane base pointers
1986 const Node *addp = n->in(AddPNode::Address);
1987 assert( !addp->is_AddP() ||
1988 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
1989 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
1990 "Base pointers must match" );
1991 break;
1992 }
1994 case Op_ModI:
1995 if (UseDivMod) {
1996 // Check if a%b and a/b both exist
1997 Node* d = n->find_similar(Op_DivI);
1998 if (d) {
1999 // Replace them with a fused divmod if supported
2000 Compile* C = Compile::current();
2001 if (Matcher::has_match_rule(Op_DivModI)) {
2002 DivModINode* divmod = DivModINode::make(C, n);
2003 d->replace_by(divmod->div_proj());
2004 n->replace_by(divmod->mod_proj());
2005 } else {
2006 // replace a%b with a-((a/b)*b)
2007 Node* mult = new (C, 3) MulINode(d, d->in(2));
2008 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2009 n->replace_by( sub );
2010 }
2011 }
2012 }
2013 break;
2015 case Op_ModL:
2016 if (UseDivMod) {
2017 // Check if a%b and a/b both exist
2018 Node* d = n->find_similar(Op_DivL);
2019 if (d) {
2020 // Replace them with a fused divmod if supported
2021 Compile* C = Compile::current();
2022 if (Matcher::has_match_rule(Op_DivModL)) {
2023 DivModLNode* divmod = DivModLNode::make(C, n);
2024 d->replace_by(divmod->div_proj());
2025 n->replace_by(divmod->mod_proj());
2026 } else {
2027 // replace a%b with a-((a/b)*b)
2028 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2029 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2030 n->replace_by( sub );
2031 }
2032 }
2033 }
2034 break;
2036 case Op_Load16B:
2037 case Op_Load8B:
2038 case Op_Load4B:
2039 case Op_Load8S:
2040 case Op_Load4S:
2041 case Op_Load2S:
2042 case Op_Load8C:
2043 case Op_Load4C:
2044 case Op_Load2C:
2045 case Op_Load4I:
2046 case Op_Load2I:
2047 case Op_Load2L:
2048 case Op_Load4F:
2049 case Op_Load2F:
2050 case Op_Load2D:
2051 case Op_Store16B:
2052 case Op_Store8B:
2053 case Op_Store4B:
2054 case Op_Store8C:
2055 case Op_Store4C:
2056 case Op_Store2C:
2057 case Op_Store4I:
2058 case Op_Store2I:
2059 case Op_Store2L:
2060 case Op_Store4F:
2061 case Op_Store2F:
2062 case Op_Store2D:
2063 break;
2065 case Op_PackB:
2066 case Op_PackS:
2067 case Op_PackC:
2068 case Op_PackI:
2069 case Op_PackF:
2070 case Op_PackL:
2071 case Op_PackD:
2072 if (n->req()-1 > 2) {
2073 // Replace many operand PackNodes with a binary tree for matching
2074 PackNode* p = (PackNode*) n;
2075 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2076 n->replace_by(btp);
2077 }
2078 break;
2079 default:
2080 assert( !n->is_Call(), "" );
2081 assert( !n->is_Mem(), "" );
2082 break;
2083 }
2085 // Collect CFG split points
2086 if (n->is_MultiBranch())
2087 fpu._tests.push(n);
2088 }
2090 //------------------------------final_graph_reshaping_walk---------------------
2091 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2092 // requires that the walk visits a node's inputs before visiting the node.
2093 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) {
2094 fpu._visited.set(root->_idx); // first, mark node as visited
2095 uint cnt = root->req();
2096 Node *n = root;
2097 uint i = 0;
2098 while (true) {
2099 if (i < cnt) {
2100 // Place all non-visited non-null inputs onto stack
2101 Node* m = n->in(i);
2102 ++i;
2103 if (m != NULL && !fpu._visited.test_set(m->_idx)) {
2104 cnt = m->req();
2105 nstack.push(n, i); // put on stack parent and next input's index
2106 n = m;
2107 i = 0;
2108 }
2109 } else {
2110 // Now do post-visit work
2111 final_graph_reshaping_impl( n, fpu );
2112 if (nstack.is_empty())
2113 break; // finished
2114 n = nstack.node(); // Get node from stack
2115 cnt = n->req();
2116 i = nstack.index();
2117 nstack.pop(); // Shift to the next node on stack
2118 }
2119 }
2120 }
2122 //------------------------------final_graph_reshaping--------------------------
2123 // Final Graph Reshaping.
2124 //
2125 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2126 // and not commoned up and forced early. Must come after regular
2127 // optimizations to avoid GVN undoing the cloning. Clone constant
2128 // inputs to Loop Phis; these will be split by the allocator anyways.
2129 // Remove Opaque nodes.
2130 // (2) Move last-uses by commutative operations to the left input to encourage
2131 // Intel update-in-place two-address operations and better register usage
2132 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2133 // calls canonicalizing them back.
2134 // (3) Count the number of double-precision FP ops, single-precision FP ops
2135 // and call sites. On Intel, we can get correct rounding either by
2136 // forcing singles to memory (requires extra stores and loads after each
2137 // FP bytecode) or we can set a rounding mode bit (requires setting and
2138 // clearing the mode bit around call sites). The mode bit is only used
2139 // if the relative frequency of single FP ops to calls is low enough.
2140 // This is a key transform for SPEC mpeg_audio.
2141 // (4) Detect infinite loops; blobs of code reachable from above but not
2142 // below. Several of the Code_Gen algorithms fail on such code shapes,
2143 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2144 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2145 // Detection is by looking for IfNodes where only 1 projection is
2146 // reachable from below or CatchNodes missing some targets.
2147 // (5) Assert for insane oop offsets in debug mode.
2149 bool Compile::final_graph_reshaping() {
2150 // an infinite loop may have been eliminated by the optimizer,
2151 // in which case the graph will be empty.
2152 if (root()->req() == 1) {
2153 record_method_not_compilable("trivial infinite loop");
2154 return true;
2155 }
2157 Final_Reshape_Counts fpu;
2159 // Visit everybody reachable!
2160 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2161 Node_Stack nstack(unique() >> 1);
2162 final_graph_reshaping_walk(nstack, root(), fpu);
2164 // Check for unreachable (from below) code (i.e., infinite loops).
2165 for( uint i = 0; i < fpu._tests.size(); i++ ) {
2166 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch();
2167 // Get number of CFG targets.
2168 // Note that PCTables include exception targets after calls.
2169 uint required_outcnt = n->required_outcnt();
2170 if (n->outcnt() != required_outcnt) {
2171 // Check for a few special cases. Rethrow Nodes never take the
2172 // 'fall-thru' path, so expected kids is 1 less.
2173 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2174 if (n->in(0)->in(0)->is_Call()) {
2175 CallNode *call = n->in(0)->in(0)->as_Call();
2176 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2177 required_outcnt--; // Rethrow always has 1 less kid
2178 } else if (call->req() > TypeFunc::Parms &&
2179 call->is_CallDynamicJava()) {
2180 // Check for null receiver. In such case, the optimizer has
2181 // detected that the virtual call will always result in a null
2182 // pointer exception. The fall-through projection of this CatchNode
2183 // will not be populated.
2184 Node *arg0 = call->in(TypeFunc::Parms);
2185 if (arg0->is_Type() &&
2186 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2187 required_outcnt--;
2188 }
2189 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2190 call->req() > TypeFunc::Parms+1 &&
2191 call->is_CallStaticJava()) {
2192 // Check for negative array length. In such case, the optimizer has
2193 // detected that the allocation attempt will always result in an
2194 // exception. There is no fall-through projection of this CatchNode .
2195 Node *arg1 = call->in(TypeFunc::Parms+1);
2196 if (arg1->is_Type() &&
2197 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2198 required_outcnt--;
2199 }
2200 }
2201 }
2202 }
2203 // Recheck with a better notion of 'required_outcnt'
2204 if (n->outcnt() != required_outcnt) {
2205 record_method_not_compilable("malformed control flow");
2206 return true; // Not all targets reachable!
2207 }
2208 }
2209 // Check that I actually visited all kids. Unreached kids
2210 // must be infinite loops.
2211 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2212 if (!fpu._visited.test(n->fast_out(j)->_idx)) {
2213 record_method_not_compilable("infinite loop");
2214 return true; // Found unvisited kid; must be unreach
2215 }
2216 }
2218 // If original bytecodes contained a mixture of floats and doubles
2219 // check if the optimizer has made it homogenous, item (3).
2220 if( Use24BitFPMode && Use24BitFP &&
2221 fpu.get_float_count() > 32 &&
2222 fpu.get_double_count() == 0 &&
2223 (10 * fpu.get_call_count() < fpu.get_float_count()) ) {
2224 set_24_bit_selection_and_mode( false, true );
2225 }
2227 set_has_java_calls(fpu.get_java_call_count() > 0);
2229 // No infinite loops, no reason to bail out.
2230 return false;
2231 }
2233 //-----------------------------too_many_traps----------------------------------
2234 // Report if there are too many traps at the current method and bci.
2235 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2236 bool Compile::too_many_traps(ciMethod* method,
2237 int bci,
2238 Deoptimization::DeoptReason reason) {
2239 ciMethodData* md = method->method_data();
2240 if (md->is_empty()) {
2241 // Assume the trap has not occurred, or that it occurred only
2242 // because of a transient condition during start-up in the interpreter.
2243 return false;
2244 }
2245 if (md->has_trap_at(bci, reason) != 0) {
2246 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2247 // Also, if there are multiple reasons, or if there is no per-BCI record,
2248 // assume the worst.
2249 if (log())
2250 log()->elem("observe trap='%s' count='%d'",
2251 Deoptimization::trap_reason_name(reason),
2252 md->trap_count(reason));
2253 return true;
2254 } else {
2255 // Ignore method/bci and see if there have been too many globally.
2256 return too_many_traps(reason, md);
2257 }
2258 }
2260 // Less-accurate variant which does not require a method and bci.
2261 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2262 ciMethodData* logmd) {
2263 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2264 // Too many traps globally.
2265 // Note that we use cumulative trap_count, not just md->trap_count.
2266 if (log()) {
2267 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2268 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2269 Deoptimization::trap_reason_name(reason),
2270 mcount, trap_count(reason));
2271 }
2272 return true;
2273 } else {
2274 // The coast is clear.
2275 return false;
2276 }
2277 }
2279 //--------------------------too_many_recompiles--------------------------------
2280 // Report if there are too many recompiles at the current method and bci.
2281 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2282 // Is not eager to return true, since this will cause the compiler to use
2283 // Action_none for a trap point, to avoid too many recompilations.
2284 bool Compile::too_many_recompiles(ciMethod* method,
2285 int bci,
2286 Deoptimization::DeoptReason reason) {
2287 ciMethodData* md = method->method_data();
2288 if (md->is_empty()) {
2289 // Assume the trap has not occurred, or that it occurred only
2290 // because of a transient condition during start-up in the interpreter.
2291 return false;
2292 }
2293 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2294 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2295 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2296 Deoptimization::DeoptReason per_bc_reason
2297 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2298 if ((per_bc_reason == Deoptimization::Reason_none
2299 || md->has_trap_at(bci, reason) != 0)
2300 // The trap frequency measure we care about is the recompile count:
2301 && md->trap_recompiled_at(bci)
2302 && md->overflow_recompile_count() >= bc_cutoff) {
2303 // Do not emit a trap here if it has already caused recompilations.
2304 // Also, if there are multiple reasons, or if there is no per-BCI record,
2305 // assume the worst.
2306 if (log())
2307 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2308 Deoptimization::trap_reason_name(reason),
2309 md->trap_count(reason),
2310 md->overflow_recompile_count());
2311 return true;
2312 } else if (trap_count(reason) != 0
2313 && decompile_count() >= m_cutoff) {
2314 // Too many recompiles globally, and we have seen this sort of trap.
2315 // Use cumulative decompile_count, not just md->decompile_count.
2316 if (log())
2317 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2318 Deoptimization::trap_reason_name(reason),
2319 md->trap_count(reason), trap_count(reason),
2320 md->decompile_count(), decompile_count());
2321 return true;
2322 } else {
2323 // The coast is clear.
2324 return false;
2325 }
2326 }
2329 #ifndef PRODUCT
2330 //------------------------------verify_graph_edges---------------------------
2331 // Walk the Graph and verify that there is a one-to-one correspondence
2332 // between Use-Def edges and Def-Use edges in the graph.
2333 void Compile::verify_graph_edges(bool no_dead_code) {
2334 if (VerifyGraphEdges) {
2335 ResourceArea *area = Thread::current()->resource_area();
2336 Unique_Node_List visited(area);
2337 // Call recursive graph walk to check edges
2338 _root->verify_edges(visited);
2339 if (no_dead_code) {
2340 // Now make sure that no visited node is used by an unvisited node.
2341 bool dead_nodes = 0;
2342 Unique_Node_List checked(area);
2343 while (visited.size() > 0) {
2344 Node* n = visited.pop();
2345 checked.push(n);
2346 for (uint i = 0; i < n->outcnt(); i++) {
2347 Node* use = n->raw_out(i);
2348 if (checked.member(use)) continue; // already checked
2349 if (visited.member(use)) continue; // already in the graph
2350 if (use->is_Con()) continue; // a dead ConNode is OK
2351 // At this point, we have found a dead node which is DU-reachable.
2352 if (dead_nodes++ == 0)
2353 tty->print_cr("*** Dead nodes reachable via DU edges:");
2354 use->dump(2);
2355 tty->print_cr("---");
2356 checked.push(use); // No repeats; pretend it is now checked.
2357 }
2358 }
2359 assert(dead_nodes == 0, "using nodes must be reachable from root");
2360 }
2361 }
2362 }
2363 #endif
2365 // The Compile object keeps track of failure reasons separately from the ciEnv.
2366 // This is required because there is not quite a 1-1 relation between the
2367 // ciEnv and its compilation task and the Compile object. Note that one
2368 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2369 // to backtrack and retry without subsuming loads. Other than this backtracking
2370 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2371 // by the logic in C2Compiler.
2372 void Compile::record_failure(const char* reason) {
2373 if (log() != NULL) {
2374 log()->elem("failure reason='%s' phase='compile'", reason);
2375 }
2376 if (_failure_reason == NULL) {
2377 // Record the first failure reason.
2378 _failure_reason = reason;
2379 }
2380 _root = NULL; // flush the graph, too
2381 }
2383 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2384 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2385 {
2386 if (dolog) {
2387 C = Compile::current();
2388 _log = C->log();
2389 } else {
2390 C = NULL;
2391 _log = NULL;
2392 }
2393 if (_log != NULL) {
2394 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2395 _log->stamp();
2396 _log->end_head();
2397 }
2398 }
2400 Compile::TracePhase::~TracePhase() {
2401 if (_log != NULL) {
2402 _log->done("phase nodes='%d'", C->unique());
2403 }
2404 }