Wed, 16 Jul 2008 16:04:39 -0700
6723160: Nightly failure: Error: meet not symmetric
Summary: Add missing _instance_id settings and other EA fixes.
Reviewed-by: rasbold
1 /*
2 * Copyright 1997-2008 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 _compile->end_method();
317 if (_compile->scratch_buffer_blob() != NULL)
318 BufferBlob::free(_compile->scratch_buffer_blob());
319 _compile->env()->set_compiler_data(NULL);
320 }
323 //----------------------------print_compile_messages---------------------------
324 void Compile::print_compile_messages() {
325 #ifndef PRODUCT
326 // Check if recompiling
327 if (_subsume_loads == false && PrintOpto) {
328 // Recompiling without allowing machine instructions to subsume loads
329 tty->print_cr("*********************************************************");
330 tty->print_cr("** Bailout: Recompile without subsuming loads **");
331 tty->print_cr("*********************************************************");
332 }
333 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
334 // Recompiling without escape analysis
335 tty->print_cr("*********************************************************");
336 tty->print_cr("** Bailout: Recompile without escape analysis **");
337 tty->print_cr("*********************************************************");
338 }
339 if (env()->break_at_compile()) {
340 // Open the debugger when compiing this method.
341 tty->print("### Breaking when compiling: ");
342 method()->print_short_name();
343 tty->cr();
344 BREAKPOINT;
345 }
347 if( PrintOpto ) {
348 if (is_osr_compilation()) {
349 tty->print("[OSR]%3d", _compile_id);
350 } else {
351 tty->print("%3d", _compile_id);
352 }
353 }
354 #endif
355 }
358 void Compile::init_scratch_buffer_blob() {
359 if( scratch_buffer_blob() != NULL ) return;
361 // Construct a temporary CodeBuffer to have it construct a BufferBlob
362 // Cache this BufferBlob for this compile.
363 ResourceMark rm;
364 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size);
365 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size);
366 // Record the buffer blob for next time.
367 set_scratch_buffer_blob(blob);
368 // Have we run out of code space?
369 if (scratch_buffer_blob() == NULL) {
370 // Let CompilerBroker disable further compilations.
371 record_failure("Not enough space for scratch buffer in CodeCache");
372 return;
373 }
375 // Initialize the relocation buffers
376 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size;
377 set_scratch_locs_memory(locs_buf);
378 }
381 //-----------------------scratch_emit_size-------------------------------------
382 // Helper function that computes size by emitting code
383 uint Compile::scratch_emit_size(const Node* n) {
384 // Emit into a trash buffer and count bytes emitted.
385 // This is a pretty expensive way to compute a size,
386 // but it works well enough if seldom used.
387 // All common fixed-size instructions are given a size
388 // method by the AD file.
389 // Note that the scratch buffer blob and locs memory are
390 // allocated at the beginning of the compile task, and
391 // may be shared by several calls to scratch_emit_size.
392 // The allocation of the scratch buffer blob is particularly
393 // expensive, since it has to grab the code cache lock.
394 BufferBlob* blob = this->scratch_buffer_blob();
395 assert(blob != NULL, "Initialize BufferBlob at start");
396 assert(blob->size() > MAX_inst_size, "sanity");
397 relocInfo* locs_buf = scratch_locs_memory();
398 address blob_begin = blob->instructions_begin();
399 address blob_end = (address)locs_buf;
400 assert(blob->instructions_contains(blob_end), "sanity");
401 CodeBuffer buf(blob_begin, blob_end - blob_begin);
402 buf.initialize_consts_size(MAX_const_size);
403 buf.initialize_stubs_size(MAX_stubs_size);
404 assert(locs_buf != NULL, "sanity");
405 int lsize = MAX_locs_size / 2;
406 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize);
407 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize);
408 n->emit(buf, this->regalloc());
409 return buf.code_size();
410 }
413 // ============================================================================
414 //------------------------------Compile standard-------------------------------
415 debug_only( int Compile::_debug_idx = 100000; )
417 // Compile a method. entry_bci is -1 for normal compilations and indicates
418 // the continuation bci for on stack replacement.
421 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
422 : Phase(Compiler),
423 _env(ci_env),
424 _log(ci_env->log()),
425 _compile_id(ci_env->compile_id()),
426 _save_argument_registers(false),
427 _stub_name(NULL),
428 _stub_function(NULL),
429 _stub_entry_point(NULL),
430 _method(target),
431 _entry_bci(osr_bci),
432 _initial_gvn(NULL),
433 _for_igvn(NULL),
434 _warm_calls(NULL),
435 _subsume_loads(subsume_loads),
436 _do_escape_analysis(do_escape_analysis),
437 _failure_reason(NULL),
438 _code_buffer("Compile::Fill_buffer"),
439 _orig_pc_slot(0),
440 _orig_pc_slot_offset_in_bytes(0),
441 _node_bundling_limit(0),
442 _node_bundling_base(NULL),
443 #ifndef PRODUCT
444 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
445 _printer(IdealGraphPrinter::printer()),
446 #endif
447 _congraph(NULL) {
448 C = this;
450 CompileWrapper cw(this);
451 #ifndef PRODUCT
452 if (TimeCompiler2) {
453 tty->print(" ");
454 target->holder()->name()->print();
455 tty->print(".");
456 target->print_short_name();
457 tty->print(" ");
458 }
459 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
460 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
461 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
462 if (!print_opto_assembly) {
463 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
464 if (print_assembly && !Disassembler::can_decode()) {
465 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
466 print_opto_assembly = true;
467 }
468 }
469 set_print_assembly(print_opto_assembly);
470 #endif
472 if (ProfileTraps) {
473 // Make sure the method being compiled gets its own MDO,
474 // so we can at least track the decompile_count().
475 method()->build_method_data();
476 }
478 Init(::AliasLevel);
481 print_compile_messages();
483 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
484 _ilt = InlineTree::build_inline_tree_root();
485 else
486 _ilt = NULL;
488 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
489 assert(num_alias_types() >= AliasIdxRaw, "");
491 #define MINIMUM_NODE_HASH 1023
492 // Node list that Iterative GVN will start with
493 Unique_Node_List for_igvn(comp_arena());
494 set_for_igvn(&for_igvn);
496 // GVN that will be run immediately on new nodes
497 uint estimated_size = method()->code_size()*4+64;
498 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
499 PhaseGVN gvn(node_arena(), estimated_size);
500 set_initial_gvn(&gvn);
502 { // Scope for timing the parser
503 TracePhase t3("parse", &_t_parser, true);
505 // Put top into the hash table ASAP.
506 initial_gvn()->transform_no_reclaim(top());
508 // Set up tf(), start(), and find a CallGenerator.
509 CallGenerator* cg;
510 if (is_osr_compilation()) {
511 const TypeTuple *domain = StartOSRNode::osr_domain();
512 const TypeTuple *range = TypeTuple::make_range(method()->signature());
513 init_tf(TypeFunc::make(domain, range));
514 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
515 initial_gvn()->set_type_bottom(s);
516 init_start(s);
517 cg = CallGenerator::for_osr(method(), entry_bci());
518 } else {
519 // Normal case.
520 init_tf(TypeFunc::make(method()));
521 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
522 initial_gvn()->set_type_bottom(s);
523 init_start(s);
524 float past_uses = method()->interpreter_invocation_count();
525 float expected_uses = past_uses;
526 cg = CallGenerator::for_inline(method(), expected_uses);
527 }
528 if (failing()) return;
529 if (cg == NULL) {
530 record_method_not_compilable_all_tiers("cannot parse method");
531 return;
532 }
533 JVMState* jvms = build_start_state(start(), tf());
534 if ((jvms = cg->generate(jvms)) == NULL) {
535 record_method_not_compilable("method parse failed");
536 return;
537 }
538 GraphKit kit(jvms);
540 if (!kit.stopped()) {
541 // Accept return values, and transfer control we know not where.
542 // This is done by a special, unique ReturnNode bound to root.
543 return_values(kit.jvms());
544 }
546 if (kit.has_exceptions()) {
547 // Any exceptions that escape from this call must be rethrown
548 // to whatever caller is dynamically above us on the stack.
549 // This is done by a special, unique RethrowNode bound to root.
550 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
551 }
553 // Remove clutter produced by parsing.
554 if (!failing()) {
555 ResourceMark rm;
556 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
557 }
558 }
560 // Note: Large methods are capped off in do_one_bytecode().
561 if (failing()) return;
563 // After parsing, node notes are no longer automagic.
564 // They must be propagated by register_new_node_with_optimizer(),
565 // clone(), or the like.
566 set_default_node_notes(NULL);
568 for (;;) {
569 int successes = Inline_Warm();
570 if (failing()) return;
571 if (successes == 0) break;
572 }
574 // Drain the list.
575 Finish_Warm();
576 #ifndef PRODUCT
577 if (_printer) {
578 _printer->print_inlining(this);
579 }
580 #endif
582 if (failing()) return;
583 NOT_PRODUCT( verify_graph_edges(); )
585 // Perform escape analysis
586 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
587 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
589 _congraph = new(comp_arena()) ConnectionGraph(this);
590 bool has_non_escaping_obj = _congraph->compute_escape();
592 #ifndef PRODUCT
593 if (PrintEscapeAnalysis) {
594 _congraph->dump();
595 }
596 #endif
597 if (!has_non_escaping_obj) {
598 _congraph = NULL;
599 }
601 if (failing()) return;
602 }
603 // Now optimize
604 Optimize();
605 if (failing()) return;
606 NOT_PRODUCT( verify_graph_edges(); )
608 print_method("Before Matching");
610 #ifndef PRODUCT
611 if (PrintIdeal) {
612 ttyLocker ttyl; // keep the following output all in one block
613 // This output goes directly to the tty, not the compiler log.
614 // To enable tools to match it up with the compilation activity,
615 // be sure to tag this tty output with the compile ID.
616 if (xtty != NULL) {
617 xtty->head("ideal compile_id='%d'%s", compile_id(),
618 is_osr_compilation() ? " compile_kind='osr'" :
619 "");
620 }
621 root()->dump(9999);
622 if (xtty != NULL) {
623 xtty->tail("ideal");
624 }
625 }
626 #endif
628 // Now that we know the size of all the monitors we can add a fixed slot
629 // for the original deopt pc.
631 _orig_pc_slot = fixed_slots();
632 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
633 set_fixed_slots(next_slot);
635 // Now generate code
636 Code_Gen();
637 if (failing()) return;
639 // Check if we want to skip execution of all compiled code.
640 {
641 #ifndef PRODUCT
642 if (OptoNoExecute) {
643 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
644 return;
645 }
646 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
647 #endif
649 if (is_osr_compilation()) {
650 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
651 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
652 } else {
653 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
654 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
655 }
657 env()->register_method(_method, _entry_bci,
658 &_code_offsets,
659 _orig_pc_slot_offset_in_bytes,
660 code_buffer(),
661 frame_size_in_words(), _oop_map_set,
662 &_handler_table, &_inc_table,
663 compiler,
664 env()->comp_level(),
665 true, /*has_debug_info*/
666 has_unsafe_access()
667 );
668 }
669 }
671 //------------------------------Compile----------------------------------------
672 // Compile a runtime stub
673 Compile::Compile( ciEnv* ci_env,
674 TypeFunc_generator generator,
675 address stub_function,
676 const char *stub_name,
677 int is_fancy_jump,
678 bool pass_tls,
679 bool save_arg_registers,
680 bool return_pc )
681 : Phase(Compiler),
682 _env(ci_env),
683 _log(ci_env->log()),
684 _compile_id(-1),
685 _save_argument_registers(save_arg_registers),
686 _method(NULL),
687 _stub_name(stub_name),
688 _stub_function(stub_function),
689 _stub_entry_point(NULL),
690 _entry_bci(InvocationEntryBci),
691 _initial_gvn(NULL),
692 _for_igvn(NULL),
693 _warm_calls(NULL),
694 _orig_pc_slot(0),
695 _orig_pc_slot_offset_in_bytes(0),
696 _subsume_loads(true),
697 _do_escape_analysis(false),
698 _failure_reason(NULL),
699 _code_buffer("Compile::Fill_buffer"),
700 _node_bundling_limit(0),
701 _node_bundling_base(NULL),
702 #ifndef PRODUCT
703 _trace_opto_output(TraceOptoOutput),
704 _printer(NULL),
705 #endif
706 _congraph(NULL) {
707 C = this;
709 #ifndef PRODUCT
710 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
711 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
712 set_print_assembly(PrintFrameConverterAssembly);
713 #endif
714 CompileWrapper cw(this);
715 Init(/*AliasLevel=*/ 0);
716 init_tf((*generator)());
718 {
719 // The following is a dummy for the sake of GraphKit::gen_stub
720 Unique_Node_List for_igvn(comp_arena());
721 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
722 PhaseGVN gvn(Thread::current()->resource_area(),255);
723 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
724 gvn.transform_no_reclaim(top());
726 GraphKit kit;
727 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
728 }
730 NOT_PRODUCT( verify_graph_edges(); )
731 Code_Gen();
732 if (failing()) return;
735 // Entry point will be accessed using compile->stub_entry_point();
736 if (code_buffer() == NULL) {
737 Matcher::soft_match_failure();
738 } else {
739 if (PrintAssembly && (WizardMode || Verbose))
740 tty->print_cr("### Stub::%s", stub_name);
742 if (!failing()) {
743 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
745 // Make the NMethod
746 // For now we mark the frame as never safe for profile stackwalking
747 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
748 code_buffer(),
749 CodeOffsets::frame_never_safe,
750 // _code_offsets.value(CodeOffsets::Frame_Complete),
751 frame_size_in_words(),
752 _oop_map_set,
753 save_arg_registers);
754 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
756 _stub_entry_point = rs->entry_point();
757 }
758 }
759 }
761 #ifndef PRODUCT
762 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
763 if(PrintOpto && Verbose) {
764 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
765 }
766 }
767 #endif
769 void Compile::print_codes() {
770 }
772 //------------------------------Init-------------------------------------------
773 // Prepare for a single compilation
774 void Compile::Init(int aliaslevel) {
775 _unique = 0;
776 _regalloc = NULL;
778 _tf = NULL; // filled in later
779 _top = NULL; // cached later
780 _matcher = NULL; // filled in later
781 _cfg = NULL; // filled in later
783 set_24_bit_selection_and_mode(Use24BitFP, false);
785 _node_note_array = NULL;
786 _default_node_notes = NULL;
788 _immutable_memory = NULL; // filled in at first inquiry
790 // Globally visible Nodes
791 // First set TOP to NULL to give safe behavior during creation of RootNode
792 set_cached_top_node(NULL);
793 set_root(new (this, 3) RootNode());
794 // Now that you have a Root to point to, create the real TOP
795 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
796 set_recent_alloc(NULL, NULL);
798 // Create Debug Information Recorder to record scopes, oopmaps, etc.
799 env()->set_oop_recorder(new OopRecorder(comp_arena()));
800 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
801 env()->set_dependencies(new Dependencies(env()));
803 _fixed_slots = 0;
804 set_has_split_ifs(false);
805 set_has_loops(has_method() && method()->has_loops()); // first approximation
806 _deopt_happens = true; // start out assuming the worst
807 _trap_can_recompile = false; // no traps emitted yet
808 _major_progress = true; // start out assuming good things will happen
809 set_has_unsafe_access(false);
810 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
811 set_decompile_count(0);
813 // Compilation level related initialization
814 if (env()->comp_level() == CompLevel_fast_compile) {
815 set_num_loop_opts(Tier1LoopOptsCount);
816 set_do_inlining(Tier1Inline != 0);
817 set_max_inline_size(Tier1MaxInlineSize);
818 set_freq_inline_size(Tier1FreqInlineSize);
819 set_do_scheduling(false);
820 set_do_count_invocations(Tier1CountInvocations);
821 set_do_method_data_update(Tier1UpdateMethodData);
822 } else {
823 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
824 set_num_loop_opts(LoopOptsCount);
825 set_do_inlining(Inline);
826 set_max_inline_size(MaxInlineSize);
827 set_freq_inline_size(FreqInlineSize);
828 set_do_scheduling(OptoScheduling);
829 set_do_count_invocations(false);
830 set_do_method_data_update(false);
831 }
833 if (debug_info()->recording_non_safepoints()) {
834 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
835 (comp_arena(), 8, 0, NULL));
836 set_default_node_notes(Node_Notes::make(this));
837 }
839 // // -- Initialize types before each compile --
840 // // Update cached type information
841 // if( _method && _method->constants() )
842 // Type::update_loaded_types(_method, _method->constants());
844 // Init alias_type map.
845 if (!_do_escape_analysis && aliaslevel == 3)
846 aliaslevel = 2; // No unique types without escape analysis
847 _AliasLevel = aliaslevel;
848 const int grow_ats = 16;
849 _max_alias_types = grow_ats;
850 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
851 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
852 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
853 {
854 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
855 }
856 // Initialize the first few types.
857 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
858 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
859 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
860 _num_alias_types = AliasIdxRaw+1;
861 // Zero out the alias type cache.
862 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
863 // A NULL adr_type hits in the cache right away. Preload the right answer.
864 probe_alias_cache(NULL)->_index = AliasIdxTop;
866 _intrinsics = NULL;
867 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
868 register_library_intrinsics();
869 }
871 //---------------------------init_start----------------------------------------
872 // Install the StartNode on this compile object.
873 void Compile::init_start(StartNode* s) {
874 if (failing())
875 return; // already failing
876 assert(s == start(), "");
877 }
879 StartNode* Compile::start() const {
880 assert(!failing(), "");
881 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
882 Node* start = root()->fast_out(i);
883 if( start->is_Start() )
884 return start->as_Start();
885 }
886 ShouldNotReachHere();
887 return NULL;
888 }
890 //-------------------------------immutable_memory-------------------------------------
891 // Access immutable memory
892 Node* Compile::immutable_memory() {
893 if (_immutable_memory != NULL) {
894 return _immutable_memory;
895 }
896 StartNode* s = start();
897 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
898 Node *p = s->fast_out(i);
899 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
900 _immutable_memory = p;
901 return _immutable_memory;
902 }
903 }
904 ShouldNotReachHere();
905 return NULL;
906 }
908 //----------------------set_cached_top_node------------------------------------
909 // Install the cached top node, and make sure Node::is_top works correctly.
910 void Compile::set_cached_top_node(Node* tn) {
911 if (tn != NULL) verify_top(tn);
912 Node* old_top = _top;
913 _top = tn;
914 // Calling Node::setup_is_top allows the nodes the chance to adjust
915 // their _out arrays.
916 if (_top != NULL) _top->setup_is_top();
917 if (old_top != NULL) old_top->setup_is_top();
918 assert(_top == NULL || top()->is_top(), "");
919 }
921 #ifndef PRODUCT
922 void Compile::verify_top(Node* tn) const {
923 if (tn != NULL) {
924 assert(tn->is_Con(), "top node must be a constant");
925 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
926 assert(tn->in(0) != NULL, "must have live top node");
927 }
928 }
929 #endif
932 ///-------------------Managing Per-Node Debug & Profile Info-------------------
934 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
935 guarantee(arr != NULL, "");
936 int num_blocks = arr->length();
937 if (grow_by < num_blocks) grow_by = num_blocks;
938 int num_notes = grow_by * _node_notes_block_size;
939 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
940 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
941 while (num_notes > 0) {
942 arr->append(notes);
943 notes += _node_notes_block_size;
944 num_notes -= _node_notes_block_size;
945 }
946 assert(num_notes == 0, "exact multiple, please");
947 }
949 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
950 if (source == NULL || dest == NULL) return false;
952 if (dest->is_Con())
953 return false; // Do not push debug info onto constants.
955 #ifdef ASSERT
956 // Leave a bread crumb trail pointing to the original node:
957 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
958 dest->set_debug_orig(source);
959 }
960 #endif
962 if (node_note_array() == NULL)
963 return false; // Not collecting any notes now.
965 // This is a copy onto a pre-existing node, which may already have notes.
966 // If both nodes have notes, do not overwrite any pre-existing notes.
967 Node_Notes* source_notes = node_notes_at(source->_idx);
968 if (source_notes == NULL || source_notes->is_clear()) return false;
969 Node_Notes* dest_notes = node_notes_at(dest->_idx);
970 if (dest_notes == NULL || dest_notes->is_clear()) {
971 return set_node_notes_at(dest->_idx, source_notes);
972 }
974 Node_Notes merged_notes = (*source_notes);
975 // The order of operations here ensures that dest notes will win...
976 merged_notes.update_from(dest_notes);
977 return set_node_notes_at(dest->_idx, &merged_notes);
978 }
981 //--------------------------allow_range_check_smearing-------------------------
982 // Gating condition for coalescing similar range checks.
983 // Sometimes we try 'speculatively' replacing a series of a range checks by a
984 // single covering check that is at least as strong as any of them.
985 // If the optimization succeeds, the simplified (strengthened) range check
986 // will always succeed. If it fails, we will deopt, and then give up
987 // on the optimization.
988 bool Compile::allow_range_check_smearing() const {
989 // If this method has already thrown a range-check,
990 // assume it was because we already tried range smearing
991 // and it failed.
992 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
993 return !already_trapped;
994 }
997 //------------------------------flatten_alias_type-----------------------------
998 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
999 int offset = tj->offset();
1000 TypePtr::PTR ptr = tj->ptr();
1002 // Known instance (scalarizable allocation) alias only with itself.
1003 bool is_known_inst = tj->isa_oopptr() != NULL &&
1004 tj->is_oopptr()->is_known_instance();
1006 // Process weird unsafe references.
1007 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1008 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1009 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1010 tj = TypeOopPtr::BOTTOM;
1011 ptr = tj->ptr();
1012 offset = tj->offset();
1013 }
1015 // Array pointers need some flattening
1016 const TypeAryPtr *ta = tj->isa_aryptr();
1017 if( ta && is_known_inst ) {
1018 if ( offset != Type::OffsetBot &&
1019 offset > arrayOopDesc::length_offset_in_bytes() ) {
1020 offset = Type::OffsetBot; // Flatten constant access into array body only
1021 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1022 }
1023 } else if( ta && _AliasLevel >= 2 ) {
1024 // For arrays indexed by constant indices, we flatten the alias
1025 // space to include all of the array body. Only the header, klass
1026 // and array length can be accessed un-aliased.
1027 if( offset != Type::OffsetBot ) {
1028 if( ta->const_oop() ) { // methodDataOop or methodOop
1029 offset = Type::OffsetBot; // Flatten constant access into array body
1030 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1031 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1032 // range is OK as-is.
1033 tj = ta = TypeAryPtr::RANGE;
1034 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1035 tj = TypeInstPtr::KLASS; // all klass loads look alike
1036 ta = TypeAryPtr::RANGE; // generic ignored junk
1037 ptr = TypePtr::BotPTR;
1038 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1039 tj = TypeInstPtr::MARK;
1040 ta = TypeAryPtr::RANGE; // generic ignored junk
1041 ptr = TypePtr::BotPTR;
1042 } else { // Random constant offset into array body
1043 offset = Type::OffsetBot; // Flatten constant access into array body
1044 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1045 }
1046 }
1047 // Arrays of fixed size alias with arrays of unknown size.
1048 if (ta->size() != TypeInt::POS) {
1049 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1050 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1051 }
1052 // Arrays of known objects become arrays of unknown objects.
1053 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1054 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1055 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1056 }
1057 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1058 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1059 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1060 }
1061 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1062 // cannot be distinguished by bytecode alone.
1063 if (ta->elem() == TypeInt::BOOL) {
1064 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1065 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1066 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1067 }
1068 // During the 2nd round of IterGVN, NotNull castings are removed.
1069 // Make sure the Bottom and NotNull variants alias the same.
1070 // Also, make sure exact and non-exact variants alias the same.
1071 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1072 if (ta->const_oop()) {
1073 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1074 } else {
1075 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1076 }
1077 }
1078 }
1080 // Oop pointers need some flattening
1081 const TypeInstPtr *to = tj->isa_instptr();
1082 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1083 if( ptr == TypePtr::Constant ) {
1084 // No constant oop pointers (such as Strings); they alias with
1085 // unknown strings.
1086 assert(!is_known_inst, "not scalarizable allocation");
1087 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1088 } else if( is_known_inst ) {
1089 tj = to; // Keep NotNull and klass_is_exact for instance type
1090 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1091 // During the 2nd round of IterGVN, NotNull castings are removed.
1092 // Make sure the Bottom and NotNull variants alias the same.
1093 // Also, make sure exact and non-exact variants alias the same.
1094 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1095 }
1096 // Canonicalize the holder of this field
1097 ciInstanceKlass *k = to->klass()->as_instance_klass();
1098 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1099 // First handle header references such as a LoadKlassNode, even if the
1100 // object's klass is unloaded at compile time (4965979).
1101 if (!is_known_inst) { // Do it only for non-instance types
1102 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1103 }
1104 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1105 to = NULL;
1106 tj = TypeOopPtr::BOTTOM;
1107 offset = tj->offset();
1108 } else {
1109 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1110 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1111 if( is_known_inst ) {
1112 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1113 } else {
1114 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1115 }
1116 }
1117 }
1118 }
1120 // Klass pointers to object array klasses need some flattening
1121 const TypeKlassPtr *tk = tj->isa_klassptr();
1122 if( tk ) {
1123 // If we are referencing a field within a Klass, we need
1124 // to assume the worst case of an Object. Both exact and
1125 // inexact types must flatten to the same alias class.
1126 // Since the flattened result for a klass is defined to be
1127 // precisely java.lang.Object, use a constant ptr.
1128 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1130 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1131 TypeKlassPtr::OBJECT->klass(),
1132 offset);
1133 }
1135 ciKlass* klass = tk->klass();
1136 if( klass->is_obj_array_klass() ) {
1137 ciKlass* k = TypeAryPtr::OOPS->klass();
1138 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1139 k = TypeInstPtr::BOTTOM->klass();
1140 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1141 }
1143 // Check for precise loads from the primary supertype array and force them
1144 // to the supertype cache alias index. Check for generic array loads from
1145 // the primary supertype array and also force them to the supertype cache
1146 // alias index. Since the same load can reach both, we need to merge
1147 // these 2 disparate memories into the same alias class. Since the
1148 // primary supertype array is read-only, there's no chance of confusion
1149 // where we bypass an array load and an array store.
1150 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1151 if( offset == Type::OffsetBot ||
1152 off2 < Klass::primary_super_limit()*wordSize ) {
1153 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1154 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1155 }
1156 }
1158 // Flatten all Raw pointers together.
1159 if (tj->base() == Type::RawPtr)
1160 tj = TypeRawPtr::BOTTOM;
1162 if (tj->base() == Type::AnyPtr)
1163 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1165 // Flatten all to bottom for now
1166 switch( _AliasLevel ) {
1167 case 0:
1168 tj = TypePtr::BOTTOM;
1169 break;
1170 case 1: // Flatten to: oop, static, field or array
1171 switch (tj->base()) {
1172 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1173 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1174 case Type::AryPtr: // do not distinguish arrays at all
1175 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1176 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1177 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1178 default: ShouldNotReachHere();
1179 }
1180 break;
1181 case 2: // No collasping at level 2; keep all splits
1182 case 3: // No collasping at level 3; keep all splits
1183 break;
1184 default:
1185 Unimplemented();
1186 }
1188 offset = tj->offset();
1189 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1191 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1192 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1193 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1194 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1195 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1196 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1197 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1198 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1199 assert( tj->ptr() != TypePtr::TopPTR &&
1200 tj->ptr() != TypePtr::AnyNull &&
1201 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1202 // assert( tj->ptr() != TypePtr::Constant ||
1203 // tj->base() == Type::RawPtr ||
1204 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1206 return tj;
1207 }
1209 void Compile::AliasType::Init(int i, const TypePtr* at) {
1210 _index = i;
1211 _adr_type = at;
1212 _field = NULL;
1213 _is_rewritable = true; // default
1214 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1215 if (atoop != NULL && atoop->is_known_instance()) {
1216 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1217 _general_index = Compile::current()->get_alias_index(gt);
1218 } else {
1219 _general_index = 0;
1220 }
1221 }
1223 //---------------------------------print_on------------------------------------
1224 #ifndef PRODUCT
1225 void Compile::AliasType::print_on(outputStream* st) {
1226 if (index() < 10)
1227 st->print("@ <%d> ", index());
1228 else st->print("@ <%d>", index());
1229 st->print(is_rewritable() ? " " : " RO");
1230 int offset = adr_type()->offset();
1231 if (offset == Type::OffsetBot)
1232 st->print(" +any");
1233 else st->print(" +%-3d", offset);
1234 st->print(" in ");
1235 adr_type()->dump_on(st);
1236 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1237 if (field() != NULL && tjp) {
1238 if (tjp->klass() != field()->holder() ||
1239 tjp->offset() != field()->offset_in_bytes()) {
1240 st->print(" != ");
1241 field()->print();
1242 st->print(" ***");
1243 }
1244 }
1245 }
1247 void print_alias_types() {
1248 Compile* C = Compile::current();
1249 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1250 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1251 C->alias_type(idx)->print_on(tty);
1252 tty->cr();
1253 }
1254 }
1255 #endif
1258 //----------------------------probe_alias_cache--------------------------------
1259 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1260 intptr_t key = (intptr_t) adr_type;
1261 key ^= key >> logAliasCacheSize;
1262 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1263 }
1266 //-----------------------------grow_alias_types--------------------------------
1267 void Compile::grow_alias_types() {
1268 const int old_ats = _max_alias_types; // how many before?
1269 const int new_ats = old_ats; // how many more?
1270 const int grow_ats = old_ats+new_ats; // how many now?
1271 _max_alias_types = grow_ats;
1272 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1273 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1274 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1275 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1276 }
1279 //--------------------------------find_alias_type------------------------------
1280 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1281 if (_AliasLevel == 0)
1282 return alias_type(AliasIdxBot);
1284 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1285 if (ace->_adr_type == adr_type) {
1286 return alias_type(ace->_index);
1287 }
1289 // Handle special cases.
1290 if (adr_type == NULL) return alias_type(AliasIdxTop);
1291 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1293 // Do it the slow way.
1294 const TypePtr* flat = flatten_alias_type(adr_type);
1296 #ifdef ASSERT
1297 assert(flat == flatten_alias_type(flat), "idempotent");
1298 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1299 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1300 const TypeOopPtr* foop = flat->is_oopptr();
1301 // Scalarizable allocations have exact klass always.
1302 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1303 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1304 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1305 }
1306 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1307 #endif
1309 int idx = AliasIdxTop;
1310 for (int i = 0; i < num_alias_types(); i++) {
1311 if (alias_type(i)->adr_type() == flat) {
1312 idx = i;
1313 break;
1314 }
1315 }
1317 if (idx == AliasIdxTop) {
1318 if (no_create) return NULL;
1319 // Grow the array if necessary.
1320 if (_num_alias_types == _max_alias_types) grow_alias_types();
1321 // Add a new alias type.
1322 idx = _num_alias_types++;
1323 _alias_types[idx]->Init(idx, flat);
1324 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1325 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1326 if (flat->isa_instptr()) {
1327 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1328 && flat->is_instptr()->klass() == env()->Class_klass())
1329 alias_type(idx)->set_rewritable(false);
1330 }
1331 if (flat->isa_klassptr()) {
1332 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1333 alias_type(idx)->set_rewritable(false);
1334 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1335 alias_type(idx)->set_rewritable(false);
1336 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1337 alias_type(idx)->set_rewritable(false);
1338 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1339 alias_type(idx)->set_rewritable(false);
1340 }
1341 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1342 // but the base pointer type is not distinctive enough to identify
1343 // references into JavaThread.)
1345 // Check for final instance fields.
1346 const TypeInstPtr* tinst = flat->isa_instptr();
1347 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1348 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1349 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1350 // Set field() and is_rewritable() attributes.
1351 if (field != NULL) alias_type(idx)->set_field(field);
1352 }
1353 const TypeKlassPtr* tklass = flat->isa_klassptr();
1354 // Check for final static fields.
1355 if (tklass && tklass->klass()->is_instance_klass()) {
1356 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1357 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1358 // Set field() and is_rewritable() attributes.
1359 if (field != NULL) alias_type(idx)->set_field(field);
1360 }
1361 }
1363 // Fill the cache for next time.
1364 ace->_adr_type = adr_type;
1365 ace->_index = idx;
1366 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1368 // Might as well try to fill the cache for the flattened version, too.
1369 AliasCacheEntry* face = probe_alias_cache(flat);
1370 if (face->_adr_type == NULL) {
1371 face->_adr_type = flat;
1372 face->_index = idx;
1373 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1374 }
1376 return alias_type(idx);
1377 }
1380 Compile::AliasType* Compile::alias_type(ciField* field) {
1381 const TypeOopPtr* t;
1382 if (field->is_static())
1383 t = TypeKlassPtr::make(field->holder());
1384 else
1385 t = TypeOopPtr::make_from_klass_raw(field->holder());
1386 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1387 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1388 return atp;
1389 }
1392 //------------------------------have_alias_type--------------------------------
1393 bool Compile::have_alias_type(const TypePtr* adr_type) {
1394 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1395 if (ace->_adr_type == adr_type) {
1396 return true;
1397 }
1399 // Handle special cases.
1400 if (adr_type == NULL) return true;
1401 if (adr_type == TypePtr::BOTTOM) return true;
1403 return find_alias_type(adr_type, true) != NULL;
1404 }
1406 //-----------------------------must_alias--------------------------------------
1407 // True if all values of the given address type are in the given alias category.
1408 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1409 if (alias_idx == AliasIdxBot) return true; // the universal category
1410 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1411 if (alias_idx == AliasIdxTop) return false; // the empty category
1412 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1414 // the only remaining possible overlap is identity
1415 int adr_idx = get_alias_index(adr_type);
1416 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1417 assert(adr_idx == alias_idx ||
1418 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1419 && adr_type != TypeOopPtr::BOTTOM),
1420 "should not be testing for overlap with an unsafe pointer");
1421 return adr_idx == alias_idx;
1422 }
1424 //------------------------------can_alias--------------------------------------
1425 // True if any values of the given address type are in the given alias category.
1426 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1427 if (alias_idx == AliasIdxTop) return false; // the empty category
1428 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1429 if (alias_idx == AliasIdxBot) return true; // the universal category
1430 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1432 // the only remaining possible overlap is identity
1433 int adr_idx = get_alias_index(adr_type);
1434 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1435 return adr_idx == alias_idx;
1436 }
1440 //---------------------------pop_warm_call-------------------------------------
1441 WarmCallInfo* Compile::pop_warm_call() {
1442 WarmCallInfo* wci = _warm_calls;
1443 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1444 return wci;
1445 }
1447 //----------------------------Inline_Warm--------------------------------------
1448 int Compile::Inline_Warm() {
1449 // If there is room, try to inline some more warm call sites.
1450 // %%% Do a graph index compaction pass when we think we're out of space?
1451 if (!InlineWarmCalls) return 0;
1453 int calls_made_hot = 0;
1454 int room_to_grow = NodeCountInliningCutoff - unique();
1455 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1456 int amount_grown = 0;
1457 WarmCallInfo* call;
1458 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1459 int est_size = (int)call->size();
1460 if (est_size > (room_to_grow - amount_grown)) {
1461 // This one won't fit anyway. Get rid of it.
1462 call->make_cold();
1463 continue;
1464 }
1465 call->make_hot();
1466 calls_made_hot++;
1467 amount_grown += est_size;
1468 amount_to_grow -= est_size;
1469 }
1471 if (calls_made_hot > 0) set_major_progress();
1472 return calls_made_hot;
1473 }
1476 //----------------------------Finish_Warm--------------------------------------
1477 void Compile::Finish_Warm() {
1478 if (!InlineWarmCalls) return;
1479 if (failing()) return;
1480 if (warm_calls() == NULL) return;
1482 // Clean up loose ends, if we are out of space for inlining.
1483 WarmCallInfo* call;
1484 while ((call = pop_warm_call()) != NULL) {
1485 call->make_cold();
1486 }
1487 }
1490 //------------------------------Optimize---------------------------------------
1491 // Given a graph, optimize it.
1492 void Compile::Optimize() {
1493 TracePhase t1("optimizer", &_t_optimizer, true);
1495 #ifndef PRODUCT
1496 if (env()->break_at_compile()) {
1497 BREAKPOINT;
1498 }
1500 #endif
1502 ResourceMark rm;
1503 int loop_opts_cnt;
1505 NOT_PRODUCT( verify_graph_edges(); )
1507 print_method("After Parsing");
1509 {
1510 // Iterative Global Value Numbering, including ideal transforms
1511 // Initialize IterGVN with types and values from parse-time GVN
1512 PhaseIterGVN igvn(initial_gvn());
1513 {
1514 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1515 igvn.optimize();
1516 }
1518 print_method("Iter GVN 1", 2);
1520 if (failing()) return;
1522 // get rid of the connection graph since it's information is not
1523 // updated by optimizations
1524 _congraph = NULL;
1527 // Loop transforms on the ideal graph. Range Check Elimination,
1528 // peeling, unrolling, etc.
1530 // Set loop opts counter
1531 loop_opts_cnt = num_loop_opts();
1532 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1533 {
1534 TracePhase t2("idealLoop", &_t_idealLoop, true);
1535 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1536 loop_opts_cnt--;
1537 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1538 if (failing()) return;
1539 }
1540 // Loop opts pass if partial peeling occurred in previous pass
1541 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1542 TracePhase t3("idealLoop", &_t_idealLoop, true);
1543 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1544 loop_opts_cnt--;
1545 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1546 if (failing()) return;
1547 }
1548 // Loop opts pass for loop-unrolling before CCP
1549 if(major_progress() && (loop_opts_cnt > 0)) {
1550 TracePhase t4("idealLoop", &_t_idealLoop, true);
1551 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1552 loop_opts_cnt--;
1553 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1554 }
1555 }
1556 if (failing()) return;
1558 // Conditional Constant Propagation;
1559 PhaseCCP ccp( &igvn );
1560 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1561 {
1562 TracePhase t2("ccp", &_t_ccp, true);
1563 ccp.do_transform();
1564 }
1565 print_method("PhaseCPP 1", 2);
1567 assert( true, "Break here to ccp.dump_old2new_map()");
1569 // Iterative Global Value Numbering, including ideal transforms
1570 {
1571 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1572 igvn = ccp;
1573 igvn.optimize();
1574 }
1576 print_method("Iter GVN 2", 2);
1578 if (failing()) return;
1580 // Loop transforms on the ideal graph. Range Check Elimination,
1581 // peeling, unrolling, etc.
1582 if(loop_opts_cnt > 0) {
1583 debug_only( int cnt = 0; );
1584 while(major_progress() && (loop_opts_cnt > 0)) {
1585 TracePhase t2("idealLoop", &_t_idealLoop, true);
1586 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1587 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1588 loop_opts_cnt--;
1589 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1590 if (failing()) return;
1591 }
1592 }
1593 {
1594 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1595 PhaseMacroExpand mex(igvn);
1596 if (mex.expand_macro_nodes()) {
1597 assert(failing(), "must bail out w/ explicit message");
1598 return;
1599 }
1600 }
1602 } // (End scope of igvn; run destructor if necessary for asserts.)
1604 // A method with only infinite loops has no edges entering loops from root
1605 {
1606 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1607 if (final_graph_reshaping()) {
1608 assert(failing(), "must bail out w/ explicit message");
1609 return;
1610 }
1611 }
1613 print_method("Optimize finished", 2);
1614 }
1617 //------------------------------Code_Gen---------------------------------------
1618 // Given a graph, generate code for it
1619 void Compile::Code_Gen() {
1620 if (failing()) return;
1622 // Perform instruction selection. You might think we could reclaim Matcher
1623 // memory PDQ, but actually the Matcher is used in generating spill code.
1624 // Internals of the Matcher (including some VectorSets) must remain live
1625 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1626 // set a bit in reclaimed memory.
1628 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1629 // nodes. Mapping is only valid at the root of each matched subtree.
1630 NOT_PRODUCT( verify_graph_edges(); )
1632 Node_List proj_list;
1633 Matcher m(proj_list);
1634 _matcher = &m;
1635 {
1636 TracePhase t2("matcher", &_t_matcher, true);
1637 m.match();
1638 }
1639 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1640 // nodes. Mapping is only valid at the root of each matched subtree.
1641 NOT_PRODUCT( verify_graph_edges(); )
1643 // If you have too many nodes, or if matching has failed, bail out
1644 check_node_count(0, "out of nodes matching instructions");
1645 if (failing()) return;
1647 // Build a proper-looking CFG
1648 PhaseCFG cfg(node_arena(), root(), m);
1649 _cfg = &cfg;
1650 {
1651 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1652 cfg.Dominators();
1653 if (failing()) return;
1655 NOT_PRODUCT( verify_graph_edges(); )
1657 cfg.Estimate_Block_Frequency();
1658 cfg.GlobalCodeMotion(m,unique(),proj_list);
1660 print_method("Global code motion", 2);
1662 if (failing()) return;
1663 NOT_PRODUCT( verify_graph_edges(); )
1665 debug_only( cfg.verify(); )
1666 }
1667 NOT_PRODUCT( verify_graph_edges(); )
1669 PhaseChaitin regalloc(unique(),cfg,m);
1670 _regalloc = ®alloc;
1671 {
1672 TracePhase t2("regalloc", &_t_registerAllocation, true);
1673 // Perform any platform dependent preallocation actions. This is used,
1674 // for example, to avoid taking an implicit null pointer exception
1675 // using the frame pointer on win95.
1676 _regalloc->pd_preallocate_hook();
1678 // Perform register allocation. After Chaitin, use-def chains are
1679 // no longer accurate (at spill code) and so must be ignored.
1680 // Node->LRG->reg mappings are still accurate.
1681 _regalloc->Register_Allocate();
1683 // Bail out if the allocator builds too many nodes
1684 if (failing()) return;
1685 }
1687 // Prior to register allocation we kept empty basic blocks in case the
1688 // the allocator needed a place to spill. After register allocation we
1689 // are not adding any new instructions. If any basic block is empty, we
1690 // can now safely remove it.
1691 {
1692 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); )
1693 cfg.RemoveEmpty();
1694 }
1696 // Perform any platform dependent postallocation verifications.
1697 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1699 // Apply peephole optimizations
1700 if( OptoPeephole ) {
1701 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1702 PhasePeephole peep( _regalloc, cfg);
1703 peep.do_transform();
1704 }
1706 // Convert Nodes to instruction bits in a buffer
1707 {
1708 // %%%% workspace merge brought two timers together for one job
1709 TracePhase t2a("output", &_t_output, true);
1710 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1711 Output();
1712 }
1714 print_method("Final Code");
1716 // He's dead, Jim.
1717 _cfg = (PhaseCFG*)0xdeadbeef;
1718 _regalloc = (PhaseChaitin*)0xdeadbeef;
1719 }
1722 //------------------------------dump_asm---------------------------------------
1723 // Dump formatted assembly
1724 #ifndef PRODUCT
1725 void Compile::dump_asm(int *pcs, uint pc_limit) {
1726 bool cut_short = false;
1727 tty->print_cr("#");
1728 tty->print("# "); _tf->dump(); tty->cr();
1729 tty->print_cr("#");
1731 // For all blocks
1732 int pc = 0x0; // Program counter
1733 char starts_bundle = ' ';
1734 _regalloc->dump_frame();
1736 Node *n = NULL;
1737 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1738 if (VMThread::should_terminate()) { cut_short = true; break; }
1739 Block *b = _cfg->_blocks[i];
1740 if (b->is_connector() && !Verbose) continue;
1741 n = b->_nodes[0];
1742 if (pcs && n->_idx < pc_limit)
1743 tty->print("%3.3x ", pcs[n->_idx]);
1744 else
1745 tty->print(" ");
1746 b->dump_head( &_cfg->_bbs );
1747 if (b->is_connector()) {
1748 tty->print_cr(" # Empty connector block");
1749 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1750 tty->print_cr(" # Block is sole successor of call");
1751 }
1753 // For all instructions
1754 Node *delay = NULL;
1755 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1756 if (VMThread::should_terminate()) { cut_short = true; break; }
1757 n = b->_nodes[j];
1758 if (valid_bundle_info(n)) {
1759 Bundle *bundle = node_bundling(n);
1760 if (bundle->used_in_unconditional_delay()) {
1761 delay = n;
1762 continue;
1763 }
1764 if (bundle->starts_bundle())
1765 starts_bundle = '+';
1766 }
1768 if (WizardMode) n->dump();
1770 if( !n->is_Region() && // Dont print in the Assembly
1771 !n->is_Phi() && // a few noisely useless nodes
1772 !n->is_Proj() &&
1773 !n->is_MachTemp() &&
1774 !n->is_Catch() && // Would be nice to print exception table targets
1775 !n->is_MergeMem() && // Not very interesting
1776 !n->is_top() && // Debug info table constants
1777 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1778 ) {
1779 if (pcs && n->_idx < pc_limit)
1780 tty->print("%3.3x", pcs[n->_idx]);
1781 else
1782 tty->print(" ");
1783 tty->print(" %c ", starts_bundle);
1784 starts_bundle = ' ';
1785 tty->print("\t");
1786 n->format(_regalloc, tty);
1787 tty->cr();
1788 }
1790 // If we have an instruction with a delay slot, and have seen a delay,
1791 // then back up and print it
1792 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1793 assert(delay != NULL, "no unconditional delay instruction");
1794 if (WizardMode) delay->dump();
1796 if (node_bundling(delay)->starts_bundle())
1797 starts_bundle = '+';
1798 if (pcs && n->_idx < pc_limit)
1799 tty->print("%3.3x", pcs[n->_idx]);
1800 else
1801 tty->print(" ");
1802 tty->print(" %c ", starts_bundle);
1803 starts_bundle = ' ';
1804 tty->print("\t");
1805 delay->format(_regalloc, tty);
1806 tty->print_cr("");
1807 delay = NULL;
1808 }
1810 // Dump the exception table as well
1811 if( n->is_Catch() && (Verbose || WizardMode) ) {
1812 // Print the exception table for this offset
1813 _handler_table.print_subtable_for(pc);
1814 }
1815 }
1817 if (pcs && n->_idx < pc_limit)
1818 tty->print_cr("%3.3x", pcs[n->_idx]);
1819 else
1820 tty->print_cr("");
1822 assert(cut_short || delay == NULL, "no unconditional delay branch");
1824 } // End of per-block dump
1825 tty->print_cr("");
1827 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1828 }
1829 #endif
1831 //------------------------------Final_Reshape_Counts---------------------------
1832 // This class defines counters to help identify when a method
1833 // may/must be executed using hardware with only 24-bit precision.
1834 struct Final_Reshape_Counts : public StackObj {
1835 int _call_count; // count non-inlined 'common' calls
1836 int _float_count; // count float ops requiring 24-bit precision
1837 int _double_count; // count double ops requiring more precision
1838 int _java_call_count; // count non-inlined 'java' calls
1839 VectorSet _visited; // Visitation flags
1840 Node_List _tests; // Set of IfNodes & PCTableNodes
1842 Final_Reshape_Counts() :
1843 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0),
1844 _visited( Thread::current()->resource_area() ) { }
1846 void inc_call_count () { _call_count ++; }
1847 void inc_float_count () { _float_count ++; }
1848 void inc_double_count() { _double_count++; }
1849 void inc_java_call_count() { _java_call_count++; }
1851 int get_call_count () const { return _call_count ; }
1852 int get_float_count () const { return _float_count ; }
1853 int get_double_count() const { return _double_count; }
1854 int get_java_call_count() const { return _java_call_count; }
1855 };
1857 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1858 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1859 // Make sure the offset goes inside the instance layout.
1860 return k->contains_field_offset(tp->offset());
1861 // Note that OffsetBot and OffsetTop are very negative.
1862 }
1864 //------------------------------final_graph_reshaping_impl----------------------
1865 // Implement items 1-5 from final_graph_reshaping below.
1866 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
1868 if ( n->outcnt() == 0 ) return; // dead node
1869 uint nop = n->Opcode();
1871 // Check for 2-input instruction with "last use" on right input.
1872 // Swap to left input. Implements item (2).
1873 if( n->req() == 3 && // two-input instruction
1874 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1875 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1876 n->in(2)->outcnt() == 1 &&// right use IS a last use
1877 !n->in(2)->is_Con() ) { // right use is not a constant
1878 // Check for commutative opcode
1879 switch( nop ) {
1880 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1881 case Op_MaxI: case Op_MinI:
1882 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1883 case Op_AndL: case Op_XorL: case Op_OrL:
1884 case Op_AndI: case Op_XorI: case Op_OrI: {
1885 // Move "last use" input to left by swapping inputs
1886 n->swap_edges(1, 2);
1887 break;
1888 }
1889 default:
1890 break;
1891 }
1892 }
1894 // Count FPU ops and common calls, implements item (3)
1895 switch( nop ) {
1896 // Count all float operations that may use FPU
1897 case Op_AddF:
1898 case Op_SubF:
1899 case Op_MulF:
1900 case Op_DivF:
1901 case Op_NegF:
1902 case Op_ModF:
1903 case Op_ConvI2F:
1904 case Op_ConF:
1905 case Op_CmpF:
1906 case Op_CmpF3:
1907 // case Op_ConvL2F: // longs are split into 32-bit halves
1908 fpu.inc_float_count();
1909 break;
1911 case Op_ConvF2D:
1912 case Op_ConvD2F:
1913 fpu.inc_float_count();
1914 fpu.inc_double_count();
1915 break;
1917 // Count all double operations that may use FPU
1918 case Op_AddD:
1919 case Op_SubD:
1920 case Op_MulD:
1921 case Op_DivD:
1922 case Op_NegD:
1923 case Op_ModD:
1924 case Op_ConvI2D:
1925 case Op_ConvD2I:
1926 // case Op_ConvL2D: // handled by leaf call
1927 // case Op_ConvD2L: // handled by leaf call
1928 case Op_ConD:
1929 case Op_CmpD:
1930 case Op_CmpD3:
1931 fpu.inc_double_count();
1932 break;
1933 case Op_Opaque1: // Remove Opaque Nodes before matching
1934 case Op_Opaque2: // Remove Opaque Nodes before matching
1935 n->subsume_by(n->in(1));
1936 break;
1937 case Op_CallStaticJava:
1938 case Op_CallJava:
1939 case Op_CallDynamicJava:
1940 fpu.inc_java_call_count(); // Count java call site;
1941 case Op_CallRuntime:
1942 case Op_CallLeaf:
1943 case Op_CallLeafNoFP: {
1944 assert( n->is_Call(), "" );
1945 CallNode *call = n->as_Call();
1946 // Count call sites where the FP mode bit would have to be flipped.
1947 // Do not count uncommon runtime calls:
1948 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
1949 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
1950 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
1951 fpu.inc_call_count(); // Count the call site
1952 } else { // See if uncommon argument is shared
1953 Node *n = call->in(TypeFunc::Parms);
1954 int nop = n->Opcode();
1955 // Clone shared simple arguments to uncommon calls, item (1).
1956 if( n->outcnt() > 1 &&
1957 !n->is_Proj() &&
1958 nop != Op_CreateEx &&
1959 nop != Op_CheckCastPP &&
1960 !n->is_Mem() ) {
1961 Node *x = n->clone();
1962 call->set_req( TypeFunc::Parms, x );
1963 }
1964 }
1965 break;
1966 }
1968 case Op_StoreD:
1969 case Op_LoadD:
1970 case Op_LoadD_unaligned:
1971 fpu.inc_double_count();
1972 goto handle_mem;
1973 case Op_StoreF:
1974 case Op_LoadF:
1975 fpu.inc_float_count();
1976 goto handle_mem;
1978 case Op_StoreB:
1979 case Op_StoreC:
1980 case Op_StoreCM:
1981 case Op_StorePConditional:
1982 case Op_StoreI:
1983 case Op_StoreL:
1984 case Op_StoreLConditional:
1985 case Op_CompareAndSwapI:
1986 case Op_CompareAndSwapL:
1987 case Op_CompareAndSwapP:
1988 case Op_CompareAndSwapN:
1989 case Op_StoreP:
1990 case Op_StoreN:
1991 case Op_LoadB:
1992 case Op_LoadC:
1993 case Op_LoadI:
1994 case Op_LoadKlass:
1995 case Op_LoadNKlass:
1996 case Op_LoadL:
1997 case Op_LoadL_unaligned:
1998 case Op_LoadPLocked:
1999 case Op_LoadLLocked:
2000 case Op_LoadP:
2001 case Op_LoadN:
2002 case Op_LoadRange:
2003 case Op_LoadS: {
2004 handle_mem:
2005 #ifdef ASSERT
2006 if( VerifyOptoOopOffsets ) {
2007 assert( n->is_Mem(), "" );
2008 MemNode *mem = (MemNode*)n;
2009 // Check to see if address types have grounded out somehow.
2010 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2011 assert( !tp || oop_offset_is_sane(tp), "" );
2012 }
2013 #endif
2014 break;
2015 }
2017 case Op_AddP: { // Assert sane base pointers
2018 Node *addp = n->in(AddPNode::Address);
2019 assert( !addp->is_AddP() ||
2020 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2021 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2022 "Base pointers must match" );
2023 #ifdef _LP64
2024 if (UseCompressedOops &&
2025 addp->Opcode() == Op_ConP &&
2026 addp == n->in(AddPNode::Base) &&
2027 n->in(AddPNode::Offset)->is_Con()) {
2028 // Use addressing with narrow klass to load with offset on x86.
2029 // On sparc loading 32-bits constant and decoding it have less
2030 // instructions (4) then load 64-bits constant (7).
2031 // Do this transformation here since IGVN will convert ConN back to ConP.
2032 const Type* t = addp->bottom_type();
2033 if (t->isa_oopptr()) {
2034 Node* nn = NULL;
2036 // Look for existing ConN node of the same exact type.
2037 Compile* C = Compile::current();
2038 Node* r = C->root();
2039 uint cnt = r->outcnt();
2040 for (uint i = 0; i < cnt; i++) {
2041 Node* m = r->raw_out(i);
2042 if (m!= NULL && m->Opcode() == Op_ConN &&
2043 m->bottom_type()->make_ptr() == t) {
2044 nn = m;
2045 break;
2046 }
2047 }
2048 if (nn != NULL) {
2049 // Decode a narrow oop to match address
2050 // [R12 + narrow_oop_reg<<3 + offset]
2051 nn = new (C, 2) DecodeNNode(nn, t);
2052 n->set_req(AddPNode::Base, nn);
2053 n->set_req(AddPNode::Address, nn);
2054 if (addp->outcnt() == 0) {
2055 addp->disconnect_inputs(NULL);
2056 }
2057 }
2058 }
2059 }
2060 #endif
2061 break;
2062 }
2064 #ifdef _LP64
2065 case Op_CmpP:
2066 // Do this transformation here to preserve CmpPNode::sub() and
2067 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2068 if( n->in(1)->is_DecodeN() ) {
2069 Compile* C = Compile::current();
2070 Node* in2 = NULL;
2071 if( n->in(2)->is_DecodeN() ) {
2072 in2 = n->in(2)->in(1);
2073 } else if ( n->in(2)->Opcode() == Op_ConP ) {
2074 const Type* t = n->in(2)->bottom_type();
2075 if (t == TypePtr::NULL_PTR) {
2076 Node *in1 = n->in(1);
2077 if (Matcher::clone_shift_expressions) {
2078 // x86, ARM and friends can handle 2 adds in addressing mode.
2079 // Decode a narrow oop and do implicit NULL check in address
2080 // [R12 + narrow_oop_reg<<3 + offset]
2081 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2082 } else {
2083 // Don't replace CmpP(o ,null) if 'o' is used in AddP
2084 // to generate implicit NULL check on Sparc where
2085 // narrow oops can't be used in address.
2086 uint i = 0;
2087 for (; i < in1->outcnt(); i++) {
2088 if (in1->raw_out(i)->is_AddP())
2089 break;
2090 }
2091 if (i >= in1->outcnt()) {
2092 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2093 }
2094 }
2095 } else if (t->isa_oopptr()) {
2096 in2 = ConNode::make(C, t->make_narrowoop());
2097 }
2098 }
2099 if( in2 != NULL ) {
2100 Node* cmpN = new (C, 3) CmpNNode(n->in(1)->in(1), in2);
2101 n->subsume_by( cmpN );
2102 }
2103 }
2104 #endif
2106 case Op_ModI:
2107 if (UseDivMod) {
2108 // Check if a%b and a/b both exist
2109 Node* d = n->find_similar(Op_DivI);
2110 if (d) {
2111 // Replace them with a fused divmod if supported
2112 Compile* C = Compile::current();
2113 if (Matcher::has_match_rule(Op_DivModI)) {
2114 DivModINode* divmod = DivModINode::make(C, n);
2115 d->subsume_by(divmod->div_proj());
2116 n->subsume_by(divmod->mod_proj());
2117 } else {
2118 // replace a%b with a-((a/b)*b)
2119 Node* mult = new (C, 3) MulINode(d, d->in(2));
2120 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2121 n->subsume_by( sub );
2122 }
2123 }
2124 }
2125 break;
2127 case Op_ModL:
2128 if (UseDivMod) {
2129 // Check if a%b and a/b both exist
2130 Node* d = n->find_similar(Op_DivL);
2131 if (d) {
2132 // Replace them with a fused divmod if supported
2133 Compile* C = Compile::current();
2134 if (Matcher::has_match_rule(Op_DivModL)) {
2135 DivModLNode* divmod = DivModLNode::make(C, n);
2136 d->subsume_by(divmod->div_proj());
2137 n->subsume_by(divmod->mod_proj());
2138 } else {
2139 // replace a%b with a-((a/b)*b)
2140 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2141 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2142 n->subsume_by( sub );
2143 }
2144 }
2145 }
2146 break;
2148 case Op_Load16B:
2149 case Op_Load8B:
2150 case Op_Load4B:
2151 case Op_Load8S:
2152 case Op_Load4S:
2153 case Op_Load2S:
2154 case Op_Load8C:
2155 case Op_Load4C:
2156 case Op_Load2C:
2157 case Op_Load4I:
2158 case Op_Load2I:
2159 case Op_Load2L:
2160 case Op_Load4F:
2161 case Op_Load2F:
2162 case Op_Load2D:
2163 case Op_Store16B:
2164 case Op_Store8B:
2165 case Op_Store4B:
2166 case Op_Store8C:
2167 case Op_Store4C:
2168 case Op_Store2C:
2169 case Op_Store4I:
2170 case Op_Store2I:
2171 case Op_Store2L:
2172 case Op_Store4F:
2173 case Op_Store2F:
2174 case Op_Store2D:
2175 break;
2177 case Op_PackB:
2178 case Op_PackS:
2179 case Op_PackC:
2180 case Op_PackI:
2181 case Op_PackF:
2182 case Op_PackL:
2183 case Op_PackD:
2184 if (n->req()-1 > 2) {
2185 // Replace many operand PackNodes with a binary tree for matching
2186 PackNode* p = (PackNode*) n;
2187 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2188 n->subsume_by(btp);
2189 }
2190 break;
2191 default:
2192 assert( !n->is_Call(), "" );
2193 assert( !n->is_Mem(), "" );
2194 break;
2195 }
2197 // Collect CFG split points
2198 if (n->is_MultiBranch())
2199 fpu._tests.push(n);
2200 }
2202 //------------------------------final_graph_reshaping_walk---------------------
2203 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2204 // requires that the walk visits a node's inputs before visiting the node.
2205 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) {
2206 fpu._visited.set(root->_idx); // first, mark node as visited
2207 uint cnt = root->req();
2208 Node *n = root;
2209 uint i = 0;
2210 while (true) {
2211 if (i < cnt) {
2212 // Place all non-visited non-null inputs onto stack
2213 Node* m = n->in(i);
2214 ++i;
2215 if (m != NULL && !fpu._visited.test_set(m->_idx)) {
2216 cnt = m->req();
2217 nstack.push(n, i); // put on stack parent and next input's index
2218 n = m;
2219 i = 0;
2220 }
2221 } else {
2222 // Now do post-visit work
2223 final_graph_reshaping_impl( n, fpu );
2224 if (nstack.is_empty())
2225 break; // finished
2226 n = nstack.node(); // Get node from stack
2227 cnt = n->req();
2228 i = nstack.index();
2229 nstack.pop(); // Shift to the next node on stack
2230 }
2231 }
2232 }
2234 //------------------------------final_graph_reshaping--------------------------
2235 // Final Graph Reshaping.
2236 //
2237 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2238 // and not commoned up and forced early. Must come after regular
2239 // optimizations to avoid GVN undoing the cloning. Clone constant
2240 // inputs to Loop Phis; these will be split by the allocator anyways.
2241 // Remove Opaque nodes.
2242 // (2) Move last-uses by commutative operations to the left input to encourage
2243 // Intel update-in-place two-address operations and better register usage
2244 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2245 // calls canonicalizing them back.
2246 // (3) Count the number of double-precision FP ops, single-precision FP ops
2247 // and call sites. On Intel, we can get correct rounding either by
2248 // forcing singles to memory (requires extra stores and loads after each
2249 // FP bytecode) or we can set a rounding mode bit (requires setting and
2250 // clearing the mode bit around call sites). The mode bit is only used
2251 // if the relative frequency of single FP ops to calls is low enough.
2252 // This is a key transform for SPEC mpeg_audio.
2253 // (4) Detect infinite loops; blobs of code reachable from above but not
2254 // below. Several of the Code_Gen algorithms fail on such code shapes,
2255 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2256 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2257 // Detection is by looking for IfNodes where only 1 projection is
2258 // reachable from below or CatchNodes missing some targets.
2259 // (5) Assert for insane oop offsets in debug mode.
2261 bool Compile::final_graph_reshaping() {
2262 // an infinite loop may have been eliminated by the optimizer,
2263 // in which case the graph will be empty.
2264 if (root()->req() == 1) {
2265 record_method_not_compilable("trivial infinite loop");
2266 return true;
2267 }
2269 Final_Reshape_Counts fpu;
2271 // Visit everybody reachable!
2272 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2273 Node_Stack nstack(unique() >> 1);
2274 final_graph_reshaping_walk(nstack, root(), fpu);
2276 // Check for unreachable (from below) code (i.e., infinite loops).
2277 for( uint i = 0; i < fpu._tests.size(); i++ ) {
2278 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch();
2279 // Get number of CFG targets.
2280 // Note that PCTables include exception targets after calls.
2281 uint required_outcnt = n->required_outcnt();
2282 if (n->outcnt() != required_outcnt) {
2283 // Check for a few special cases. Rethrow Nodes never take the
2284 // 'fall-thru' path, so expected kids is 1 less.
2285 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2286 if (n->in(0)->in(0)->is_Call()) {
2287 CallNode *call = n->in(0)->in(0)->as_Call();
2288 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2289 required_outcnt--; // Rethrow always has 1 less kid
2290 } else if (call->req() > TypeFunc::Parms &&
2291 call->is_CallDynamicJava()) {
2292 // Check for null receiver. In such case, the optimizer has
2293 // detected that the virtual call will always result in a null
2294 // pointer exception. The fall-through projection of this CatchNode
2295 // will not be populated.
2296 Node *arg0 = call->in(TypeFunc::Parms);
2297 if (arg0->is_Type() &&
2298 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2299 required_outcnt--;
2300 }
2301 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2302 call->req() > TypeFunc::Parms+1 &&
2303 call->is_CallStaticJava()) {
2304 // Check for negative array length. In such case, the optimizer has
2305 // detected that the allocation attempt will always result in an
2306 // exception. There is no fall-through projection of this CatchNode .
2307 Node *arg1 = call->in(TypeFunc::Parms+1);
2308 if (arg1->is_Type() &&
2309 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2310 required_outcnt--;
2311 }
2312 }
2313 }
2314 }
2315 // Recheck with a better notion of 'required_outcnt'
2316 if (n->outcnt() != required_outcnt) {
2317 record_method_not_compilable("malformed control flow");
2318 return true; // Not all targets reachable!
2319 }
2320 }
2321 // Check that I actually visited all kids. Unreached kids
2322 // must be infinite loops.
2323 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2324 if (!fpu._visited.test(n->fast_out(j)->_idx)) {
2325 record_method_not_compilable("infinite loop");
2326 return true; // Found unvisited kid; must be unreach
2327 }
2328 }
2330 // If original bytecodes contained a mixture of floats and doubles
2331 // check if the optimizer has made it homogenous, item (3).
2332 if( Use24BitFPMode && Use24BitFP &&
2333 fpu.get_float_count() > 32 &&
2334 fpu.get_double_count() == 0 &&
2335 (10 * fpu.get_call_count() < fpu.get_float_count()) ) {
2336 set_24_bit_selection_and_mode( false, true );
2337 }
2339 set_has_java_calls(fpu.get_java_call_count() > 0);
2341 // No infinite loops, no reason to bail out.
2342 return false;
2343 }
2345 //-----------------------------too_many_traps----------------------------------
2346 // Report if there are too many traps at the current method and bci.
2347 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2348 bool Compile::too_many_traps(ciMethod* method,
2349 int bci,
2350 Deoptimization::DeoptReason reason) {
2351 ciMethodData* md = method->method_data();
2352 if (md->is_empty()) {
2353 // Assume the trap has not occurred, or that it occurred only
2354 // because of a transient condition during start-up in the interpreter.
2355 return false;
2356 }
2357 if (md->has_trap_at(bci, reason) != 0) {
2358 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2359 // Also, if there are multiple reasons, or if there is no per-BCI record,
2360 // assume the worst.
2361 if (log())
2362 log()->elem("observe trap='%s' count='%d'",
2363 Deoptimization::trap_reason_name(reason),
2364 md->trap_count(reason));
2365 return true;
2366 } else {
2367 // Ignore method/bci and see if there have been too many globally.
2368 return too_many_traps(reason, md);
2369 }
2370 }
2372 // Less-accurate variant which does not require a method and bci.
2373 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2374 ciMethodData* logmd) {
2375 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2376 // Too many traps globally.
2377 // Note that we use cumulative trap_count, not just md->trap_count.
2378 if (log()) {
2379 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2380 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2381 Deoptimization::trap_reason_name(reason),
2382 mcount, trap_count(reason));
2383 }
2384 return true;
2385 } else {
2386 // The coast is clear.
2387 return false;
2388 }
2389 }
2391 //--------------------------too_many_recompiles--------------------------------
2392 // Report if there are too many recompiles at the current method and bci.
2393 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2394 // Is not eager to return true, since this will cause the compiler to use
2395 // Action_none for a trap point, to avoid too many recompilations.
2396 bool Compile::too_many_recompiles(ciMethod* method,
2397 int bci,
2398 Deoptimization::DeoptReason reason) {
2399 ciMethodData* md = method->method_data();
2400 if (md->is_empty()) {
2401 // Assume the trap has not occurred, or that it occurred only
2402 // because of a transient condition during start-up in the interpreter.
2403 return false;
2404 }
2405 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2406 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2407 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2408 Deoptimization::DeoptReason per_bc_reason
2409 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2410 if ((per_bc_reason == Deoptimization::Reason_none
2411 || md->has_trap_at(bci, reason) != 0)
2412 // The trap frequency measure we care about is the recompile count:
2413 && md->trap_recompiled_at(bci)
2414 && md->overflow_recompile_count() >= bc_cutoff) {
2415 // Do not emit a trap here if it has already caused recompilations.
2416 // Also, if there are multiple reasons, or if there is no per-BCI record,
2417 // assume the worst.
2418 if (log())
2419 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2420 Deoptimization::trap_reason_name(reason),
2421 md->trap_count(reason),
2422 md->overflow_recompile_count());
2423 return true;
2424 } else if (trap_count(reason) != 0
2425 && decompile_count() >= m_cutoff) {
2426 // Too many recompiles globally, and we have seen this sort of trap.
2427 // Use cumulative decompile_count, not just md->decompile_count.
2428 if (log())
2429 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2430 Deoptimization::trap_reason_name(reason),
2431 md->trap_count(reason), trap_count(reason),
2432 md->decompile_count(), decompile_count());
2433 return true;
2434 } else {
2435 // The coast is clear.
2436 return false;
2437 }
2438 }
2441 #ifndef PRODUCT
2442 //------------------------------verify_graph_edges---------------------------
2443 // Walk the Graph and verify that there is a one-to-one correspondence
2444 // between Use-Def edges and Def-Use edges in the graph.
2445 void Compile::verify_graph_edges(bool no_dead_code) {
2446 if (VerifyGraphEdges) {
2447 ResourceArea *area = Thread::current()->resource_area();
2448 Unique_Node_List visited(area);
2449 // Call recursive graph walk to check edges
2450 _root->verify_edges(visited);
2451 if (no_dead_code) {
2452 // Now make sure that no visited node is used by an unvisited node.
2453 bool dead_nodes = 0;
2454 Unique_Node_List checked(area);
2455 while (visited.size() > 0) {
2456 Node* n = visited.pop();
2457 checked.push(n);
2458 for (uint i = 0; i < n->outcnt(); i++) {
2459 Node* use = n->raw_out(i);
2460 if (checked.member(use)) continue; // already checked
2461 if (visited.member(use)) continue; // already in the graph
2462 if (use->is_Con()) continue; // a dead ConNode is OK
2463 // At this point, we have found a dead node which is DU-reachable.
2464 if (dead_nodes++ == 0)
2465 tty->print_cr("*** Dead nodes reachable via DU edges:");
2466 use->dump(2);
2467 tty->print_cr("---");
2468 checked.push(use); // No repeats; pretend it is now checked.
2469 }
2470 }
2471 assert(dead_nodes == 0, "using nodes must be reachable from root");
2472 }
2473 }
2474 }
2475 #endif
2477 // The Compile object keeps track of failure reasons separately from the ciEnv.
2478 // This is required because there is not quite a 1-1 relation between the
2479 // ciEnv and its compilation task and the Compile object. Note that one
2480 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2481 // to backtrack and retry without subsuming loads. Other than this backtracking
2482 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2483 // by the logic in C2Compiler.
2484 void Compile::record_failure(const char* reason) {
2485 if (log() != NULL) {
2486 log()->elem("failure reason='%s' phase='compile'", reason);
2487 }
2488 if (_failure_reason == NULL) {
2489 // Record the first failure reason.
2490 _failure_reason = reason;
2491 }
2492 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2493 C->print_method(_failure_reason);
2494 }
2495 _root = NULL; // flush the graph, too
2496 }
2498 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2499 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2500 {
2501 if (dolog) {
2502 C = Compile::current();
2503 _log = C->log();
2504 } else {
2505 C = NULL;
2506 _log = NULL;
2507 }
2508 if (_log != NULL) {
2509 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2510 _log->stamp();
2511 _log->end_head();
2512 }
2513 }
2515 Compile::TracePhase::~TracePhase() {
2516 if (_log != NULL) {
2517 _log->done("phase nodes='%d'", C->unique());
2518 }
2519 }