Tue, 02 Sep 2008 15:18:26 -0400
6741004: UseLargePages + UseCompressedOops breaks implicit null checking guard page
Summary: Turn off c2 implicit null checking on windows and large pages specified.
Reviewed-by: jrose, xlu
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);
588 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction.
589 PhaseGVN* igvn = initial_gvn();
590 Node* oop_null = igvn->zerocon(T_OBJECT);
591 Node* noop_null = igvn->zerocon(T_NARROWOOP);
593 _congraph = new(comp_arena()) ConnectionGraph(this);
594 bool has_non_escaping_obj = _congraph->compute_escape();
596 #ifndef PRODUCT
597 if (PrintEscapeAnalysis) {
598 _congraph->dump();
599 }
600 #endif
601 // Cleanup.
602 if (oop_null->outcnt() == 0)
603 igvn->hash_delete(oop_null);
604 if (noop_null->outcnt() == 0)
605 igvn->hash_delete(noop_null);
607 if (!has_non_escaping_obj) {
608 _congraph = NULL;
609 }
611 if (failing()) return;
612 }
613 // Now optimize
614 Optimize();
615 if (failing()) return;
616 NOT_PRODUCT( verify_graph_edges(); )
618 print_method("Before Matching");
620 #ifndef PRODUCT
621 if (PrintIdeal) {
622 ttyLocker ttyl; // keep the following output all in one block
623 // This output goes directly to the tty, not the compiler log.
624 // To enable tools to match it up with the compilation activity,
625 // be sure to tag this tty output with the compile ID.
626 if (xtty != NULL) {
627 xtty->head("ideal compile_id='%d'%s", compile_id(),
628 is_osr_compilation() ? " compile_kind='osr'" :
629 "");
630 }
631 root()->dump(9999);
632 if (xtty != NULL) {
633 xtty->tail("ideal");
634 }
635 }
636 #endif
638 // Now that we know the size of all the monitors we can add a fixed slot
639 // for the original deopt pc.
641 _orig_pc_slot = fixed_slots();
642 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
643 set_fixed_slots(next_slot);
645 // Now generate code
646 Code_Gen();
647 if (failing()) return;
649 // Check if we want to skip execution of all compiled code.
650 {
651 #ifndef PRODUCT
652 if (OptoNoExecute) {
653 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
654 return;
655 }
656 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
657 #endif
659 if (is_osr_compilation()) {
660 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
661 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
662 } else {
663 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
664 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
665 }
667 env()->register_method(_method, _entry_bci,
668 &_code_offsets,
669 _orig_pc_slot_offset_in_bytes,
670 code_buffer(),
671 frame_size_in_words(), _oop_map_set,
672 &_handler_table, &_inc_table,
673 compiler,
674 env()->comp_level(),
675 true, /*has_debug_info*/
676 has_unsafe_access()
677 );
678 }
679 }
681 //------------------------------Compile----------------------------------------
682 // Compile a runtime stub
683 Compile::Compile( ciEnv* ci_env,
684 TypeFunc_generator generator,
685 address stub_function,
686 const char *stub_name,
687 int is_fancy_jump,
688 bool pass_tls,
689 bool save_arg_registers,
690 bool return_pc )
691 : Phase(Compiler),
692 _env(ci_env),
693 _log(ci_env->log()),
694 _compile_id(-1),
695 _save_argument_registers(save_arg_registers),
696 _method(NULL),
697 _stub_name(stub_name),
698 _stub_function(stub_function),
699 _stub_entry_point(NULL),
700 _entry_bci(InvocationEntryBci),
701 _initial_gvn(NULL),
702 _for_igvn(NULL),
703 _warm_calls(NULL),
704 _orig_pc_slot(0),
705 _orig_pc_slot_offset_in_bytes(0),
706 _subsume_loads(true),
707 _do_escape_analysis(false),
708 _failure_reason(NULL),
709 _code_buffer("Compile::Fill_buffer"),
710 _node_bundling_limit(0),
711 _node_bundling_base(NULL),
712 #ifndef PRODUCT
713 _trace_opto_output(TraceOptoOutput),
714 _printer(NULL),
715 #endif
716 _congraph(NULL) {
717 C = this;
719 #ifndef PRODUCT
720 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
721 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
722 set_print_assembly(PrintFrameConverterAssembly);
723 #endif
724 CompileWrapper cw(this);
725 Init(/*AliasLevel=*/ 0);
726 init_tf((*generator)());
728 {
729 // The following is a dummy for the sake of GraphKit::gen_stub
730 Unique_Node_List for_igvn(comp_arena());
731 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
732 PhaseGVN gvn(Thread::current()->resource_area(),255);
733 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
734 gvn.transform_no_reclaim(top());
736 GraphKit kit;
737 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
738 }
740 NOT_PRODUCT( verify_graph_edges(); )
741 Code_Gen();
742 if (failing()) return;
745 // Entry point will be accessed using compile->stub_entry_point();
746 if (code_buffer() == NULL) {
747 Matcher::soft_match_failure();
748 } else {
749 if (PrintAssembly && (WizardMode || Verbose))
750 tty->print_cr("### Stub::%s", stub_name);
752 if (!failing()) {
753 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
755 // Make the NMethod
756 // For now we mark the frame as never safe for profile stackwalking
757 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
758 code_buffer(),
759 CodeOffsets::frame_never_safe,
760 // _code_offsets.value(CodeOffsets::Frame_Complete),
761 frame_size_in_words(),
762 _oop_map_set,
763 save_arg_registers);
764 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
766 _stub_entry_point = rs->entry_point();
767 }
768 }
769 }
771 #ifndef PRODUCT
772 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
773 if(PrintOpto && Verbose) {
774 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
775 }
776 }
777 #endif
779 void Compile::print_codes() {
780 }
782 //------------------------------Init-------------------------------------------
783 // Prepare for a single compilation
784 void Compile::Init(int aliaslevel) {
785 _unique = 0;
786 _regalloc = NULL;
788 _tf = NULL; // filled in later
789 _top = NULL; // cached later
790 _matcher = NULL; // filled in later
791 _cfg = NULL; // filled in later
793 set_24_bit_selection_and_mode(Use24BitFP, false);
795 _node_note_array = NULL;
796 _default_node_notes = NULL;
798 _immutable_memory = NULL; // filled in at first inquiry
800 // Globally visible Nodes
801 // First set TOP to NULL to give safe behavior during creation of RootNode
802 set_cached_top_node(NULL);
803 set_root(new (this, 3) RootNode());
804 // Now that you have a Root to point to, create the real TOP
805 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
806 set_recent_alloc(NULL, NULL);
808 // Create Debug Information Recorder to record scopes, oopmaps, etc.
809 env()->set_oop_recorder(new OopRecorder(comp_arena()));
810 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
811 env()->set_dependencies(new Dependencies(env()));
813 _fixed_slots = 0;
814 set_has_split_ifs(false);
815 set_has_loops(has_method() && method()->has_loops()); // first approximation
816 _deopt_happens = true; // start out assuming the worst
817 _trap_can_recompile = false; // no traps emitted yet
818 _major_progress = true; // start out assuming good things will happen
819 set_has_unsafe_access(false);
820 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
821 set_decompile_count(0);
823 // Compilation level related initialization
824 if (env()->comp_level() == CompLevel_fast_compile) {
825 set_num_loop_opts(Tier1LoopOptsCount);
826 set_do_inlining(Tier1Inline != 0);
827 set_max_inline_size(Tier1MaxInlineSize);
828 set_freq_inline_size(Tier1FreqInlineSize);
829 set_do_scheduling(false);
830 set_do_count_invocations(Tier1CountInvocations);
831 set_do_method_data_update(Tier1UpdateMethodData);
832 } else {
833 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
834 set_num_loop_opts(LoopOptsCount);
835 set_do_inlining(Inline);
836 set_max_inline_size(MaxInlineSize);
837 set_freq_inline_size(FreqInlineSize);
838 set_do_scheduling(OptoScheduling);
839 set_do_count_invocations(false);
840 set_do_method_data_update(false);
841 }
843 if (debug_info()->recording_non_safepoints()) {
844 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
845 (comp_arena(), 8, 0, NULL));
846 set_default_node_notes(Node_Notes::make(this));
847 }
849 // // -- Initialize types before each compile --
850 // // Update cached type information
851 // if( _method && _method->constants() )
852 // Type::update_loaded_types(_method, _method->constants());
854 // Init alias_type map.
855 if (!_do_escape_analysis && aliaslevel == 3)
856 aliaslevel = 2; // No unique types without escape analysis
857 _AliasLevel = aliaslevel;
858 const int grow_ats = 16;
859 _max_alias_types = grow_ats;
860 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
861 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
862 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
863 {
864 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
865 }
866 // Initialize the first few types.
867 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
868 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
869 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
870 _num_alias_types = AliasIdxRaw+1;
871 // Zero out the alias type cache.
872 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
873 // A NULL adr_type hits in the cache right away. Preload the right answer.
874 probe_alias_cache(NULL)->_index = AliasIdxTop;
876 _intrinsics = NULL;
877 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
878 register_library_intrinsics();
879 }
881 //---------------------------init_start----------------------------------------
882 // Install the StartNode on this compile object.
883 void Compile::init_start(StartNode* s) {
884 if (failing())
885 return; // already failing
886 assert(s == start(), "");
887 }
889 StartNode* Compile::start() const {
890 assert(!failing(), "");
891 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
892 Node* start = root()->fast_out(i);
893 if( start->is_Start() )
894 return start->as_Start();
895 }
896 ShouldNotReachHere();
897 return NULL;
898 }
900 //-------------------------------immutable_memory-------------------------------------
901 // Access immutable memory
902 Node* Compile::immutable_memory() {
903 if (_immutable_memory != NULL) {
904 return _immutable_memory;
905 }
906 StartNode* s = start();
907 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
908 Node *p = s->fast_out(i);
909 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
910 _immutable_memory = p;
911 return _immutable_memory;
912 }
913 }
914 ShouldNotReachHere();
915 return NULL;
916 }
918 //----------------------set_cached_top_node------------------------------------
919 // Install the cached top node, and make sure Node::is_top works correctly.
920 void Compile::set_cached_top_node(Node* tn) {
921 if (tn != NULL) verify_top(tn);
922 Node* old_top = _top;
923 _top = tn;
924 // Calling Node::setup_is_top allows the nodes the chance to adjust
925 // their _out arrays.
926 if (_top != NULL) _top->setup_is_top();
927 if (old_top != NULL) old_top->setup_is_top();
928 assert(_top == NULL || top()->is_top(), "");
929 }
931 #ifndef PRODUCT
932 void Compile::verify_top(Node* tn) const {
933 if (tn != NULL) {
934 assert(tn->is_Con(), "top node must be a constant");
935 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
936 assert(tn->in(0) != NULL, "must have live top node");
937 }
938 }
939 #endif
942 ///-------------------Managing Per-Node Debug & Profile Info-------------------
944 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
945 guarantee(arr != NULL, "");
946 int num_blocks = arr->length();
947 if (grow_by < num_blocks) grow_by = num_blocks;
948 int num_notes = grow_by * _node_notes_block_size;
949 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
950 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
951 while (num_notes > 0) {
952 arr->append(notes);
953 notes += _node_notes_block_size;
954 num_notes -= _node_notes_block_size;
955 }
956 assert(num_notes == 0, "exact multiple, please");
957 }
959 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
960 if (source == NULL || dest == NULL) return false;
962 if (dest->is_Con())
963 return false; // Do not push debug info onto constants.
965 #ifdef ASSERT
966 // Leave a bread crumb trail pointing to the original node:
967 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
968 dest->set_debug_orig(source);
969 }
970 #endif
972 if (node_note_array() == NULL)
973 return false; // Not collecting any notes now.
975 // This is a copy onto a pre-existing node, which may already have notes.
976 // If both nodes have notes, do not overwrite any pre-existing notes.
977 Node_Notes* source_notes = node_notes_at(source->_idx);
978 if (source_notes == NULL || source_notes->is_clear()) return false;
979 Node_Notes* dest_notes = node_notes_at(dest->_idx);
980 if (dest_notes == NULL || dest_notes->is_clear()) {
981 return set_node_notes_at(dest->_idx, source_notes);
982 }
984 Node_Notes merged_notes = (*source_notes);
985 // The order of operations here ensures that dest notes will win...
986 merged_notes.update_from(dest_notes);
987 return set_node_notes_at(dest->_idx, &merged_notes);
988 }
991 //--------------------------allow_range_check_smearing-------------------------
992 // Gating condition for coalescing similar range checks.
993 // Sometimes we try 'speculatively' replacing a series of a range checks by a
994 // single covering check that is at least as strong as any of them.
995 // If the optimization succeeds, the simplified (strengthened) range check
996 // will always succeed. If it fails, we will deopt, and then give up
997 // on the optimization.
998 bool Compile::allow_range_check_smearing() const {
999 // If this method has already thrown a range-check,
1000 // assume it was because we already tried range smearing
1001 // and it failed.
1002 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1003 return !already_trapped;
1004 }
1007 //------------------------------flatten_alias_type-----------------------------
1008 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1009 int offset = tj->offset();
1010 TypePtr::PTR ptr = tj->ptr();
1012 // Known instance (scalarizable allocation) alias only with itself.
1013 bool is_known_inst = tj->isa_oopptr() != NULL &&
1014 tj->is_oopptr()->is_known_instance();
1016 // Process weird unsafe references.
1017 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1018 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1019 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1020 tj = TypeOopPtr::BOTTOM;
1021 ptr = tj->ptr();
1022 offset = tj->offset();
1023 }
1025 // Array pointers need some flattening
1026 const TypeAryPtr *ta = tj->isa_aryptr();
1027 if( ta && is_known_inst ) {
1028 if ( offset != Type::OffsetBot &&
1029 offset > arrayOopDesc::length_offset_in_bytes() ) {
1030 offset = Type::OffsetBot; // Flatten constant access into array body only
1031 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1032 }
1033 } else if( ta && _AliasLevel >= 2 ) {
1034 // For arrays indexed by constant indices, we flatten the alias
1035 // space to include all of the array body. Only the header, klass
1036 // and array length can be accessed un-aliased.
1037 if( offset != Type::OffsetBot ) {
1038 if( ta->const_oop() ) { // methodDataOop or methodOop
1039 offset = Type::OffsetBot; // Flatten constant access into array body
1040 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1041 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1042 // range is OK as-is.
1043 tj = ta = TypeAryPtr::RANGE;
1044 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1045 tj = TypeInstPtr::KLASS; // all klass loads look alike
1046 ta = TypeAryPtr::RANGE; // generic ignored junk
1047 ptr = TypePtr::BotPTR;
1048 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1049 tj = TypeInstPtr::MARK;
1050 ta = TypeAryPtr::RANGE; // generic ignored junk
1051 ptr = TypePtr::BotPTR;
1052 } else { // Random constant offset into array body
1053 offset = Type::OffsetBot; // Flatten constant access into array body
1054 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1055 }
1056 }
1057 // Arrays of fixed size alias with arrays of unknown size.
1058 if (ta->size() != TypeInt::POS) {
1059 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1060 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1061 }
1062 // Arrays of known objects become arrays of unknown objects.
1063 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1064 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1065 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1066 }
1067 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1068 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1069 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1070 }
1071 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1072 // cannot be distinguished by bytecode alone.
1073 if (ta->elem() == TypeInt::BOOL) {
1074 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1075 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1076 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1077 }
1078 // During the 2nd round of IterGVN, NotNull castings are removed.
1079 // Make sure the Bottom and NotNull variants alias the same.
1080 // Also, make sure exact and non-exact variants alias the same.
1081 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1082 if (ta->const_oop()) {
1083 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1084 } else {
1085 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1086 }
1087 }
1088 }
1090 // Oop pointers need some flattening
1091 const TypeInstPtr *to = tj->isa_instptr();
1092 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1093 if( ptr == TypePtr::Constant ) {
1094 // No constant oop pointers (such as Strings); they alias with
1095 // unknown strings.
1096 assert(!is_known_inst, "not scalarizable allocation");
1097 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1098 } else if( is_known_inst ) {
1099 tj = to; // Keep NotNull and klass_is_exact for instance type
1100 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1101 // During the 2nd round of IterGVN, NotNull castings are removed.
1102 // Make sure the Bottom and NotNull variants alias the same.
1103 // Also, make sure exact and non-exact variants alias the same.
1104 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1105 }
1106 // Canonicalize the holder of this field
1107 ciInstanceKlass *k = to->klass()->as_instance_klass();
1108 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1109 // First handle header references such as a LoadKlassNode, even if the
1110 // object's klass is unloaded at compile time (4965979).
1111 if (!is_known_inst) { // Do it only for non-instance types
1112 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1113 }
1114 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1115 to = NULL;
1116 tj = TypeOopPtr::BOTTOM;
1117 offset = tj->offset();
1118 } else {
1119 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1120 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1121 if( is_known_inst ) {
1122 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1123 } else {
1124 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1125 }
1126 }
1127 }
1128 }
1130 // Klass pointers to object array klasses need some flattening
1131 const TypeKlassPtr *tk = tj->isa_klassptr();
1132 if( tk ) {
1133 // If we are referencing a field within a Klass, we need
1134 // to assume the worst case of an Object. Both exact and
1135 // inexact types must flatten to the same alias class.
1136 // Since the flattened result for a klass is defined to be
1137 // precisely java.lang.Object, use a constant ptr.
1138 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1140 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1141 TypeKlassPtr::OBJECT->klass(),
1142 offset);
1143 }
1145 ciKlass* klass = tk->klass();
1146 if( klass->is_obj_array_klass() ) {
1147 ciKlass* k = TypeAryPtr::OOPS->klass();
1148 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1149 k = TypeInstPtr::BOTTOM->klass();
1150 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1151 }
1153 // Check for precise loads from the primary supertype array and force them
1154 // to the supertype cache alias index. Check for generic array loads from
1155 // the primary supertype array and also force them to the supertype cache
1156 // alias index. Since the same load can reach both, we need to merge
1157 // these 2 disparate memories into the same alias class. Since the
1158 // primary supertype array is read-only, there's no chance of confusion
1159 // where we bypass an array load and an array store.
1160 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1161 if( offset == Type::OffsetBot ||
1162 off2 < Klass::primary_super_limit()*wordSize ) {
1163 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1164 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1165 }
1166 }
1168 // Flatten all Raw pointers together.
1169 if (tj->base() == Type::RawPtr)
1170 tj = TypeRawPtr::BOTTOM;
1172 if (tj->base() == Type::AnyPtr)
1173 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1175 // Flatten all to bottom for now
1176 switch( _AliasLevel ) {
1177 case 0:
1178 tj = TypePtr::BOTTOM;
1179 break;
1180 case 1: // Flatten to: oop, static, field or array
1181 switch (tj->base()) {
1182 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1183 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1184 case Type::AryPtr: // do not distinguish arrays at all
1185 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1186 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1187 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1188 default: ShouldNotReachHere();
1189 }
1190 break;
1191 case 2: // No collasping at level 2; keep all splits
1192 case 3: // No collasping at level 3; keep all splits
1193 break;
1194 default:
1195 Unimplemented();
1196 }
1198 offset = tj->offset();
1199 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1201 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1202 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1203 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1204 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1205 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1206 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1207 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1208 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1209 assert( tj->ptr() != TypePtr::TopPTR &&
1210 tj->ptr() != TypePtr::AnyNull &&
1211 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1212 // assert( tj->ptr() != TypePtr::Constant ||
1213 // tj->base() == Type::RawPtr ||
1214 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1216 return tj;
1217 }
1219 void Compile::AliasType::Init(int i, const TypePtr* at) {
1220 _index = i;
1221 _adr_type = at;
1222 _field = NULL;
1223 _is_rewritable = true; // default
1224 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1225 if (atoop != NULL && atoop->is_known_instance()) {
1226 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1227 _general_index = Compile::current()->get_alias_index(gt);
1228 } else {
1229 _general_index = 0;
1230 }
1231 }
1233 //---------------------------------print_on------------------------------------
1234 #ifndef PRODUCT
1235 void Compile::AliasType::print_on(outputStream* st) {
1236 if (index() < 10)
1237 st->print("@ <%d> ", index());
1238 else st->print("@ <%d>", index());
1239 st->print(is_rewritable() ? " " : " RO");
1240 int offset = adr_type()->offset();
1241 if (offset == Type::OffsetBot)
1242 st->print(" +any");
1243 else st->print(" +%-3d", offset);
1244 st->print(" in ");
1245 adr_type()->dump_on(st);
1246 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1247 if (field() != NULL && tjp) {
1248 if (tjp->klass() != field()->holder() ||
1249 tjp->offset() != field()->offset_in_bytes()) {
1250 st->print(" != ");
1251 field()->print();
1252 st->print(" ***");
1253 }
1254 }
1255 }
1257 void print_alias_types() {
1258 Compile* C = Compile::current();
1259 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1260 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1261 C->alias_type(idx)->print_on(tty);
1262 tty->cr();
1263 }
1264 }
1265 #endif
1268 //----------------------------probe_alias_cache--------------------------------
1269 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1270 intptr_t key = (intptr_t) adr_type;
1271 key ^= key >> logAliasCacheSize;
1272 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1273 }
1276 //-----------------------------grow_alias_types--------------------------------
1277 void Compile::grow_alias_types() {
1278 const int old_ats = _max_alias_types; // how many before?
1279 const int new_ats = old_ats; // how many more?
1280 const int grow_ats = old_ats+new_ats; // how many now?
1281 _max_alias_types = grow_ats;
1282 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1283 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1284 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1285 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1286 }
1289 //--------------------------------find_alias_type------------------------------
1290 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1291 if (_AliasLevel == 0)
1292 return alias_type(AliasIdxBot);
1294 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1295 if (ace->_adr_type == adr_type) {
1296 return alias_type(ace->_index);
1297 }
1299 // Handle special cases.
1300 if (adr_type == NULL) return alias_type(AliasIdxTop);
1301 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1303 // Do it the slow way.
1304 const TypePtr* flat = flatten_alias_type(adr_type);
1306 #ifdef ASSERT
1307 assert(flat == flatten_alias_type(flat), "idempotent");
1308 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1309 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1310 const TypeOopPtr* foop = flat->is_oopptr();
1311 // Scalarizable allocations have exact klass always.
1312 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1313 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1314 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1315 }
1316 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1317 #endif
1319 int idx = AliasIdxTop;
1320 for (int i = 0; i < num_alias_types(); i++) {
1321 if (alias_type(i)->adr_type() == flat) {
1322 idx = i;
1323 break;
1324 }
1325 }
1327 if (idx == AliasIdxTop) {
1328 if (no_create) return NULL;
1329 // Grow the array if necessary.
1330 if (_num_alias_types == _max_alias_types) grow_alias_types();
1331 // Add a new alias type.
1332 idx = _num_alias_types++;
1333 _alias_types[idx]->Init(idx, flat);
1334 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1335 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1336 if (flat->isa_instptr()) {
1337 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1338 && flat->is_instptr()->klass() == env()->Class_klass())
1339 alias_type(idx)->set_rewritable(false);
1340 }
1341 if (flat->isa_klassptr()) {
1342 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1343 alias_type(idx)->set_rewritable(false);
1344 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1345 alias_type(idx)->set_rewritable(false);
1346 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1347 alias_type(idx)->set_rewritable(false);
1348 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1349 alias_type(idx)->set_rewritable(false);
1350 }
1351 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1352 // but the base pointer type is not distinctive enough to identify
1353 // references into JavaThread.)
1355 // Check for final instance fields.
1356 const TypeInstPtr* tinst = flat->isa_instptr();
1357 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1358 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1359 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1360 // Set field() and is_rewritable() attributes.
1361 if (field != NULL) alias_type(idx)->set_field(field);
1362 }
1363 const TypeKlassPtr* tklass = flat->isa_klassptr();
1364 // Check for final static fields.
1365 if (tklass && tklass->klass()->is_instance_klass()) {
1366 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1367 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1368 // Set field() and is_rewritable() attributes.
1369 if (field != NULL) alias_type(idx)->set_field(field);
1370 }
1371 }
1373 // Fill the cache for next time.
1374 ace->_adr_type = adr_type;
1375 ace->_index = idx;
1376 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1378 // Might as well try to fill the cache for the flattened version, too.
1379 AliasCacheEntry* face = probe_alias_cache(flat);
1380 if (face->_adr_type == NULL) {
1381 face->_adr_type = flat;
1382 face->_index = idx;
1383 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1384 }
1386 return alias_type(idx);
1387 }
1390 Compile::AliasType* Compile::alias_type(ciField* field) {
1391 const TypeOopPtr* t;
1392 if (field->is_static())
1393 t = TypeKlassPtr::make(field->holder());
1394 else
1395 t = TypeOopPtr::make_from_klass_raw(field->holder());
1396 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1397 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1398 return atp;
1399 }
1402 //------------------------------have_alias_type--------------------------------
1403 bool Compile::have_alias_type(const TypePtr* adr_type) {
1404 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1405 if (ace->_adr_type == adr_type) {
1406 return true;
1407 }
1409 // Handle special cases.
1410 if (adr_type == NULL) return true;
1411 if (adr_type == TypePtr::BOTTOM) return true;
1413 return find_alias_type(adr_type, true) != NULL;
1414 }
1416 //-----------------------------must_alias--------------------------------------
1417 // True if all values of the given address type are in the given alias category.
1418 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1419 if (alias_idx == AliasIdxBot) return true; // the universal category
1420 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1421 if (alias_idx == AliasIdxTop) return false; // the empty category
1422 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1424 // the only remaining possible overlap is identity
1425 int adr_idx = get_alias_index(adr_type);
1426 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1427 assert(adr_idx == alias_idx ||
1428 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1429 && adr_type != TypeOopPtr::BOTTOM),
1430 "should not be testing for overlap with an unsafe pointer");
1431 return adr_idx == alias_idx;
1432 }
1434 //------------------------------can_alias--------------------------------------
1435 // True if any values of the given address type are in the given alias category.
1436 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1437 if (alias_idx == AliasIdxTop) return false; // the empty category
1438 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1439 if (alias_idx == AliasIdxBot) return true; // the universal category
1440 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1442 // the only remaining possible overlap is identity
1443 int adr_idx = get_alias_index(adr_type);
1444 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1445 return adr_idx == alias_idx;
1446 }
1450 //---------------------------pop_warm_call-------------------------------------
1451 WarmCallInfo* Compile::pop_warm_call() {
1452 WarmCallInfo* wci = _warm_calls;
1453 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1454 return wci;
1455 }
1457 //----------------------------Inline_Warm--------------------------------------
1458 int Compile::Inline_Warm() {
1459 // If there is room, try to inline some more warm call sites.
1460 // %%% Do a graph index compaction pass when we think we're out of space?
1461 if (!InlineWarmCalls) return 0;
1463 int calls_made_hot = 0;
1464 int room_to_grow = NodeCountInliningCutoff - unique();
1465 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1466 int amount_grown = 0;
1467 WarmCallInfo* call;
1468 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1469 int est_size = (int)call->size();
1470 if (est_size > (room_to_grow - amount_grown)) {
1471 // This one won't fit anyway. Get rid of it.
1472 call->make_cold();
1473 continue;
1474 }
1475 call->make_hot();
1476 calls_made_hot++;
1477 amount_grown += est_size;
1478 amount_to_grow -= est_size;
1479 }
1481 if (calls_made_hot > 0) set_major_progress();
1482 return calls_made_hot;
1483 }
1486 //----------------------------Finish_Warm--------------------------------------
1487 void Compile::Finish_Warm() {
1488 if (!InlineWarmCalls) return;
1489 if (failing()) return;
1490 if (warm_calls() == NULL) return;
1492 // Clean up loose ends, if we are out of space for inlining.
1493 WarmCallInfo* call;
1494 while ((call = pop_warm_call()) != NULL) {
1495 call->make_cold();
1496 }
1497 }
1500 //------------------------------Optimize---------------------------------------
1501 // Given a graph, optimize it.
1502 void Compile::Optimize() {
1503 TracePhase t1("optimizer", &_t_optimizer, true);
1505 #ifndef PRODUCT
1506 if (env()->break_at_compile()) {
1507 BREAKPOINT;
1508 }
1510 #endif
1512 ResourceMark rm;
1513 int loop_opts_cnt;
1515 NOT_PRODUCT( verify_graph_edges(); )
1517 print_method("After Parsing");
1519 {
1520 // Iterative Global Value Numbering, including ideal transforms
1521 // Initialize IterGVN with types and values from parse-time GVN
1522 PhaseIterGVN igvn(initial_gvn());
1523 {
1524 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1525 igvn.optimize();
1526 }
1528 print_method("Iter GVN 1", 2);
1530 if (failing()) return;
1532 // get rid of the connection graph since it's information is not
1533 // updated by optimizations
1534 _congraph = NULL;
1537 // Loop transforms on the ideal graph. Range Check Elimination,
1538 // peeling, unrolling, etc.
1540 // Set loop opts counter
1541 loop_opts_cnt = num_loop_opts();
1542 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1543 {
1544 TracePhase t2("idealLoop", &_t_idealLoop, true);
1545 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1546 loop_opts_cnt--;
1547 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1548 if (failing()) return;
1549 }
1550 // Loop opts pass if partial peeling occurred in previous pass
1551 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1552 TracePhase t3("idealLoop", &_t_idealLoop, true);
1553 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1554 loop_opts_cnt--;
1555 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1556 if (failing()) return;
1557 }
1558 // Loop opts pass for loop-unrolling before CCP
1559 if(major_progress() && (loop_opts_cnt > 0)) {
1560 TracePhase t4("idealLoop", &_t_idealLoop, true);
1561 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1562 loop_opts_cnt--;
1563 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1564 }
1565 }
1566 if (failing()) return;
1568 // Conditional Constant Propagation;
1569 PhaseCCP ccp( &igvn );
1570 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1571 {
1572 TracePhase t2("ccp", &_t_ccp, true);
1573 ccp.do_transform();
1574 }
1575 print_method("PhaseCPP 1", 2);
1577 assert( true, "Break here to ccp.dump_old2new_map()");
1579 // Iterative Global Value Numbering, including ideal transforms
1580 {
1581 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1582 igvn = ccp;
1583 igvn.optimize();
1584 }
1586 print_method("Iter GVN 2", 2);
1588 if (failing()) return;
1590 // Loop transforms on the ideal graph. Range Check Elimination,
1591 // peeling, unrolling, etc.
1592 if(loop_opts_cnt > 0) {
1593 debug_only( int cnt = 0; );
1594 while(major_progress() && (loop_opts_cnt > 0)) {
1595 TracePhase t2("idealLoop", &_t_idealLoop, true);
1596 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1597 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1598 loop_opts_cnt--;
1599 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1600 if (failing()) return;
1601 }
1602 }
1603 {
1604 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1605 PhaseMacroExpand mex(igvn);
1606 if (mex.expand_macro_nodes()) {
1607 assert(failing(), "must bail out w/ explicit message");
1608 return;
1609 }
1610 }
1612 } // (End scope of igvn; run destructor if necessary for asserts.)
1614 // A method with only infinite loops has no edges entering loops from root
1615 {
1616 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1617 if (final_graph_reshaping()) {
1618 assert(failing(), "must bail out w/ explicit message");
1619 return;
1620 }
1621 }
1623 print_method("Optimize finished", 2);
1624 }
1627 //------------------------------Code_Gen---------------------------------------
1628 // Given a graph, generate code for it
1629 void Compile::Code_Gen() {
1630 if (failing()) return;
1632 // Perform instruction selection. You might think we could reclaim Matcher
1633 // memory PDQ, but actually the Matcher is used in generating spill code.
1634 // Internals of the Matcher (including some VectorSets) must remain live
1635 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1636 // set a bit in reclaimed memory.
1638 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1639 // nodes. Mapping is only valid at the root of each matched subtree.
1640 NOT_PRODUCT( verify_graph_edges(); )
1642 Node_List proj_list;
1643 Matcher m(proj_list);
1644 _matcher = &m;
1645 {
1646 TracePhase t2("matcher", &_t_matcher, true);
1647 m.match();
1648 }
1649 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1650 // nodes. Mapping is only valid at the root of each matched subtree.
1651 NOT_PRODUCT( verify_graph_edges(); )
1653 // If you have too many nodes, or if matching has failed, bail out
1654 check_node_count(0, "out of nodes matching instructions");
1655 if (failing()) return;
1657 // Build a proper-looking CFG
1658 PhaseCFG cfg(node_arena(), root(), m);
1659 _cfg = &cfg;
1660 {
1661 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1662 cfg.Dominators();
1663 if (failing()) return;
1665 NOT_PRODUCT( verify_graph_edges(); )
1667 cfg.Estimate_Block_Frequency();
1668 cfg.GlobalCodeMotion(m,unique(),proj_list);
1670 print_method("Global code motion", 2);
1672 if (failing()) return;
1673 NOT_PRODUCT( verify_graph_edges(); )
1675 debug_only( cfg.verify(); )
1676 }
1677 NOT_PRODUCT( verify_graph_edges(); )
1679 PhaseChaitin regalloc(unique(),cfg,m);
1680 _regalloc = ®alloc;
1681 {
1682 TracePhase t2("regalloc", &_t_registerAllocation, true);
1683 // Perform any platform dependent preallocation actions. This is used,
1684 // for example, to avoid taking an implicit null pointer exception
1685 // using the frame pointer on win95.
1686 _regalloc->pd_preallocate_hook();
1688 // Perform register allocation. After Chaitin, use-def chains are
1689 // no longer accurate (at spill code) and so must be ignored.
1690 // Node->LRG->reg mappings are still accurate.
1691 _regalloc->Register_Allocate();
1693 // Bail out if the allocator builds too many nodes
1694 if (failing()) return;
1695 }
1697 // Prior to register allocation we kept empty basic blocks in case the
1698 // the allocator needed a place to spill. After register allocation we
1699 // are not adding any new instructions. If any basic block is empty, we
1700 // can now safely remove it.
1701 {
1702 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); )
1703 cfg.RemoveEmpty();
1704 }
1706 // Perform any platform dependent postallocation verifications.
1707 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1709 // Apply peephole optimizations
1710 if( OptoPeephole ) {
1711 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1712 PhasePeephole peep( _regalloc, cfg);
1713 peep.do_transform();
1714 }
1716 // Convert Nodes to instruction bits in a buffer
1717 {
1718 // %%%% workspace merge brought two timers together for one job
1719 TracePhase t2a("output", &_t_output, true);
1720 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1721 Output();
1722 }
1724 print_method("Final Code");
1726 // He's dead, Jim.
1727 _cfg = (PhaseCFG*)0xdeadbeef;
1728 _regalloc = (PhaseChaitin*)0xdeadbeef;
1729 }
1732 //------------------------------dump_asm---------------------------------------
1733 // Dump formatted assembly
1734 #ifndef PRODUCT
1735 void Compile::dump_asm(int *pcs, uint pc_limit) {
1736 bool cut_short = false;
1737 tty->print_cr("#");
1738 tty->print("# "); _tf->dump(); tty->cr();
1739 tty->print_cr("#");
1741 // For all blocks
1742 int pc = 0x0; // Program counter
1743 char starts_bundle = ' ';
1744 _regalloc->dump_frame();
1746 Node *n = NULL;
1747 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1748 if (VMThread::should_terminate()) { cut_short = true; break; }
1749 Block *b = _cfg->_blocks[i];
1750 if (b->is_connector() && !Verbose) continue;
1751 n = b->_nodes[0];
1752 if (pcs && n->_idx < pc_limit)
1753 tty->print("%3.3x ", pcs[n->_idx]);
1754 else
1755 tty->print(" ");
1756 b->dump_head( &_cfg->_bbs );
1757 if (b->is_connector()) {
1758 tty->print_cr(" # Empty connector block");
1759 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1760 tty->print_cr(" # Block is sole successor of call");
1761 }
1763 // For all instructions
1764 Node *delay = NULL;
1765 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1766 if (VMThread::should_terminate()) { cut_short = true; break; }
1767 n = b->_nodes[j];
1768 if (valid_bundle_info(n)) {
1769 Bundle *bundle = node_bundling(n);
1770 if (bundle->used_in_unconditional_delay()) {
1771 delay = n;
1772 continue;
1773 }
1774 if (bundle->starts_bundle())
1775 starts_bundle = '+';
1776 }
1778 if (WizardMode) n->dump();
1780 if( !n->is_Region() && // Dont print in the Assembly
1781 !n->is_Phi() && // a few noisely useless nodes
1782 !n->is_Proj() &&
1783 !n->is_MachTemp() &&
1784 !n->is_Catch() && // Would be nice to print exception table targets
1785 !n->is_MergeMem() && // Not very interesting
1786 !n->is_top() && // Debug info table constants
1787 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1788 ) {
1789 if (pcs && n->_idx < pc_limit)
1790 tty->print("%3.3x", pcs[n->_idx]);
1791 else
1792 tty->print(" ");
1793 tty->print(" %c ", starts_bundle);
1794 starts_bundle = ' ';
1795 tty->print("\t");
1796 n->format(_regalloc, tty);
1797 tty->cr();
1798 }
1800 // If we have an instruction with a delay slot, and have seen a delay,
1801 // then back up and print it
1802 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1803 assert(delay != NULL, "no unconditional delay instruction");
1804 if (WizardMode) delay->dump();
1806 if (node_bundling(delay)->starts_bundle())
1807 starts_bundle = '+';
1808 if (pcs && n->_idx < pc_limit)
1809 tty->print("%3.3x", pcs[n->_idx]);
1810 else
1811 tty->print(" ");
1812 tty->print(" %c ", starts_bundle);
1813 starts_bundle = ' ';
1814 tty->print("\t");
1815 delay->format(_regalloc, tty);
1816 tty->print_cr("");
1817 delay = NULL;
1818 }
1820 // Dump the exception table as well
1821 if( n->is_Catch() && (Verbose || WizardMode) ) {
1822 // Print the exception table for this offset
1823 _handler_table.print_subtable_for(pc);
1824 }
1825 }
1827 if (pcs && n->_idx < pc_limit)
1828 tty->print_cr("%3.3x", pcs[n->_idx]);
1829 else
1830 tty->print_cr("");
1832 assert(cut_short || delay == NULL, "no unconditional delay branch");
1834 } // End of per-block dump
1835 tty->print_cr("");
1837 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1838 }
1839 #endif
1841 //------------------------------Final_Reshape_Counts---------------------------
1842 // This class defines counters to help identify when a method
1843 // may/must be executed using hardware with only 24-bit precision.
1844 struct Final_Reshape_Counts : public StackObj {
1845 int _call_count; // count non-inlined 'common' calls
1846 int _float_count; // count float ops requiring 24-bit precision
1847 int _double_count; // count double ops requiring more precision
1848 int _java_call_count; // count non-inlined 'java' calls
1849 VectorSet _visited; // Visitation flags
1850 Node_List _tests; // Set of IfNodes & PCTableNodes
1852 Final_Reshape_Counts() :
1853 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0),
1854 _visited( Thread::current()->resource_area() ) { }
1856 void inc_call_count () { _call_count ++; }
1857 void inc_float_count () { _float_count ++; }
1858 void inc_double_count() { _double_count++; }
1859 void inc_java_call_count() { _java_call_count++; }
1861 int get_call_count () const { return _call_count ; }
1862 int get_float_count () const { return _float_count ; }
1863 int get_double_count() const { return _double_count; }
1864 int get_java_call_count() const { return _java_call_count; }
1865 };
1867 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1868 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1869 // Make sure the offset goes inside the instance layout.
1870 return k->contains_field_offset(tp->offset());
1871 // Note that OffsetBot and OffsetTop are very negative.
1872 }
1874 //------------------------------final_graph_reshaping_impl----------------------
1875 // Implement items 1-5 from final_graph_reshaping below.
1876 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
1878 if ( n->outcnt() == 0 ) return; // dead node
1879 uint nop = n->Opcode();
1881 // Check for 2-input instruction with "last use" on right input.
1882 // Swap to left input. Implements item (2).
1883 if( n->req() == 3 && // two-input instruction
1884 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1885 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1886 n->in(2)->outcnt() == 1 &&// right use IS a last use
1887 !n->in(2)->is_Con() ) { // right use is not a constant
1888 // Check for commutative opcode
1889 switch( nop ) {
1890 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1891 case Op_MaxI: case Op_MinI:
1892 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1893 case Op_AndL: case Op_XorL: case Op_OrL:
1894 case Op_AndI: case Op_XorI: case Op_OrI: {
1895 // Move "last use" input to left by swapping inputs
1896 n->swap_edges(1, 2);
1897 break;
1898 }
1899 default:
1900 break;
1901 }
1902 }
1904 // Count FPU ops and common calls, implements item (3)
1905 switch( nop ) {
1906 // Count all float operations that may use FPU
1907 case Op_AddF:
1908 case Op_SubF:
1909 case Op_MulF:
1910 case Op_DivF:
1911 case Op_NegF:
1912 case Op_ModF:
1913 case Op_ConvI2F:
1914 case Op_ConF:
1915 case Op_CmpF:
1916 case Op_CmpF3:
1917 // case Op_ConvL2F: // longs are split into 32-bit halves
1918 fpu.inc_float_count();
1919 break;
1921 case Op_ConvF2D:
1922 case Op_ConvD2F:
1923 fpu.inc_float_count();
1924 fpu.inc_double_count();
1925 break;
1927 // Count all double operations that may use FPU
1928 case Op_AddD:
1929 case Op_SubD:
1930 case Op_MulD:
1931 case Op_DivD:
1932 case Op_NegD:
1933 case Op_ModD:
1934 case Op_ConvI2D:
1935 case Op_ConvD2I:
1936 // case Op_ConvL2D: // handled by leaf call
1937 // case Op_ConvD2L: // handled by leaf call
1938 case Op_ConD:
1939 case Op_CmpD:
1940 case Op_CmpD3:
1941 fpu.inc_double_count();
1942 break;
1943 case Op_Opaque1: // Remove Opaque Nodes before matching
1944 case Op_Opaque2: // Remove Opaque Nodes before matching
1945 n->subsume_by(n->in(1));
1946 break;
1947 case Op_CallStaticJava:
1948 case Op_CallJava:
1949 case Op_CallDynamicJava:
1950 fpu.inc_java_call_count(); // Count java call site;
1951 case Op_CallRuntime:
1952 case Op_CallLeaf:
1953 case Op_CallLeafNoFP: {
1954 assert( n->is_Call(), "" );
1955 CallNode *call = n->as_Call();
1956 // Count call sites where the FP mode bit would have to be flipped.
1957 // Do not count uncommon runtime calls:
1958 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
1959 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
1960 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
1961 fpu.inc_call_count(); // Count the call site
1962 } else { // See if uncommon argument is shared
1963 Node *n = call->in(TypeFunc::Parms);
1964 int nop = n->Opcode();
1965 // Clone shared simple arguments to uncommon calls, item (1).
1966 if( n->outcnt() > 1 &&
1967 !n->is_Proj() &&
1968 nop != Op_CreateEx &&
1969 nop != Op_CheckCastPP &&
1970 !n->is_Mem() ) {
1971 Node *x = n->clone();
1972 call->set_req( TypeFunc::Parms, x );
1973 }
1974 }
1975 break;
1976 }
1978 case Op_StoreD:
1979 case Op_LoadD:
1980 case Op_LoadD_unaligned:
1981 fpu.inc_double_count();
1982 goto handle_mem;
1983 case Op_StoreF:
1984 case Op_LoadF:
1985 fpu.inc_float_count();
1986 goto handle_mem;
1988 case Op_StoreB:
1989 case Op_StoreC:
1990 case Op_StoreCM:
1991 case Op_StorePConditional:
1992 case Op_StoreI:
1993 case Op_StoreL:
1994 case Op_StoreLConditional:
1995 case Op_CompareAndSwapI:
1996 case Op_CompareAndSwapL:
1997 case Op_CompareAndSwapP:
1998 case Op_CompareAndSwapN:
1999 case Op_StoreP:
2000 case Op_StoreN:
2001 case Op_LoadB:
2002 case Op_LoadC:
2003 case Op_LoadI:
2004 case Op_LoadKlass:
2005 case Op_LoadNKlass:
2006 case Op_LoadL:
2007 case Op_LoadL_unaligned:
2008 case Op_LoadPLocked:
2009 case Op_LoadLLocked:
2010 case Op_LoadP:
2011 case Op_LoadN:
2012 case Op_LoadRange:
2013 case Op_LoadS: {
2014 handle_mem:
2015 #ifdef ASSERT
2016 if( VerifyOptoOopOffsets ) {
2017 assert( n->is_Mem(), "" );
2018 MemNode *mem = (MemNode*)n;
2019 // Check to see if address types have grounded out somehow.
2020 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2021 assert( !tp || oop_offset_is_sane(tp), "" );
2022 }
2023 #endif
2024 break;
2025 }
2027 case Op_AddP: { // Assert sane base pointers
2028 Node *addp = n->in(AddPNode::Address);
2029 assert( !addp->is_AddP() ||
2030 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2031 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2032 "Base pointers must match" );
2033 #ifdef _LP64
2034 if (UseCompressedOops &&
2035 addp->Opcode() == Op_ConP &&
2036 addp == n->in(AddPNode::Base) &&
2037 n->in(AddPNode::Offset)->is_Con()) {
2038 // Use addressing with narrow klass to load with offset on x86.
2039 // On sparc loading 32-bits constant and decoding it have less
2040 // instructions (4) then load 64-bits constant (7).
2041 // Do this transformation here since IGVN will convert ConN back to ConP.
2042 const Type* t = addp->bottom_type();
2043 if (t->isa_oopptr()) {
2044 Node* nn = NULL;
2046 // Look for existing ConN node of the same exact type.
2047 Compile* C = Compile::current();
2048 Node* r = C->root();
2049 uint cnt = r->outcnt();
2050 for (uint i = 0; i < cnt; i++) {
2051 Node* m = r->raw_out(i);
2052 if (m!= NULL && m->Opcode() == Op_ConN &&
2053 m->bottom_type()->make_ptr() == t) {
2054 nn = m;
2055 break;
2056 }
2057 }
2058 if (nn != NULL) {
2059 // Decode a narrow oop to match address
2060 // [R12 + narrow_oop_reg<<3 + offset]
2061 nn = new (C, 2) DecodeNNode(nn, t);
2062 n->set_req(AddPNode::Base, nn);
2063 n->set_req(AddPNode::Address, nn);
2064 if (addp->outcnt() == 0) {
2065 addp->disconnect_inputs(NULL);
2066 }
2067 }
2068 }
2069 }
2070 #endif
2071 break;
2072 }
2074 #ifdef _LP64
2075 case Op_CmpP:
2076 // Do this transformation here to preserve CmpPNode::sub() and
2077 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2078 if( n->in(1)->is_DecodeN() ) {
2079 Compile* C = Compile::current();
2080 Node* in2 = NULL;
2081 if( n->in(2)->is_DecodeN() ) {
2082 in2 = n->in(2)->in(1);
2083 } else if ( n->in(2)->Opcode() == Op_ConP ) {
2084 const Type* t = n->in(2)->bottom_type();
2085 if (t == TypePtr::NULL_PTR && UseImplicitNullCheckForNarrowOop) {
2086 Node *in1 = n->in(1);
2087 if (Matcher::clone_shift_expressions) {
2088 // x86, ARM and friends can handle 2 adds in addressing mode.
2089 // Decode a narrow oop and do implicit NULL check in address
2090 // [R12 + narrow_oop_reg<<3 + offset]
2091 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2092 } else {
2093 // Don't replace CmpP(o ,null) if 'o' is used in AddP
2094 // to generate implicit NULL check on Sparc where
2095 // narrow oops can't be used in address.
2096 uint i = 0;
2097 for (; i < in1->outcnt(); i++) {
2098 if (in1->raw_out(i)->is_AddP())
2099 break;
2100 }
2101 if (i >= in1->outcnt()) {
2102 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2103 }
2104 }
2105 } else if (t->isa_oopptr()) {
2106 in2 = ConNode::make(C, t->make_narrowoop());
2107 }
2108 }
2109 if( in2 != NULL ) {
2110 Node* cmpN = new (C, 3) CmpNNode(n->in(1)->in(1), in2);
2111 n->subsume_by( cmpN );
2112 }
2113 }
2114 break;
2115 #endif
2117 case Op_ModI:
2118 if (UseDivMod) {
2119 // Check if a%b and a/b both exist
2120 Node* d = n->find_similar(Op_DivI);
2121 if (d) {
2122 // Replace them with a fused divmod if supported
2123 Compile* C = Compile::current();
2124 if (Matcher::has_match_rule(Op_DivModI)) {
2125 DivModINode* divmod = DivModINode::make(C, n);
2126 d->subsume_by(divmod->div_proj());
2127 n->subsume_by(divmod->mod_proj());
2128 } else {
2129 // replace a%b with a-((a/b)*b)
2130 Node* mult = new (C, 3) MulINode(d, d->in(2));
2131 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2132 n->subsume_by( sub );
2133 }
2134 }
2135 }
2136 break;
2138 case Op_ModL:
2139 if (UseDivMod) {
2140 // Check if a%b and a/b both exist
2141 Node* d = n->find_similar(Op_DivL);
2142 if (d) {
2143 // Replace them with a fused divmod if supported
2144 Compile* C = Compile::current();
2145 if (Matcher::has_match_rule(Op_DivModL)) {
2146 DivModLNode* divmod = DivModLNode::make(C, n);
2147 d->subsume_by(divmod->div_proj());
2148 n->subsume_by(divmod->mod_proj());
2149 } else {
2150 // replace a%b with a-((a/b)*b)
2151 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2152 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2153 n->subsume_by( sub );
2154 }
2155 }
2156 }
2157 break;
2159 case Op_Load16B:
2160 case Op_Load8B:
2161 case Op_Load4B:
2162 case Op_Load8S:
2163 case Op_Load4S:
2164 case Op_Load2S:
2165 case Op_Load8C:
2166 case Op_Load4C:
2167 case Op_Load2C:
2168 case Op_Load4I:
2169 case Op_Load2I:
2170 case Op_Load2L:
2171 case Op_Load4F:
2172 case Op_Load2F:
2173 case Op_Load2D:
2174 case Op_Store16B:
2175 case Op_Store8B:
2176 case Op_Store4B:
2177 case Op_Store8C:
2178 case Op_Store4C:
2179 case Op_Store2C:
2180 case Op_Store4I:
2181 case Op_Store2I:
2182 case Op_Store2L:
2183 case Op_Store4F:
2184 case Op_Store2F:
2185 case Op_Store2D:
2186 break;
2188 case Op_PackB:
2189 case Op_PackS:
2190 case Op_PackC:
2191 case Op_PackI:
2192 case Op_PackF:
2193 case Op_PackL:
2194 case Op_PackD:
2195 if (n->req()-1 > 2) {
2196 // Replace many operand PackNodes with a binary tree for matching
2197 PackNode* p = (PackNode*) n;
2198 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2199 n->subsume_by(btp);
2200 }
2201 break;
2202 default:
2203 assert( !n->is_Call(), "" );
2204 assert( !n->is_Mem(), "" );
2205 break;
2206 }
2208 // Collect CFG split points
2209 if (n->is_MultiBranch())
2210 fpu._tests.push(n);
2211 }
2213 //------------------------------final_graph_reshaping_walk---------------------
2214 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2215 // requires that the walk visits a node's inputs before visiting the node.
2216 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) {
2217 fpu._visited.set(root->_idx); // first, mark node as visited
2218 uint cnt = root->req();
2219 Node *n = root;
2220 uint i = 0;
2221 while (true) {
2222 if (i < cnt) {
2223 // Place all non-visited non-null inputs onto stack
2224 Node* m = n->in(i);
2225 ++i;
2226 if (m != NULL && !fpu._visited.test_set(m->_idx)) {
2227 cnt = m->req();
2228 nstack.push(n, i); // put on stack parent and next input's index
2229 n = m;
2230 i = 0;
2231 }
2232 } else {
2233 // Now do post-visit work
2234 final_graph_reshaping_impl( n, fpu );
2235 if (nstack.is_empty())
2236 break; // finished
2237 n = nstack.node(); // Get node from stack
2238 cnt = n->req();
2239 i = nstack.index();
2240 nstack.pop(); // Shift to the next node on stack
2241 }
2242 }
2243 }
2245 //------------------------------final_graph_reshaping--------------------------
2246 // Final Graph Reshaping.
2247 //
2248 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2249 // and not commoned up and forced early. Must come after regular
2250 // optimizations to avoid GVN undoing the cloning. Clone constant
2251 // inputs to Loop Phis; these will be split by the allocator anyways.
2252 // Remove Opaque nodes.
2253 // (2) Move last-uses by commutative operations to the left input to encourage
2254 // Intel update-in-place two-address operations and better register usage
2255 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2256 // calls canonicalizing them back.
2257 // (3) Count the number of double-precision FP ops, single-precision FP ops
2258 // and call sites. On Intel, we can get correct rounding either by
2259 // forcing singles to memory (requires extra stores and loads after each
2260 // FP bytecode) or we can set a rounding mode bit (requires setting and
2261 // clearing the mode bit around call sites). The mode bit is only used
2262 // if the relative frequency of single FP ops to calls is low enough.
2263 // This is a key transform for SPEC mpeg_audio.
2264 // (4) Detect infinite loops; blobs of code reachable from above but not
2265 // below. Several of the Code_Gen algorithms fail on such code shapes,
2266 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2267 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2268 // Detection is by looking for IfNodes where only 1 projection is
2269 // reachable from below or CatchNodes missing some targets.
2270 // (5) Assert for insane oop offsets in debug mode.
2272 bool Compile::final_graph_reshaping() {
2273 // an infinite loop may have been eliminated by the optimizer,
2274 // in which case the graph will be empty.
2275 if (root()->req() == 1) {
2276 record_method_not_compilable("trivial infinite loop");
2277 return true;
2278 }
2280 Final_Reshape_Counts fpu;
2282 // Visit everybody reachable!
2283 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2284 Node_Stack nstack(unique() >> 1);
2285 final_graph_reshaping_walk(nstack, root(), fpu);
2287 // Check for unreachable (from below) code (i.e., infinite loops).
2288 for( uint i = 0; i < fpu._tests.size(); i++ ) {
2289 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch();
2290 // Get number of CFG targets.
2291 // Note that PCTables include exception targets after calls.
2292 uint required_outcnt = n->required_outcnt();
2293 if (n->outcnt() != required_outcnt) {
2294 // Check for a few special cases. Rethrow Nodes never take the
2295 // 'fall-thru' path, so expected kids is 1 less.
2296 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2297 if (n->in(0)->in(0)->is_Call()) {
2298 CallNode *call = n->in(0)->in(0)->as_Call();
2299 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2300 required_outcnt--; // Rethrow always has 1 less kid
2301 } else if (call->req() > TypeFunc::Parms &&
2302 call->is_CallDynamicJava()) {
2303 // Check for null receiver. In such case, the optimizer has
2304 // detected that the virtual call will always result in a null
2305 // pointer exception. The fall-through projection of this CatchNode
2306 // will not be populated.
2307 Node *arg0 = call->in(TypeFunc::Parms);
2308 if (arg0->is_Type() &&
2309 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2310 required_outcnt--;
2311 }
2312 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2313 call->req() > TypeFunc::Parms+1 &&
2314 call->is_CallStaticJava()) {
2315 // Check for negative array length. In such case, the optimizer has
2316 // detected that the allocation attempt will always result in an
2317 // exception. There is no fall-through projection of this CatchNode .
2318 Node *arg1 = call->in(TypeFunc::Parms+1);
2319 if (arg1->is_Type() &&
2320 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2321 required_outcnt--;
2322 }
2323 }
2324 }
2325 }
2326 // Recheck with a better notion of 'required_outcnt'
2327 if (n->outcnt() != required_outcnt) {
2328 record_method_not_compilable("malformed control flow");
2329 return true; // Not all targets reachable!
2330 }
2331 }
2332 // Check that I actually visited all kids. Unreached kids
2333 // must be infinite loops.
2334 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2335 if (!fpu._visited.test(n->fast_out(j)->_idx)) {
2336 record_method_not_compilable("infinite loop");
2337 return true; // Found unvisited kid; must be unreach
2338 }
2339 }
2341 // If original bytecodes contained a mixture of floats and doubles
2342 // check if the optimizer has made it homogenous, item (3).
2343 if( Use24BitFPMode && Use24BitFP &&
2344 fpu.get_float_count() > 32 &&
2345 fpu.get_double_count() == 0 &&
2346 (10 * fpu.get_call_count() < fpu.get_float_count()) ) {
2347 set_24_bit_selection_and_mode( false, true );
2348 }
2350 set_has_java_calls(fpu.get_java_call_count() > 0);
2352 // No infinite loops, no reason to bail out.
2353 return false;
2354 }
2356 //-----------------------------too_many_traps----------------------------------
2357 // Report if there are too many traps at the current method and bci.
2358 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2359 bool Compile::too_many_traps(ciMethod* method,
2360 int bci,
2361 Deoptimization::DeoptReason reason) {
2362 ciMethodData* md = method->method_data();
2363 if (md->is_empty()) {
2364 // Assume the trap has not occurred, or that it occurred only
2365 // because of a transient condition during start-up in the interpreter.
2366 return false;
2367 }
2368 if (md->has_trap_at(bci, reason) != 0) {
2369 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2370 // Also, if there are multiple reasons, or if there is no per-BCI record,
2371 // assume the worst.
2372 if (log())
2373 log()->elem("observe trap='%s' count='%d'",
2374 Deoptimization::trap_reason_name(reason),
2375 md->trap_count(reason));
2376 return true;
2377 } else {
2378 // Ignore method/bci and see if there have been too many globally.
2379 return too_many_traps(reason, md);
2380 }
2381 }
2383 // Less-accurate variant which does not require a method and bci.
2384 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2385 ciMethodData* logmd) {
2386 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2387 // Too many traps globally.
2388 // Note that we use cumulative trap_count, not just md->trap_count.
2389 if (log()) {
2390 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2391 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2392 Deoptimization::trap_reason_name(reason),
2393 mcount, trap_count(reason));
2394 }
2395 return true;
2396 } else {
2397 // The coast is clear.
2398 return false;
2399 }
2400 }
2402 //--------------------------too_many_recompiles--------------------------------
2403 // Report if there are too many recompiles at the current method and bci.
2404 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2405 // Is not eager to return true, since this will cause the compiler to use
2406 // Action_none for a trap point, to avoid too many recompilations.
2407 bool Compile::too_many_recompiles(ciMethod* method,
2408 int bci,
2409 Deoptimization::DeoptReason reason) {
2410 ciMethodData* md = method->method_data();
2411 if (md->is_empty()) {
2412 // Assume the trap has not occurred, or that it occurred only
2413 // because of a transient condition during start-up in the interpreter.
2414 return false;
2415 }
2416 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2417 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2418 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2419 Deoptimization::DeoptReason per_bc_reason
2420 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2421 if ((per_bc_reason == Deoptimization::Reason_none
2422 || md->has_trap_at(bci, reason) != 0)
2423 // The trap frequency measure we care about is the recompile count:
2424 && md->trap_recompiled_at(bci)
2425 && md->overflow_recompile_count() >= bc_cutoff) {
2426 // Do not emit a trap here if it has already caused recompilations.
2427 // Also, if there are multiple reasons, or if there is no per-BCI record,
2428 // assume the worst.
2429 if (log())
2430 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2431 Deoptimization::trap_reason_name(reason),
2432 md->trap_count(reason),
2433 md->overflow_recompile_count());
2434 return true;
2435 } else if (trap_count(reason) != 0
2436 && decompile_count() >= m_cutoff) {
2437 // Too many recompiles globally, and we have seen this sort of trap.
2438 // Use cumulative decompile_count, not just md->decompile_count.
2439 if (log())
2440 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2441 Deoptimization::trap_reason_name(reason),
2442 md->trap_count(reason), trap_count(reason),
2443 md->decompile_count(), decompile_count());
2444 return true;
2445 } else {
2446 // The coast is clear.
2447 return false;
2448 }
2449 }
2452 #ifndef PRODUCT
2453 //------------------------------verify_graph_edges---------------------------
2454 // Walk the Graph and verify that there is a one-to-one correspondence
2455 // between Use-Def edges and Def-Use edges in the graph.
2456 void Compile::verify_graph_edges(bool no_dead_code) {
2457 if (VerifyGraphEdges) {
2458 ResourceArea *area = Thread::current()->resource_area();
2459 Unique_Node_List visited(area);
2460 // Call recursive graph walk to check edges
2461 _root->verify_edges(visited);
2462 if (no_dead_code) {
2463 // Now make sure that no visited node is used by an unvisited node.
2464 bool dead_nodes = 0;
2465 Unique_Node_List checked(area);
2466 while (visited.size() > 0) {
2467 Node* n = visited.pop();
2468 checked.push(n);
2469 for (uint i = 0; i < n->outcnt(); i++) {
2470 Node* use = n->raw_out(i);
2471 if (checked.member(use)) continue; // already checked
2472 if (visited.member(use)) continue; // already in the graph
2473 if (use->is_Con()) continue; // a dead ConNode is OK
2474 // At this point, we have found a dead node which is DU-reachable.
2475 if (dead_nodes++ == 0)
2476 tty->print_cr("*** Dead nodes reachable via DU edges:");
2477 use->dump(2);
2478 tty->print_cr("---");
2479 checked.push(use); // No repeats; pretend it is now checked.
2480 }
2481 }
2482 assert(dead_nodes == 0, "using nodes must be reachable from root");
2483 }
2484 }
2485 }
2486 #endif
2488 // The Compile object keeps track of failure reasons separately from the ciEnv.
2489 // This is required because there is not quite a 1-1 relation between the
2490 // ciEnv and its compilation task and the Compile object. Note that one
2491 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2492 // to backtrack and retry without subsuming loads. Other than this backtracking
2493 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2494 // by the logic in C2Compiler.
2495 void Compile::record_failure(const char* reason) {
2496 if (log() != NULL) {
2497 log()->elem("failure reason='%s' phase='compile'", reason);
2498 }
2499 if (_failure_reason == NULL) {
2500 // Record the first failure reason.
2501 _failure_reason = reason;
2502 }
2503 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2504 C->print_method(_failure_reason);
2505 }
2506 _root = NULL; // flush the graph, too
2507 }
2509 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2510 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2511 {
2512 if (dolog) {
2513 C = Compile::current();
2514 _log = C->log();
2515 } else {
2516 C = NULL;
2517 _log = NULL;
2518 }
2519 if (_log != NULL) {
2520 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2521 _log->stamp();
2522 _log->end_head();
2523 }
2524 }
2526 Compile::TracePhase::~TracePhase() {
2527 if (_log != NULL) {
2528 _log->done("phase nodes='%d'", C->unique());
2529 }
2530 }