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 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_compile.cpp.incl"
27
28 /// Support for intrinsics.
29
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 }
66
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 }
92
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 }
116
117 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
118 // in library_call.cpp.
119
120
121 #ifndef PRODUCT
122 // statistics gathering...
123
124 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
125 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
126
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 }
163
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 }
175
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 }
196
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
214
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 }
220
221 bool Compile::valid_bundle_info(const Node *n) {
222 return (_node_bundling_limit > n->_idx);
223 }
224
225
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
232
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()); }
237
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 }
250
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 }
275
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 }
285
286 // ============================================================================
287 //------------------------------CompileWrapper---------------------------------
288 class CompileWrapper : public StackObj {
289 Compile *const _compile;
290 public:
291 CompileWrapper(Compile* compile);
292
293 ~CompileWrapper();
294 };
295
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");
303
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 }
321
322
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 }
346
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 }
356
357
358 void Compile::init_scratch_buffer_blob() {
359 if( scratch_buffer_blob() != NULL ) return;
360
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 }
374
375 // Initialize the relocation buffers
376 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size;
377 set_scratch_locs_memory(locs_buf);
378 }
379
380
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 }
411
412
413 // ============================================================================
414 //------------------------------Compile standard-------------------------------
415 debug_only( int Compile::_debug_idx = 100000; )
416
417 // Compile a method. entry_bci is -1 for normal compilations and indicates
418 // the continuation bci for on stack replacement.
419
420
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;
449
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 set_parsed_irreducible_loop(false);
471 #endif
472
473 if (ProfileTraps) {
474 // Make sure the method being compiled gets its own MDO,
475 // so we can at least track the decompile_count().
476 method()->build_method_data();
477 }
478
479 Init(::AliasLevel);
480
481
482 print_compile_messages();
483
484 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
485 _ilt = InlineTree::build_inline_tree_root();
486 else
487 _ilt = NULL;
488
489 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
490 assert(num_alias_types() >= AliasIdxRaw, "");
491
492 #define MINIMUM_NODE_HASH 1023
493 // Node list that Iterative GVN will start with
494 Unique_Node_List for_igvn(comp_arena());
495 set_for_igvn(&for_igvn);
496
497 // GVN that will be run immediately on new nodes
498 uint estimated_size = method()->code_size()*4+64;
499 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
500 PhaseGVN gvn(node_arena(), estimated_size);
501 set_initial_gvn(&gvn);
502
503 { // Scope for timing the parser
504 TracePhase t3("parse", &_t_parser, true);
505
506 // Put top into the hash table ASAP.
507 initial_gvn()->transform_no_reclaim(top());
508
509 // Set up tf(), start(), and find a CallGenerator.
510 CallGenerator* cg;
511 if (is_osr_compilation()) {
512 const TypeTuple *domain = StartOSRNode::osr_domain();
513 const TypeTuple *range = TypeTuple::make_range(method()->signature());
514 init_tf(TypeFunc::make(domain, range));
515 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
516 initial_gvn()->set_type_bottom(s);
517 init_start(s);
518 cg = CallGenerator::for_osr(method(), entry_bci());
519 } else {
520 // Normal case.
521 init_tf(TypeFunc::make(method()));
522 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
523 initial_gvn()->set_type_bottom(s);
524 init_start(s);
525 float past_uses = method()->interpreter_invocation_count();
526 float expected_uses = past_uses;
527 cg = CallGenerator::for_inline(method(), expected_uses);
528 }
529 if (failing()) return;
530 if (cg == NULL) {
531 record_method_not_compilable_all_tiers("cannot parse method");
532 return;
533 }
534 JVMState* jvms = build_start_state(start(), tf());
535 if ((jvms = cg->generate(jvms)) == NULL) {
536 record_method_not_compilable("method parse failed");
537 return;
538 }
539 GraphKit kit(jvms);
540
541 if (!kit.stopped()) {
542 // Accept return values, and transfer control we know not where.
543 // This is done by a special, unique ReturnNode bound to root.
544 return_values(kit.jvms());
545 }
546
547 if (kit.has_exceptions()) {
548 // Any exceptions that escape from this call must be rethrown
549 // to whatever caller is dynamically above us on the stack.
550 // This is done by a special, unique RethrowNode bound to root.
551 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
552 }
553
554 print_method("Before RemoveUseless");
555
556 // Remove clutter produced by parsing.
557 if (!failing()) {
558 ResourceMark rm;
559 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
560 }
561 }
562
563 // Note: Large methods are capped off in do_one_bytecode().
564 if (failing()) return;
565
566 // After parsing, node notes are no longer automagic.
567 // They must be propagated by register_new_node_with_optimizer(),
568 // clone(), or the like.
569 set_default_node_notes(NULL);
570
571 for (;;) {
572 int successes = Inline_Warm();
573 if (failing()) return;
574 if (successes == 0) break;
575 }
576
577 // Drain the list.
578 Finish_Warm();
579 #ifndef PRODUCT
580 if (_printer) {
581 _printer->print_inlining(this);
582 }
583 #endif
584
585 if (failing()) return;
586 NOT_PRODUCT( verify_graph_edges(); )
587
588 // Perform escape analysis
589 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
590 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
591 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction.
592 PhaseGVN* igvn = initial_gvn();
593 Node* oop_null = igvn->zerocon(T_OBJECT);
594 Node* noop_null = igvn->zerocon(T_NARROWOOP);
595
596 _congraph = new(comp_arena()) ConnectionGraph(this);
597 bool has_non_escaping_obj = _congraph->compute_escape();
598
599 #ifndef PRODUCT
600 if (PrintEscapeAnalysis) {
601 _congraph->dump();
602 }
603 #endif
604 // Cleanup.
605 if (oop_null->outcnt() == 0)
606 igvn->hash_delete(oop_null);
607 if (noop_null->outcnt() == 0)
608 igvn->hash_delete(noop_null);
609
610 if (!has_non_escaping_obj) {
611 _congraph = NULL;
612 }
613
614 if (failing()) return;
615 }
616 // Now optimize
617 Optimize();
618 if (failing()) return;
619 NOT_PRODUCT( verify_graph_edges(); )
620
621 #ifndef PRODUCT
622 if (PrintIdeal) {
623 ttyLocker ttyl; // keep the following output all in one block
624 // This output goes directly to the tty, not the compiler log.
625 // To enable tools to match it up with the compilation activity,
626 // be sure to tag this tty output with the compile ID.
627 if (xtty != NULL) {
628 xtty->head("ideal compile_id='%d'%s", compile_id(),
629 is_osr_compilation() ? " compile_kind='osr'" :
630 "");
631 }
632 root()->dump(9999);
633 if (xtty != NULL) {
634 xtty->tail("ideal");
635 }
636 }
637 #endif
638
639 // Now that we know the size of all the monitors we can add a fixed slot
640 // for the original deopt pc.
641
642 _orig_pc_slot = fixed_slots();
643 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
644 set_fixed_slots(next_slot);
645
646 // Now generate code
647 Code_Gen();
648 if (failing()) return;
649
650 // Check if we want to skip execution of all compiled code.
651 {
652 #ifndef PRODUCT
653 if (OptoNoExecute) {
654 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
655 return;
656 }
657 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
658 #endif
659
660 if (is_osr_compilation()) {
661 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
662 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
663 } else {
664 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
665 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
666 }
667
668 env()->register_method(_method, _entry_bci,
669 &_code_offsets,
670 _orig_pc_slot_offset_in_bytes,
671 code_buffer(),
672 frame_size_in_words(), _oop_map_set,
673 &_handler_table, &_inc_table,
674 compiler,
675 env()->comp_level(),
676 true, /*has_debug_info*/
677 has_unsafe_access()
678 );
679 }
680 }
681
682 //------------------------------Compile----------------------------------------
683 // Compile a runtime stub
684 Compile::Compile( ciEnv* ci_env,
685 TypeFunc_generator generator,
686 address stub_function,
687 const char *stub_name,
688 int is_fancy_jump,
689 bool pass_tls,
690 bool save_arg_registers,
691 bool return_pc )
692 : Phase(Compiler),
693 _env(ci_env),
694 _log(ci_env->log()),
695 _compile_id(-1),
696 _save_argument_registers(save_arg_registers),
697 _method(NULL),
698 _stub_name(stub_name),
699 _stub_function(stub_function),
700 _stub_entry_point(NULL),
701 _entry_bci(InvocationEntryBci),
702 _initial_gvn(NULL),
703 _for_igvn(NULL),
704 _warm_calls(NULL),
705 _orig_pc_slot(0),
706 _orig_pc_slot_offset_in_bytes(0),
707 _subsume_loads(true),
708 _do_escape_analysis(false),
709 _failure_reason(NULL),
710 _code_buffer("Compile::Fill_buffer"),
711 _node_bundling_limit(0),
712 _node_bundling_base(NULL),
713 #ifndef PRODUCT
714 _trace_opto_output(TraceOptoOutput),
715 _printer(NULL),
716 #endif
717 _congraph(NULL) {
718 C = this;
719
720 #ifndef PRODUCT
721 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
722 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
723 set_print_assembly(PrintFrameConverterAssembly);
724 set_parsed_irreducible_loop(false);
725 #endif
726 CompileWrapper cw(this);
727 Init(/*AliasLevel=*/ 0);
728 init_tf((*generator)());
729
730 {
731 // The following is a dummy for the sake of GraphKit::gen_stub
732 Unique_Node_List for_igvn(comp_arena());
733 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
734 PhaseGVN gvn(Thread::current()->resource_area(),255);
735 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
736 gvn.transform_no_reclaim(top());
737
738 GraphKit kit;
739 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
740 }
741
742 NOT_PRODUCT( verify_graph_edges(); )
743 Code_Gen();
744 if (failing()) return;
745
746
747 // Entry point will be accessed using compile->stub_entry_point();
748 if (code_buffer() == NULL) {
749 Matcher::soft_match_failure();
750 } else {
751 if (PrintAssembly && (WizardMode || Verbose))
752 tty->print_cr("### Stub::%s", stub_name);
753
754 if (!failing()) {
755 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
756
757 // Make the NMethod
758 // For now we mark the frame as never safe for profile stackwalking
759 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
760 code_buffer(),
761 CodeOffsets::frame_never_safe,
762 // _code_offsets.value(CodeOffsets::Frame_Complete),
763 frame_size_in_words(),
764 _oop_map_set,
765 save_arg_registers);
766 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
767
768 _stub_entry_point = rs->entry_point();
769 }
770 }
771 }
772
773 #ifndef PRODUCT
774 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
775 if(PrintOpto && Verbose) {
776 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
777 }
778 }
779 #endif
780
781 void Compile::print_codes() {
782 }
783
784 //------------------------------Init-------------------------------------------
785 // Prepare for a single compilation
786 void Compile::Init(int aliaslevel) {
787 _unique = 0;
788 _regalloc = NULL;
789
790 _tf = NULL; // filled in later
791 _top = NULL; // cached later
792 _matcher = NULL; // filled in later
793 _cfg = NULL; // filled in later
794
795 set_24_bit_selection_and_mode(Use24BitFP, false);
796
797 _node_note_array = NULL;
798 _default_node_notes = NULL;
799
800 _immutable_memory = NULL; // filled in at first inquiry
801
802 // Globally visible Nodes
803 // First set TOP to NULL to give safe behavior during creation of RootNode
804 set_cached_top_node(NULL);
805 set_root(new (this, 3) RootNode());
806 // Now that you have a Root to point to, create the real TOP
807 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
808 set_recent_alloc(NULL, NULL);
809
810 // Create Debug Information Recorder to record scopes, oopmaps, etc.
811 env()->set_oop_recorder(new OopRecorder(comp_arena()));
812 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
813 env()->set_dependencies(new Dependencies(env()));
814
815 _fixed_slots = 0;
816 set_has_split_ifs(false);
817 set_has_loops(has_method() && method()->has_loops()); // first approximation
818 _deopt_happens = true; // start out assuming the worst
819 _trap_can_recompile = false; // no traps emitted yet
820 _major_progress = true; // start out assuming good things will happen
821 set_has_unsafe_access(false);
822 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
823 set_decompile_count(0);
824
825 // Compilation level related initialization
826 if (env()->comp_level() == CompLevel_fast_compile) {
827 set_num_loop_opts(Tier1LoopOptsCount);
828 set_do_inlining(Tier1Inline != 0);
829 set_max_inline_size(Tier1MaxInlineSize);
830 set_freq_inline_size(Tier1FreqInlineSize);
831 set_do_scheduling(false);
832 set_do_count_invocations(Tier1CountInvocations);
833 set_do_method_data_update(Tier1UpdateMethodData);
834 } else {
835 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
836 set_num_loop_opts(LoopOptsCount);
837 set_do_inlining(Inline);
838 set_max_inline_size(MaxInlineSize);
839 set_freq_inline_size(FreqInlineSize);
840 set_do_scheduling(OptoScheduling);
841 set_do_count_invocations(false);
842 set_do_method_data_update(false);
843 }
844
845 if (debug_info()->recording_non_safepoints()) {
846 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
847 (comp_arena(), 8, 0, NULL));
848 set_default_node_notes(Node_Notes::make(this));
849 }
850
851 // // -- Initialize types before each compile --
852 // // Update cached type information
853 // if( _method && _method->constants() )
854 // Type::update_loaded_types(_method, _method->constants());
855
856 // Init alias_type map.
857 if (!_do_escape_analysis && aliaslevel == 3)
858 aliaslevel = 2; // No unique types without escape analysis
859 _AliasLevel = aliaslevel;
860 const int grow_ats = 16;
861 _max_alias_types = grow_ats;
862 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
863 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
864 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
865 {
866 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
867 }
868 // Initialize the first few types.
869 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
870 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
871 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
872 _num_alias_types = AliasIdxRaw+1;
873 // Zero out the alias type cache.
874 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
875 // A NULL adr_type hits in the cache right away. Preload the right answer.
876 probe_alias_cache(NULL)->_index = AliasIdxTop;
877
878 _intrinsics = NULL;
879 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
880 register_library_intrinsics();
881 }
882
883 //---------------------------init_start----------------------------------------
884 // Install the StartNode on this compile object.
885 void Compile::init_start(StartNode* s) {
886 if (failing())
887 return; // already failing
888 assert(s == start(), "");
889 }
890
891 StartNode* Compile::start() const {
892 assert(!failing(), "");
893 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
894 Node* start = root()->fast_out(i);
895 if( start->is_Start() )
896 return start->as_Start();
897 }
898 ShouldNotReachHere();
899 return NULL;
900 }
901
902 //-------------------------------immutable_memory-------------------------------------
903 // Access immutable memory
904 Node* Compile::immutable_memory() {
905 if (_immutable_memory != NULL) {
906 return _immutable_memory;
907 }
908 StartNode* s = start();
909 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
910 Node *p = s->fast_out(i);
911 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
912 _immutable_memory = p;
913 return _immutable_memory;
914 }
915 }
916 ShouldNotReachHere();
917 return NULL;
918 }
919
920 //----------------------set_cached_top_node------------------------------------
921 // Install the cached top node, and make sure Node::is_top works correctly.
922 void Compile::set_cached_top_node(Node* tn) {
923 if (tn != NULL) verify_top(tn);
924 Node* old_top = _top;
925 _top = tn;
926 // Calling Node::setup_is_top allows the nodes the chance to adjust
927 // their _out arrays.
928 if (_top != NULL) _top->setup_is_top();
929 if (old_top != NULL) old_top->setup_is_top();
930 assert(_top == NULL || top()->is_top(), "");
931 }
932
933 #ifndef PRODUCT
934 void Compile::verify_top(Node* tn) const {
935 if (tn != NULL) {
936 assert(tn->is_Con(), "top node must be a constant");
937 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
938 assert(tn->in(0) != NULL, "must have live top node");
939 }
940 }
941 #endif
942
943
944 ///-------------------Managing Per-Node Debug & Profile Info-------------------
945
946 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
947 guarantee(arr != NULL, "");
948 int num_blocks = arr->length();
949 if (grow_by < num_blocks) grow_by = num_blocks;
950 int num_notes = grow_by * _node_notes_block_size;
951 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
952 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
953 while (num_notes > 0) {
954 arr->append(notes);
955 notes += _node_notes_block_size;
956 num_notes -= _node_notes_block_size;
957 }
958 assert(num_notes == 0, "exact multiple, please");
959 }
960
961 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
962 if (source == NULL || dest == NULL) return false;
963
964 if (dest->is_Con())
965 return false; // Do not push debug info onto constants.
966
967 #ifdef ASSERT
968 // Leave a bread crumb trail pointing to the original node:
969 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
970 dest->set_debug_orig(source);
971 }
972 #endif
973
974 if (node_note_array() == NULL)
975 return false; // Not collecting any notes now.
976
977 // This is a copy onto a pre-existing node, which may already have notes.
978 // If both nodes have notes, do not overwrite any pre-existing notes.
979 Node_Notes* source_notes = node_notes_at(source->_idx);
980 if (source_notes == NULL || source_notes->is_clear()) return false;
981 Node_Notes* dest_notes = node_notes_at(dest->_idx);
982 if (dest_notes == NULL || dest_notes->is_clear()) {
983 return set_node_notes_at(dest->_idx, source_notes);
984 }
985
986 Node_Notes merged_notes = (*source_notes);
987 // The order of operations here ensures that dest notes will win...
988 merged_notes.update_from(dest_notes);
989 return set_node_notes_at(dest->_idx, &merged_notes);
990 }
991
992
993 //--------------------------allow_range_check_smearing-------------------------
994 // Gating condition for coalescing similar range checks.
995 // Sometimes we try 'speculatively' replacing a series of a range checks by a
996 // single covering check that is at least as strong as any of them.
997 // If the optimization succeeds, the simplified (strengthened) range check
998 // will always succeed. If it fails, we will deopt, and then give up
999 // on the optimization.
1000 bool Compile::allow_range_check_smearing() const {
1001 // If this method has already thrown a range-check,
1002 // assume it was because we already tried range smearing
1003 // and it failed.
1004 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1005 return !already_trapped;
1006 }
1007
1008
1009 //------------------------------flatten_alias_type-----------------------------
1010 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1011 int offset = tj->offset();
1012 TypePtr::PTR ptr = tj->ptr();
1013
1014 // Known instance (scalarizable allocation) alias only with itself.
1015 bool is_known_inst = tj->isa_oopptr() != NULL &&
1016 tj->is_oopptr()->is_known_instance();
1017
1018 // Process weird unsafe references.
1019 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1020 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1021 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1022 tj = TypeOopPtr::BOTTOM;
1023 ptr = tj->ptr();
1024 offset = tj->offset();
1025 }
1026
1027 // Array pointers need some flattening
1028 const TypeAryPtr *ta = tj->isa_aryptr();
1029 if( ta && is_known_inst ) {
1030 if ( offset != Type::OffsetBot &&
1031 offset > arrayOopDesc::length_offset_in_bytes() ) {
1032 offset = Type::OffsetBot; // Flatten constant access into array body only
1033 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1034 }
1035 } else if( ta && _AliasLevel >= 2 ) {
1036 // For arrays indexed by constant indices, we flatten the alias
1037 // space to include all of the array body. Only the header, klass
1038 // and array length can be accessed un-aliased.
1039 if( offset != Type::OffsetBot ) {
1040 if( ta->const_oop() ) { // methodDataOop or methodOop
1041 offset = Type::OffsetBot; // Flatten constant access into array body
1042 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1043 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1044 // range is OK as-is.
1045 tj = ta = TypeAryPtr::RANGE;
1046 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1047 tj = TypeInstPtr::KLASS; // all klass loads look alike
1048 ta = TypeAryPtr::RANGE; // generic ignored junk
1049 ptr = TypePtr::BotPTR;
1050 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1051 tj = TypeInstPtr::MARK;
1052 ta = TypeAryPtr::RANGE; // generic ignored junk
1053 ptr = TypePtr::BotPTR;
1054 } else { // Random constant offset into array body
1055 offset = Type::OffsetBot; // Flatten constant access into array body
1056 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1057 }
1058 }
1059 // Arrays of fixed size alias with arrays of unknown size.
1060 if (ta->size() != TypeInt::POS) {
1061 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1062 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1063 }
1064 // Arrays of known objects become arrays of unknown objects.
1065 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1066 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1067 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1068 }
1069 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1070 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1071 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1072 }
1073 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1074 // cannot be distinguished by bytecode alone.
1075 if (ta->elem() == TypeInt::BOOL) {
1076 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1077 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1078 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1079 }
1080 // During the 2nd round of IterGVN, NotNull castings are removed.
1081 // Make sure the Bottom and NotNull variants alias the same.
1082 // Also, make sure exact and non-exact variants alias the same.
1083 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1084 if (ta->const_oop()) {
1085 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1086 } else {
1087 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1088 }
1089 }
1090 }
1091
1092 // Oop pointers need some flattening
1093 const TypeInstPtr *to = tj->isa_instptr();
1094 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1095 if( ptr == TypePtr::Constant ) {
1096 // No constant oop pointers (such as Strings); they alias with
1097 // unknown strings.
1098 assert(!is_known_inst, "not scalarizable allocation");
1099 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1100 } else if( is_known_inst ) {
1101 tj = to; // Keep NotNull and klass_is_exact for instance type
1102 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1103 // During the 2nd round of IterGVN, NotNull castings are removed.
1104 // Make sure the Bottom and NotNull variants alias the same.
1105 // Also, make sure exact and non-exact variants alias the same.
1106 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1107 }
1108 // Canonicalize the holder of this field
1109 ciInstanceKlass *k = to->klass()->as_instance_klass();
1110 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1111 // First handle header references such as a LoadKlassNode, even if the
1112 // object's klass is unloaded at compile time (4965979).
1113 if (!is_known_inst) { // Do it only for non-instance types
1114 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1115 }
1116 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1117 to = NULL;
1118 tj = TypeOopPtr::BOTTOM;
1119 offset = tj->offset();
1120 } else {
1121 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1122 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1123 if( is_known_inst ) {
1124 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1125 } else {
1126 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1127 }
1128 }
1129 }
1130 }
1131
1132 // Klass pointers to object array klasses need some flattening
1133 const TypeKlassPtr *tk = tj->isa_klassptr();
1134 if( tk ) {
1135 // If we are referencing a field within a Klass, we need
1136 // to assume the worst case of an Object. Both exact and
1137 // inexact types must flatten to the same alias class.
1138 // Since the flattened result for a klass is defined to be
1139 // precisely java.lang.Object, use a constant ptr.
1140 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1141
1142 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1143 TypeKlassPtr::OBJECT->klass(),
1144 offset);
1145 }
1146
1147 ciKlass* klass = tk->klass();
1148 if( klass->is_obj_array_klass() ) {
1149 ciKlass* k = TypeAryPtr::OOPS->klass();
1150 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1151 k = TypeInstPtr::BOTTOM->klass();
1152 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1153 }
1154
1155 // Check for precise loads from the primary supertype array and force them
1156 // to the supertype cache alias index. Check for generic array loads from
1157 // the primary supertype array and also force them to the supertype cache
1158 // alias index. Since the same load can reach both, we need to merge
1159 // these 2 disparate memories into the same alias class. Since the
1160 // primary supertype array is read-only, there's no chance of confusion
1161 // where we bypass an array load and an array store.
1162 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1163 if( offset == Type::OffsetBot ||
1164 off2 < Klass::primary_super_limit()*wordSize ) {
1165 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1166 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1167 }
1168 }
1169
1170 // Flatten all Raw pointers together.
1171 if (tj->base() == Type::RawPtr)
1172 tj = TypeRawPtr::BOTTOM;
1173
1174 if (tj->base() == Type::AnyPtr)
1175 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1176
1177 // Flatten all to bottom for now
1178 switch( _AliasLevel ) {
1179 case 0:
1180 tj = TypePtr::BOTTOM;
1181 break;
1182 case 1: // Flatten to: oop, static, field or array
1183 switch (tj->base()) {
1184 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1185 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1186 case Type::AryPtr: // do not distinguish arrays at all
1187 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1188 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1189 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1190 default: ShouldNotReachHere();
1191 }
1192 break;
1193 case 2: // No collasping at level 2; keep all splits
1194 case 3: // No collasping at level 3; keep all splits
1195 break;
1196 default:
1197 Unimplemented();
1198 }
1199
1200 offset = tj->offset();
1201 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1202
1203 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1204 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1205 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1206 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1207 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1208 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1209 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1210 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1211 assert( tj->ptr() != TypePtr::TopPTR &&
1212 tj->ptr() != TypePtr::AnyNull &&
1213 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1214 // assert( tj->ptr() != TypePtr::Constant ||
1215 // tj->base() == Type::RawPtr ||
1216 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1217
1218 return tj;
1219 }
1220
1221 void Compile::AliasType::Init(int i, const TypePtr* at) {
1222 _index = i;
1223 _adr_type = at;
1224 _field = NULL;
1225 _is_rewritable = true; // default
1226 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1227 if (atoop != NULL && atoop->is_known_instance()) {
1228 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1229 _general_index = Compile::current()->get_alias_index(gt);
1230 } else {
1231 _general_index = 0;
1232 }
1233 }
1234
1235 //---------------------------------print_on------------------------------------
1236 #ifndef PRODUCT
1237 void Compile::AliasType::print_on(outputStream* st) {
1238 if (index() < 10)
1239 st->print("@ <%d> ", index());
1240 else st->print("@ <%d>", index());
1241 st->print(is_rewritable() ? " " : " RO");
1242 int offset = adr_type()->offset();
1243 if (offset == Type::OffsetBot)
1244 st->print(" +any");
1245 else st->print(" +%-3d", offset);
1246 st->print(" in ");
1247 adr_type()->dump_on(st);
1248 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1249 if (field() != NULL && tjp) {
1250 if (tjp->klass() != field()->holder() ||
1251 tjp->offset() != field()->offset_in_bytes()) {
1252 st->print(" != ");
1253 field()->print();
1254 st->print(" ***");
1255 }
1256 }
1257 }
1258
1259 void print_alias_types() {
1260 Compile* C = Compile::current();
1261 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1262 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1263 C->alias_type(idx)->print_on(tty);
1264 tty->cr();
1265 }
1266 }
1267 #endif
1268
1269
1270 //----------------------------probe_alias_cache--------------------------------
1271 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1272 intptr_t key = (intptr_t) adr_type;
1273 key ^= key >> logAliasCacheSize;
1274 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1275 }
1276
1277
1278 //-----------------------------grow_alias_types--------------------------------
1279 void Compile::grow_alias_types() {
1280 const int old_ats = _max_alias_types; // how many before?
1281 const int new_ats = old_ats; // how many more?
1282 const int grow_ats = old_ats+new_ats; // how many now?
1283 _max_alias_types = grow_ats;
1284 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1285 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1286 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1287 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1288 }
1289
1290
1291 //--------------------------------find_alias_type------------------------------
1292 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1293 if (_AliasLevel == 0)
1294 return alias_type(AliasIdxBot);
1295
1296 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1297 if (ace->_adr_type == adr_type) {
1298 return alias_type(ace->_index);
1299 }
1300
1301 // Handle special cases.
1302 if (adr_type == NULL) return alias_type(AliasIdxTop);
1303 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1304
1305 // Do it the slow way.
1306 const TypePtr* flat = flatten_alias_type(adr_type);
1307
1308 #ifdef ASSERT
1309 assert(flat == flatten_alias_type(flat), "idempotent");
1310 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1311 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1312 const TypeOopPtr* foop = flat->is_oopptr();
1313 // Scalarizable allocations have exact klass always.
1314 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1315 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1316 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1317 }
1318 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1319 #endif
1320
1321 int idx = AliasIdxTop;
1322 for (int i = 0; i < num_alias_types(); i++) {
1323 if (alias_type(i)->adr_type() == flat) {
1324 idx = i;
1325 break;
1326 }
1327 }
1328
1329 if (idx == AliasIdxTop) {
1330 if (no_create) return NULL;
1331 // Grow the array if necessary.
1332 if (_num_alias_types == _max_alias_types) grow_alias_types();
1333 // Add a new alias type.
1334 idx = _num_alias_types++;
1335 _alias_types[idx]->Init(idx, flat);
1336 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1337 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1338 if (flat->isa_instptr()) {
1339 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1340 && flat->is_instptr()->klass() == env()->Class_klass())
1341 alias_type(idx)->set_rewritable(false);
1342 }
1343 if (flat->isa_klassptr()) {
1344 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1345 alias_type(idx)->set_rewritable(false);
1346 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1347 alias_type(idx)->set_rewritable(false);
1348 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1349 alias_type(idx)->set_rewritable(false);
1350 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1351 alias_type(idx)->set_rewritable(false);
1352 }
1353 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1354 // but the base pointer type is not distinctive enough to identify
1355 // references into JavaThread.)
1356
1357 // Check for final instance fields.
1358 const TypeInstPtr* tinst = flat->isa_instptr();
1359 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1360 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1361 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1362 // Set field() and is_rewritable() attributes.
1363 if (field != NULL) alias_type(idx)->set_field(field);
1364 }
1365 const TypeKlassPtr* tklass = flat->isa_klassptr();
1366 // Check for final static fields.
1367 if (tklass && tklass->klass()->is_instance_klass()) {
1368 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1369 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1370 // Set field() and is_rewritable() attributes.
1371 if (field != NULL) alias_type(idx)->set_field(field);
1372 }
1373 }
1374
1375 // Fill the cache for next time.
1376 ace->_adr_type = adr_type;
1377 ace->_index = idx;
1378 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1379
1380 // Might as well try to fill the cache for the flattened version, too.
1381 AliasCacheEntry* face = probe_alias_cache(flat);
1382 if (face->_adr_type == NULL) {
1383 face->_adr_type = flat;
1384 face->_index = idx;
1385 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1386 }
1387
1388 return alias_type(idx);
1389 }
1390
1391
1392 Compile::AliasType* Compile::alias_type(ciField* field) {
1393 const TypeOopPtr* t;
1394 if (field->is_static())
1395 t = TypeKlassPtr::make(field->holder());
1396 else
1397 t = TypeOopPtr::make_from_klass_raw(field->holder());
1398 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1399 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1400 return atp;
1401 }
1402
1403
1404 //------------------------------have_alias_type--------------------------------
1405 bool Compile::have_alias_type(const TypePtr* adr_type) {
1406 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1407 if (ace->_adr_type == adr_type) {
1408 return true;
1409 }
1410
1411 // Handle special cases.
1412 if (adr_type == NULL) return true;
1413 if (adr_type == TypePtr::BOTTOM) return true;
1414
1415 return find_alias_type(adr_type, true) != NULL;
1416 }
1417
1418 //-----------------------------must_alias--------------------------------------
1419 // True if all values of the given address type are in the given alias category.
1420 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1421 if (alias_idx == AliasIdxBot) return true; // the universal category
1422 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1423 if (alias_idx == AliasIdxTop) return false; // the empty category
1424 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1425
1426 // the only remaining possible overlap is identity
1427 int adr_idx = get_alias_index(adr_type);
1428 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1429 assert(adr_idx == alias_idx ||
1430 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1431 && adr_type != TypeOopPtr::BOTTOM),
1432 "should not be testing for overlap with an unsafe pointer");
1433 return adr_idx == alias_idx;
1434 }
1435
1436 //------------------------------can_alias--------------------------------------
1437 // True if any values of the given address type are in the given alias category.
1438 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1439 if (alias_idx == AliasIdxTop) return false; // the empty category
1440 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1441 if (alias_idx == AliasIdxBot) return true; // the universal category
1442 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1443
1444 // the only remaining possible overlap is identity
1445 int adr_idx = get_alias_index(adr_type);
1446 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1447 return adr_idx == alias_idx;
1448 }
1449
1450
1451
1452 //---------------------------pop_warm_call-------------------------------------
1453 WarmCallInfo* Compile::pop_warm_call() {
1454 WarmCallInfo* wci = _warm_calls;
1455 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1456 return wci;
1457 }
1458
1459 //----------------------------Inline_Warm--------------------------------------
1460 int Compile::Inline_Warm() {
1461 // If there is room, try to inline some more warm call sites.
1462 // %%% Do a graph index compaction pass when we think we're out of space?
1463 if (!InlineWarmCalls) return 0;
1464
1465 int calls_made_hot = 0;
1466 int room_to_grow = NodeCountInliningCutoff - unique();
1467 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1468 int amount_grown = 0;
1469 WarmCallInfo* call;
1470 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1471 int est_size = (int)call->size();
1472 if (est_size > (room_to_grow - amount_grown)) {
1473 // This one won't fit anyway. Get rid of it.
1474 call->make_cold();
1475 continue;
1476 }
1477 call->make_hot();
1478 calls_made_hot++;
1479 amount_grown += est_size;
1480 amount_to_grow -= est_size;
1481 }
1482
1483 if (calls_made_hot > 0) set_major_progress();
1484 return calls_made_hot;
1485 }
1486
1487
1488 //----------------------------Finish_Warm--------------------------------------
1489 void Compile::Finish_Warm() {
1490 if (!InlineWarmCalls) return;
1491 if (failing()) return;
1492 if (warm_calls() == NULL) return;
1493
1494 // Clean up loose ends, if we are out of space for inlining.
1495 WarmCallInfo* call;
1496 while ((call = pop_warm_call()) != NULL) {
1497 call->make_cold();
1498 }
1499 }
1500
1501
1502 //------------------------------Optimize---------------------------------------
1503 // Given a graph, optimize it.
1504 void Compile::Optimize() {
1505 TracePhase t1("optimizer", &_t_optimizer, true);
1506
1507 #ifndef PRODUCT
1508 if (env()->break_at_compile()) {
1509 BREAKPOINT;
1510 }
1511
1512 #endif
1513
1514 ResourceMark rm;
1515 int loop_opts_cnt;
1516
1517 NOT_PRODUCT( verify_graph_edges(); )
1518
1519 print_method("After Parsing");
1520
1521 {
1522 // Iterative Global Value Numbering, including ideal transforms
1523 // Initialize IterGVN with types and values from parse-time GVN
1524 PhaseIterGVN igvn(initial_gvn());
1525 {
1526 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1527 igvn.optimize();
1528 }
1529
1530 print_method("Iter GVN 1", 2);
1531
1532 if (failing()) return;
1533
1534 // get rid of the connection graph since it's information is not
1535 // updated by optimizations
1536 _congraph = NULL;
1537
1538
1539 // Loop transforms on the ideal graph. Range Check Elimination,
1540 // peeling, unrolling, etc.
1541
1542 // Set loop opts counter
1543 loop_opts_cnt = num_loop_opts();
1544 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1545 {
1546 TracePhase t2("idealLoop", &_t_idealLoop, true);
1547 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1548 loop_opts_cnt--;
1549 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1550 if (failing()) return;
1551 }
1552 // Loop opts pass if partial peeling occurred in previous pass
1553 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1554 TracePhase t3("idealLoop", &_t_idealLoop, true);
1555 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1556 loop_opts_cnt--;
1557 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1558 if (failing()) return;
1559 }
1560 // Loop opts pass for loop-unrolling before CCP
1561 if(major_progress() && (loop_opts_cnt > 0)) {
1562 TracePhase t4("idealLoop", &_t_idealLoop, true);
1563 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1564 loop_opts_cnt--;
1565 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1566 }
1567 }
1568 if (failing()) return;
1569
1570 // Conditional Constant Propagation;
1571 PhaseCCP ccp( &igvn );
1572 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1573 {
1574 TracePhase t2("ccp", &_t_ccp, true);
1575 ccp.do_transform();
1576 }
1577 print_method("PhaseCPP 1", 2);
1578
1579 assert( true, "Break here to ccp.dump_old2new_map()");
1580
1581 // Iterative Global Value Numbering, including ideal transforms
1582 {
1583 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1584 igvn = ccp;
1585 igvn.optimize();
1586 }
1587
1588 print_method("Iter GVN 2", 2);
1589
1590 if (failing()) return;
1591
1592 // Loop transforms on the ideal graph. Range Check Elimination,
1593 // peeling, unrolling, etc.
1594 if(loop_opts_cnt > 0) {
1595 debug_only( int cnt = 0; );
1596 while(major_progress() && (loop_opts_cnt > 0)) {
1597 TracePhase t2("idealLoop", &_t_idealLoop, true);
1598 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1599 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1600 loop_opts_cnt--;
1601 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1602 if (failing()) return;
1603 }
1604 }
1605 {
1606 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1607 PhaseMacroExpand mex(igvn);
1608 if (mex.expand_macro_nodes()) {
1609 assert(failing(), "must bail out w/ explicit message");
1610 return;
1611 }
1612 }
1613
1614 } // (End scope of igvn; run destructor if necessary for asserts.)
1615
1616 // A method with only infinite loops has no edges entering loops from root
1617 {
1618 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1619 if (final_graph_reshaping()) {
1620 assert(failing(), "must bail out w/ explicit message");
1621 return;
1622 }
1623 }
1624
1625 print_method("Optimize finished", 2);
1626 }
1627
1628
1629 //------------------------------Code_Gen---------------------------------------
1630 // Given a graph, generate code for it
1631 void Compile::Code_Gen() {
1632 if (failing()) return;
1633
1634 // Perform instruction selection. You might think we could reclaim Matcher
1635 // memory PDQ, but actually the Matcher is used in generating spill code.
1636 // Internals of the Matcher (including some VectorSets) must remain live
1637 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1638 // set a bit in reclaimed memory.
1639
1640 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1641 // nodes. Mapping is only valid at the root of each matched subtree.
1642 NOT_PRODUCT( verify_graph_edges(); )
1643
1644 Node_List proj_list;
1645 Matcher m(proj_list);
1646 _matcher = &m;
1647 {
1648 TracePhase t2("matcher", &_t_matcher, true);
1649 m.match();
1650 }
1651 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1652 // nodes. Mapping is only valid at the root of each matched subtree.
1653 NOT_PRODUCT( verify_graph_edges(); )
1654
1655 // If you have too many nodes, or if matching has failed, bail out
1656 check_node_count(0, "out of nodes matching instructions");
1657 if (failing()) return;
1658
1659 // Build a proper-looking CFG
1660 PhaseCFG cfg(node_arena(), root(), m);
1661 _cfg = &cfg;
1662 {
1663 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1664 cfg.Dominators();
1665 if (failing()) return;
1666
1667 NOT_PRODUCT( verify_graph_edges(); )
1668
1669 cfg.Estimate_Block_Frequency();
1670 cfg.GlobalCodeMotion(m,unique(),proj_list);
1671
1672 print_method("Global code motion", 2);
1673
1674 if (failing()) return;
1675 NOT_PRODUCT( verify_graph_edges(); )
1676
1677 debug_only( cfg.verify(); )
1678 }
1679 NOT_PRODUCT( verify_graph_edges(); )
1680
1681 PhaseChaitin regalloc(unique(),cfg,m);
1682 _regalloc = ®alloc;
1683 {
1684 TracePhase t2("regalloc", &_t_registerAllocation, true);
1685 // Perform any platform dependent preallocation actions. This is used,
1686 // for example, to avoid taking an implicit null pointer exception
1687 // using the frame pointer on win95.
1688 _regalloc->pd_preallocate_hook();
1689
1690 // Perform register allocation. After Chaitin, use-def chains are
1691 // no longer accurate (at spill code) and so must be ignored.
1692 // Node->LRG->reg mappings are still accurate.
1693 _regalloc->Register_Allocate();
1694
1695 // Bail out if the allocator builds too many nodes
1696 if (failing()) return;
1697 }
1698
1699 // Prior to register allocation we kept empty basic blocks in case the
1700 // the allocator needed a place to spill. After register allocation we
1701 // are not adding any new instructions. If any basic block is empty, we
1702 // can now safely remove it.
1703 {
1704 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); )
1705 cfg.RemoveEmpty();
1706 }
1707
1708 // Perform any platform dependent postallocation verifications.
1709 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1710
1711 // Apply peephole optimizations
1712 if( OptoPeephole ) {
1713 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1714 PhasePeephole peep( _regalloc, cfg);
1715 peep.do_transform();
1716 }
1717
1718 // Convert Nodes to instruction bits in a buffer
1719 {
1720 // %%%% workspace merge brought two timers together for one job
1721 TracePhase t2a("output", &_t_output, true);
1722 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1723 Output();
1724 }
1725
1726 print_method("Final Code");
1727
1728 // He's dead, Jim.
1729 _cfg = (PhaseCFG*)0xdeadbeef;
1730 _regalloc = (PhaseChaitin*)0xdeadbeef;
1731 }
1732
1733
1734 //------------------------------dump_asm---------------------------------------
1735 // Dump formatted assembly
1736 #ifndef PRODUCT
1737 void Compile::dump_asm(int *pcs, uint pc_limit) {
1738 bool cut_short = false;
1739 tty->print_cr("#");
1740 tty->print("# "); _tf->dump(); tty->cr();
1741 tty->print_cr("#");
1742
1743 // For all blocks
1744 int pc = 0x0; // Program counter
1745 char starts_bundle = ' ';
1746 _regalloc->dump_frame();
1747
1748 Node *n = NULL;
1749 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1750 if (VMThread::should_terminate()) { cut_short = true; break; }
1751 Block *b = _cfg->_blocks[i];
1752 if (b->is_connector() && !Verbose) continue;
1753 n = b->_nodes[0];
1754 if (pcs && n->_idx < pc_limit)
1755 tty->print("%3.3x ", pcs[n->_idx]);
1756 else
1757 tty->print(" ");
1758 b->dump_head( &_cfg->_bbs );
1759 if (b->is_connector()) {
1760 tty->print_cr(" # Empty connector block");
1761 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1762 tty->print_cr(" # Block is sole successor of call");
1763 }
1764
1765 // For all instructions
1766 Node *delay = NULL;
1767 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1768 if (VMThread::should_terminate()) { cut_short = true; break; }
1769 n = b->_nodes[j];
1770 if (valid_bundle_info(n)) {
1771 Bundle *bundle = node_bundling(n);
1772 if (bundle->used_in_unconditional_delay()) {
1773 delay = n;
1774 continue;
1775 }
1776 if (bundle->starts_bundle())
1777 starts_bundle = '+';
1778 }
1779
1780 if (WizardMode) n->dump();
1781
1782 if( !n->is_Region() && // Dont print in the Assembly
1783 !n->is_Phi() && // a few noisely useless nodes
1784 !n->is_Proj() &&
1785 !n->is_MachTemp() &&
1786 !n->is_Catch() && // Would be nice to print exception table targets
1787 !n->is_MergeMem() && // Not very interesting
1788 !n->is_top() && // Debug info table constants
1789 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1790 ) {
1791 if (pcs && n->_idx < pc_limit)
1792 tty->print("%3.3x", pcs[n->_idx]);
1793 else
1794 tty->print(" ");
1795 tty->print(" %c ", starts_bundle);
1796 starts_bundle = ' ';
1797 tty->print("\t");
1798 n->format(_regalloc, tty);
1799 tty->cr();
1800 }
1801
1802 // If we have an instruction with a delay slot, and have seen a delay,
1803 // then back up and print it
1804 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1805 assert(delay != NULL, "no unconditional delay instruction");
1806 if (WizardMode) delay->dump();
1807
1808 if (node_bundling(delay)->starts_bundle())
1809 starts_bundle = '+';
1810 if (pcs && n->_idx < pc_limit)
1811 tty->print("%3.3x", pcs[n->_idx]);
1812 else
1813 tty->print(" ");
1814 tty->print(" %c ", starts_bundle);
1815 starts_bundle = ' ';
1816 tty->print("\t");
1817 delay->format(_regalloc, tty);
1818 tty->print_cr("");
1819 delay = NULL;
1820 }
1821
1822 // Dump the exception table as well
1823 if( n->is_Catch() && (Verbose || WizardMode) ) {
1824 // Print the exception table for this offset
1825 _handler_table.print_subtable_for(pc);
1826 }
1827 }
1828
1829 if (pcs && n->_idx < pc_limit)
1830 tty->print_cr("%3.3x", pcs[n->_idx]);
1831 else
1832 tty->print_cr("");
1833
1834 assert(cut_short || delay == NULL, "no unconditional delay branch");
1835
1836 } // End of per-block dump
1837 tty->print_cr("");
1838
1839 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1840 }
1841 #endif
1842
1843 //------------------------------Final_Reshape_Counts---------------------------
1844 // This class defines counters to help identify when a method
1845 // may/must be executed using hardware with only 24-bit precision.
1846 struct Final_Reshape_Counts : public StackObj {
1847 int _call_count; // count non-inlined 'common' calls
1848 int _float_count; // count float ops requiring 24-bit precision
1849 int _double_count; // count double ops requiring more precision
1850 int _java_call_count; // count non-inlined 'java' calls
1851 VectorSet _visited; // Visitation flags
1852 Node_List _tests; // Set of IfNodes & PCTableNodes
1853
1854 Final_Reshape_Counts() :
1855 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0),
1856 _visited( Thread::current()->resource_area() ) { }
1857
1858 void inc_call_count () { _call_count ++; }
1859 void inc_float_count () { _float_count ++; }
1860 void inc_double_count() { _double_count++; }
1861 void inc_java_call_count() { _java_call_count++; }
1862
1863 int get_call_count () const { return _call_count ; }
1864 int get_float_count () const { return _float_count ; }
1865 int get_double_count() const { return _double_count; }
1866 int get_java_call_count() const { return _java_call_count; }
1867 };
1868
1869 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1870 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1871 // Make sure the offset goes inside the instance layout.
1872 return k->contains_field_offset(tp->offset());
1873 // Note that OffsetBot and OffsetTop are very negative.
1874 }
1875
1876 //------------------------------final_graph_reshaping_impl----------------------
1877 // Implement items 1-5 from final_graph_reshaping below.
1878 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
1879
1880 if ( n->outcnt() == 0 ) return; // dead node
1881 uint nop = n->Opcode();
1882
1883 // Check for 2-input instruction with "last use" on right input.
1884 // Swap to left input. Implements item (2).
1885 if( n->req() == 3 && // two-input instruction
1886 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1887 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1888 n->in(2)->outcnt() == 1 &&// right use IS a last use
1889 !n->in(2)->is_Con() ) { // right use is not a constant
1890 // Check for commutative opcode
1891 switch( nop ) {
1892 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1893 case Op_MaxI: case Op_MinI:
1894 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1895 case Op_AndL: case Op_XorL: case Op_OrL:
1896 case Op_AndI: case Op_XorI: case Op_OrI: {
1897 // Move "last use" input to left by swapping inputs
1898 n->swap_edges(1, 2);
1899 break;
1900 }
1901 default:
1902 break;
1903 }
1904 }
1905
1906 // Count FPU ops and common calls, implements item (3)
1907 switch( nop ) {
1908 // Count all float operations that may use FPU
1909 case Op_AddF:
1910 case Op_SubF:
1911 case Op_MulF:
1912 case Op_DivF:
1913 case Op_NegF:
1914 case Op_ModF:
1915 case Op_ConvI2F:
1916 case Op_ConF:
1917 case Op_CmpF:
1918 case Op_CmpF3:
1919 // case Op_ConvL2F: // longs are split into 32-bit halves
1920 fpu.inc_float_count();
1921 break;
1922
1923 case Op_ConvF2D:
1924 case Op_ConvD2F:
1925 fpu.inc_float_count();
1926 fpu.inc_double_count();
1927 break;
1928
1929 // Count all double operations that may use FPU
1930 case Op_AddD:
1931 case Op_SubD:
1932 case Op_MulD:
1933 case Op_DivD:
1934 case Op_NegD:
1935 case Op_ModD:
1936 case Op_ConvI2D:
1937 case Op_ConvD2I:
1938 // case Op_ConvL2D: // handled by leaf call
1939 // case Op_ConvD2L: // handled by leaf call
1940 case Op_ConD:
1941 case Op_CmpD:
1942 case Op_CmpD3:
1943 fpu.inc_double_count();
1944 break;
1945 case Op_Opaque1: // Remove Opaque Nodes before matching
1946 case Op_Opaque2: // Remove Opaque Nodes before matching
1947 n->subsume_by(n->in(1));
1948 break;
1949 case Op_CallStaticJava:
1950 case Op_CallJava:
1951 case Op_CallDynamicJava:
1952 fpu.inc_java_call_count(); // Count java call site;
1953 case Op_CallRuntime:
1954 case Op_CallLeaf:
1955 case Op_CallLeafNoFP: {
1956 assert( n->is_Call(), "" );
1957 CallNode *call = n->as_Call();
1958 // Count call sites where the FP mode bit would have to be flipped.
1959 // Do not count uncommon runtime calls:
1960 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
1961 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
1962 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
1963 fpu.inc_call_count(); // Count the call site
1964 } else { // See if uncommon argument is shared
1965 Node *n = call->in(TypeFunc::Parms);
1966 int nop = n->Opcode();
1967 // Clone shared simple arguments to uncommon calls, item (1).
1968 if( n->outcnt() > 1 &&
1969 !n->is_Proj() &&
1970 nop != Op_CreateEx &&
1971 nop != Op_CheckCastPP &&
1972 nop != Op_DecodeN &&
1973 !n->is_Mem() ) {
1974 Node *x = n->clone();
1975 call->set_req( TypeFunc::Parms, x );
1976 }
1977 }
1978 break;
1979 }
1980
1981 case Op_StoreD:
1982 case Op_LoadD:
1983 case Op_LoadD_unaligned:
1984 fpu.inc_double_count();
1985 goto handle_mem;
1986 case Op_StoreF:
1987 case Op_LoadF:
1988 fpu.inc_float_count();
1989 goto handle_mem;
1990
1991 case Op_StoreB:
1992 case Op_StoreC:
1993 case Op_StoreCM:
1994 case Op_StorePConditional:
1995 case Op_StoreI:
1996 case Op_StoreL:
1997 case Op_StoreLConditional:
1998 case Op_CompareAndSwapI:
1999 case Op_CompareAndSwapL:
2000 case Op_CompareAndSwapP:
2001 case Op_CompareAndSwapN:
2002 case Op_StoreP:
2003 case Op_StoreN:
2004 case Op_LoadB:
2005 case Op_LoadC:
2006 case Op_LoadI:
2007 case Op_LoadKlass:
2008 case Op_LoadNKlass:
2009 case Op_LoadL:
2010 case Op_LoadL_unaligned:
2011 case Op_LoadPLocked:
2012 case Op_LoadLLocked:
2013 case Op_LoadP:
2014 case Op_LoadN:
2015 case Op_LoadRange:
2016 case Op_LoadS: {
2017 handle_mem:
2018 #ifdef ASSERT
2019 if( VerifyOptoOopOffsets ) {
2020 assert( n->is_Mem(), "" );
2021 MemNode *mem = (MemNode*)n;
2022 // Check to see if address types have grounded out somehow.
2023 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2024 assert( !tp || oop_offset_is_sane(tp), "" );
2025 }
2026 #endif
2027 break;
2028 }
2029
2030 case Op_AddP: { // Assert sane base pointers
2031 Node *addp = n->in(AddPNode::Address);
2032 assert( !addp->is_AddP() ||
2033 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2034 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2035 "Base pointers must match" );
2036 #ifdef _LP64
2037 if (UseCompressedOops &&
2038 addp->Opcode() == Op_ConP &&
2039 addp == n->in(AddPNode::Base) &&
2040 n->in(AddPNode::Offset)->is_Con()) {
2041 // Use addressing with narrow klass to load with offset on x86.
2042 // On sparc loading 32-bits constant and decoding it have less
2043 // instructions (4) then load 64-bits constant (7).
2044 // Do this transformation here since IGVN will convert ConN back to ConP.
2045 const Type* t = addp->bottom_type();
2046 if (t->isa_oopptr()) {
2047 Node* nn = NULL;
2048
2049 // Look for existing ConN node of the same exact type.
2050 Compile* C = Compile::current();
2051 Node* r = C->root();
2052 uint cnt = r->outcnt();
2053 for (uint i = 0; i < cnt; i++) {
2054 Node* m = r->raw_out(i);
2055 if (m!= NULL && m->Opcode() == Op_ConN &&
2056 m->bottom_type()->make_ptr() == t) {
2057 nn = m;
2058 break;
2059 }
2060 }
2061 if (nn != NULL) {
2062 // Decode a narrow oop to match address
2063 // [R12 + narrow_oop_reg<<3 + offset]
2064 nn = new (C, 2) DecodeNNode(nn, t);
2065 n->set_req(AddPNode::Base, nn);
2066 n->set_req(AddPNode::Address, nn);
2067 if (addp->outcnt() == 0) {
2068 addp->disconnect_inputs(NULL);
2069 }
2070 }
2071 }
2072 }
2073 #endif
2074 break;
2075 }
2076
2077 #ifdef _LP64
2078 case Op_CastPP:
2079 if (n->in(1)->is_DecodeN() && UseImplicitNullCheckForNarrowOop) {
2080 Compile* C = Compile::current();
2081 Node* in1 = n->in(1);
2082 const Type* t = n->bottom_type();
2083 Node* new_in1 = in1->clone();
2084 new_in1->as_DecodeN()->set_type(t);
2085
2086 if (!Matcher::clone_shift_expressions) {
2087 //
2088 // x86, ARM and friends can handle 2 adds in addressing mode
2089 // and Matcher can fold a DecodeN node into address by using
2090 // a narrow oop directly and do implicit NULL check in address:
2091 //
2092 // [R12 + narrow_oop_reg<<3 + offset]
2093 // NullCheck narrow_oop_reg
2094 //
2095 // On other platforms (Sparc) we have to keep new DecodeN node and
2096 // use it to do implicit NULL check in address:
2097 //
2098 // decode_not_null narrow_oop_reg, base_reg
2099 // [base_reg + offset]
2100 // NullCheck base_reg
2101 //
2102 // Pin the new DecodeN node to non-null path on these patforms (Sparc)
2103 // to keep the information to which NULL check the new DecodeN node
2104 // corresponds to use it as value in implicit_null_check().
2105 //
2106 new_in1->set_req(0, n->in(0));
2107 }
2108
2109 n->subsume_by(new_in1);
2110 if (in1->outcnt() == 0) {
2111 in1->disconnect_inputs(NULL);
2112 }
2113 }
2114 break;
2115
2116 case Op_CmpP:
2117 // Do this transformation here to preserve CmpPNode::sub() and
2118 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2119 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2120 Node* in1 = n->in(1);
2121 Node* in2 = n->in(2);
2122 if (!in1->is_DecodeN()) {
2123 in2 = in1;
2124 in1 = n->in(2);
2125 }
2126 assert(in1->is_DecodeN(), "sanity");
2127
2128 Compile* C = Compile::current();
2129 Node* new_in2 = NULL;
2130 if (in2->is_DecodeN()) {
2131 new_in2 = in2->in(1);
2132 } else if (in2->Opcode() == Op_ConP) {
2133 const Type* t = in2->bottom_type();
2134 if (t == TypePtr::NULL_PTR && UseImplicitNullCheckForNarrowOop) {
2135 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2136 //
2137 // This transformation together with CastPP transformation above
2138 // will generated code for implicit NULL checks for compressed oops.
2139 //
2140 // The original code after Optimize()
2141 //
2142 // LoadN memory, narrow_oop_reg
2143 // decode narrow_oop_reg, base_reg
2144 // CmpP base_reg, NULL
2145 // CastPP base_reg // NotNull
2146 // Load [base_reg + offset], val_reg
2147 //
2148 // after these transformations will be
2149 //
2150 // LoadN memory, narrow_oop_reg
2151 // CmpN narrow_oop_reg, NULL
2152 // decode_not_null narrow_oop_reg, base_reg
2153 // Load [base_reg + offset], val_reg
2154 //
2155 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2156 // since narrow oops can be used in debug info now (see the code in
2157 // final_graph_reshaping_walk()).
2158 //
2159 // At the end the code will be matched to
2160 // on x86:
2161 //
2162 // Load_narrow_oop memory, narrow_oop_reg
2163 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2164 // NullCheck narrow_oop_reg
2165 //
2166 // and on sparc:
2167 //
2168 // Load_narrow_oop memory, narrow_oop_reg
2169 // decode_not_null narrow_oop_reg, base_reg
2170 // Load [base_reg + offset], val_reg
2171 // NullCheck base_reg
2172 //
2173 } else if (t->isa_oopptr()) {
2174 new_in2 = ConNode::make(C, t->make_narrowoop());
2175 }
2176 }
2177 if (new_in2 != NULL) {
2178 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2179 n->subsume_by( cmpN );
2180 if (in1->outcnt() == 0) {
2181 in1->disconnect_inputs(NULL);
2182 }
2183 if (in2->outcnt() == 0) {
2184 in2->disconnect_inputs(NULL);
2185 }
2186 }
2187 }
2188 break;
2189
2190 case Op_DecodeN:
2191 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2192 break;
2193
2194 case Op_EncodeP: {
2195 Node* in1 = n->in(1);
2196 if (in1->is_DecodeN()) {
2197 n->subsume_by(in1->in(1));
2198 } else if (in1->Opcode() == Op_ConP) {
2199 Compile* C = Compile::current();
2200 const Type* t = in1->bottom_type();
2201 if (t == TypePtr::NULL_PTR) {
2202 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2203 } else if (t->isa_oopptr()) {
2204 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2205 }
2206 }
2207 if (in1->outcnt() == 0) {
2208 in1->disconnect_inputs(NULL);
2209 }
2210 break;
2211 }
2212
2213 case Op_Phi:
2214 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2215 // The EncodeP optimization may create Phi with the same edges
2216 // for all paths. It is not handled well by Register Allocator.
2217 Node* unique_in = n->in(1);
2218 assert(unique_in != NULL, "");
2219 uint cnt = n->req();
2220 for (uint i = 2; i < cnt; i++) {
2221 Node* m = n->in(i);
2222 assert(m != NULL, "");
2223 if (unique_in != m)
2224 unique_in = NULL;
2225 }
2226 if (unique_in != NULL) {
2227 n->subsume_by(unique_in);
2228 }
2229 }
2230 break;
2231
2232 #endif
2233
2234 case Op_ModI:
2235 if (UseDivMod) {
2236 // Check if a%b and a/b both exist
2237 Node* d = n->find_similar(Op_DivI);
2238 if (d) {
2239 // Replace them with a fused divmod if supported
2240 Compile* C = Compile::current();
2241 if (Matcher::has_match_rule(Op_DivModI)) {
2242 DivModINode* divmod = DivModINode::make(C, n);
2243 d->subsume_by(divmod->div_proj());
2244 n->subsume_by(divmod->mod_proj());
2245 } else {
2246 // replace a%b with a-((a/b)*b)
2247 Node* mult = new (C, 3) MulINode(d, d->in(2));
2248 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2249 n->subsume_by( sub );
2250 }
2251 }
2252 }
2253 break;
2254
2255 case Op_ModL:
2256 if (UseDivMod) {
2257 // Check if a%b and a/b both exist
2258 Node* d = n->find_similar(Op_DivL);
2259 if (d) {
2260 // Replace them with a fused divmod if supported
2261 Compile* C = Compile::current();
2262 if (Matcher::has_match_rule(Op_DivModL)) {
2263 DivModLNode* divmod = DivModLNode::make(C, n);
2264 d->subsume_by(divmod->div_proj());
2265 n->subsume_by(divmod->mod_proj());
2266 } else {
2267 // replace a%b with a-((a/b)*b)
2268 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2269 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2270 n->subsume_by( sub );
2271 }
2272 }
2273 }
2274 break;
2275
2276 case Op_Load16B:
2277 case Op_Load8B:
2278 case Op_Load4B:
2279 case Op_Load8S:
2280 case Op_Load4S:
2281 case Op_Load2S:
2282 case Op_Load8C:
2283 case Op_Load4C:
2284 case Op_Load2C:
2285 case Op_Load4I:
2286 case Op_Load2I:
2287 case Op_Load2L:
2288 case Op_Load4F:
2289 case Op_Load2F:
2290 case Op_Load2D:
2291 case Op_Store16B:
2292 case Op_Store8B:
2293 case Op_Store4B:
2294 case Op_Store8C:
2295 case Op_Store4C:
2296 case Op_Store2C:
2297 case Op_Store4I:
2298 case Op_Store2I:
2299 case Op_Store2L:
2300 case Op_Store4F:
2301 case Op_Store2F:
2302 case Op_Store2D:
2303 break;
2304
2305 case Op_PackB:
2306 case Op_PackS:
2307 case Op_PackC:
2308 case Op_PackI:
2309 case Op_PackF:
2310 case Op_PackL:
2311 case Op_PackD:
2312 if (n->req()-1 > 2) {
2313 // Replace many operand PackNodes with a binary tree for matching
2314 PackNode* p = (PackNode*) n;
2315 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2316 n->subsume_by(btp);
2317 }
2318 break;
2319 default:
2320 assert( !n->is_Call(), "" );
2321 assert( !n->is_Mem(), "" );
2322 break;
2323 }
2324
2325 // Collect CFG split points
2326 if (n->is_MultiBranch())
2327 fpu._tests.push(n);
2328 }
2329
2330 //------------------------------final_graph_reshaping_walk---------------------
2331 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2332 // requires that the walk visits a node's inputs before visiting the node.
2333 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) {
2334 ResourceArea *area = Thread::current()->resource_area();
2335 Unique_Node_List sfpt(area);
2336
2337 fpu._visited.set(root->_idx); // first, mark node as visited
2338 uint cnt = root->req();
2339 Node *n = root;
2340 uint i = 0;
2341 while (true) {
2342 if (i < cnt) {
2343 // Place all non-visited non-null inputs onto stack
2344 Node* m = n->in(i);
2345 ++i;
2346 if (m != NULL && !fpu._visited.test_set(m->_idx)) {
2347 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2348 sfpt.push(m);
2349 cnt = m->req();
2350 nstack.push(n, i); // put on stack parent and next input's index
2351 n = m;
2352 i = 0;
2353 }
2354 } else {
2355 // Now do post-visit work
2356 final_graph_reshaping_impl( n, fpu );
2357 if (nstack.is_empty())
2358 break; // finished
2359 n = nstack.node(); // Get node from stack
2360 cnt = n->req();
2361 i = nstack.index();
2362 nstack.pop(); // Shift to the next node on stack
2363 }
2364 }
2365
2366 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2367 // It could be done for an uncommon traps or any safepoints/calls
2368 // if the DecodeN node is referenced only in a debug info.
2369 while (sfpt.size() > 0) {
2370 n = sfpt.pop();
2371 JVMState *jvms = n->as_SafePoint()->jvms();
2372 assert(jvms != NULL, "sanity");
2373 int start = jvms->debug_start();
2374 int end = n->req();
2375 bool is_uncommon = (n->is_CallStaticJava() &&
2376 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2377 for (int j = start; j < end; j++) {
2378 Node* in = n->in(j);
2379 if (in->is_DecodeN()) {
2380 bool safe_to_skip = true;
2381 if (!is_uncommon ) {
2382 // Is it safe to skip?
2383 for (uint i = 0; i < in->outcnt(); i++) {
2384 Node* u = in->raw_out(i);
2385 if (!u->is_SafePoint() ||
2386 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2387 safe_to_skip = false;
2388 }
2389 }
2390 }
2391 if (safe_to_skip) {
2392 n->set_req(j, in->in(1));
2393 }
2394 if (in->outcnt() == 0) {
2395 in->disconnect_inputs(NULL);
2396 }
2397 }
2398 }
2399 }
2400 }
2401
2402 //------------------------------final_graph_reshaping--------------------------
2403 // Final Graph Reshaping.
2404 //
2405 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2406 // and not commoned up and forced early. Must come after regular
2407 // optimizations to avoid GVN undoing the cloning. Clone constant
2408 // inputs to Loop Phis; these will be split by the allocator anyways.
2409 // Remove Opaque nodes.
2410 // (2) Move last-uses by commutative operations to the left input to encourage
2411 // Intel update-in-place two-address operations and better register usage
2412 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2413 // calls canonicalizing them back.
2414 // (3) Count the number of double-precision FP ops, single-precision FP ops
2415 // and call sites. On Intel, we can get correct rounding either by
2416 // forcing singles to memory (requires extra stores and loads after each
2417 // FP bytecode) or we can set a rounding mode bit (requires setting and
2418 // clearing the mode bit around call sites). The mode bit is only used
2419 // if the relative frequency of single FP ops to calls is low enough.
2420 // This is a key transform for SPEC mpeg_audio.
2421 // (4) Detect infinite loops; blobs of code reachable from above but not
2422 // below. Several of the Code_Gen algorithms fail on such code shapes,
2423 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2424 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2425 // Detection is by looking for IfNodes where only 1 projection is
2426 // reachable from below or CatchNodes missing some targets.
2427 // (5) Assert for insane oop offsets in debug mode.
2428
2429 bool Compile::final_graph_reshaping() {
2430 // an infinite loop may have been eliminated by the optimizer,
2431 // in which case the graph will be empty.
2432 if (root()->req() == 1) {
2433 record_method_not_compilable("trivial infinite loop");
2434 return true;
2435 }
2436
2437 Final_Reshape_Counts fpu;
2438
2439 // Visit everybody reachable!
2440 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2441 Node_Stack nstack(unique() >> 1);
2442 final_graph_reshaping_walk(nstack, root(), fpu);
2443
2444 // Check for unreachable (from below) code (i.e., infinite loops).
2445 for( uint i = 0; i < fpu._tests.size(); i++ ) {
2446 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch();
2447 // Get number of CFG targets.
2448 // Note that PCTables include exception targets after calls.
2449 uint required_outcnt = n->required_outcnt();
2450 if (n->outcnt() != required_outcnt) {
2451 // Check for a few special cases. Rethrow Nodes never take the
2452 // 'fall-thru' path, so expected kids is 1 less.
2453 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2454 if (n->in(0)->in(0)->is_Call()) {
2455 CallNode *call = n->in(0)->in(0)->as_Call();
2456 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2457 required_outcnt--; // Rethrow always has 1 less kid
2458 } else if (call->req() > TypeFunc::Parms &&
2459 call->is_CallDynamicJava()) {
2460 // Check for null receiver. In such case, the optimizer has
2461 // detected that the virtual call will always result in a null
2462 // pointer exception. The fall-through projection of this CatchNode
2463 // will not be populated.
2464 Node *arg0 = call->in(TypeFunc::Parms);
2465 if (arg0->is_Type() &&
2466 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2467 required_outcnt--;
2468 }
2469 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2470 call->req() > TypeFunc::Parms+1 &&
2471 call->is_CallStaticJava()) {
2472 // Check for negative array length. In such case, the optimizer has
2473 // detected that the allocation attempt will always result in an
2474 // exception. There is no fall-through projection of this CatchNode .
2475 Node *arg1 = call->in(TypeFunc::Parms+1);
2476 if (arg1->is_Type() &&
2477 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2478 required_outcnt--;
2479 }
2480 }
2481 }
2482 }
2483 // Recheck with a better notion of 'required_outcnt'
2484 if (n->outcnt() != required_outcnt) {
2485 record_method_not_compilable("malformed control flow");
2486 return true; // Not all targets reachable!
2487 }
2488 }
2489 // Check that I actually visited all kids. Unreached kids
2490 // must be infinite loops.
2491 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2492 if (!fpu._visited.test(n->fast_out(j)->_idx)) {
2493 record_method_not_compilable("infinite loop");
2494 return true; // Found unvisited kid; must be unreach
2495 }
2496 }
2497
2498 // If original bytecodes contained a mixture of floats and doubles
2499 // check if the optimizer has made it homogenous, item (3).
2500 if( Use24BitFPMode && Use24BitFP &&
2501 fpu.get_float_count() > 32 &&
2502 fpu.get_double_count() == 0 &&
2503 (10 * fpu.get_call_count() < fpu.get_float_count()) ) {
2504 set_24_bit_selection_and_mode( false, true );
2505 }
2506
2507 set_has_java_calls(fpu.get_java_call_count() > 0);
2508
2509 // No infinite loops, no reason to bail out.
2510 return false;
2511 }
2512
2513 //-----------------------------too_many_traps----------------------------------
2514 // Report if there are too many traps at the current method and bci.
2515 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2516 bool Compile::too_many_traps(ciMethod* method,
2517 int bci,
2518 Deoptimization::DeoptReason reason) {
2519 ciMethodData* md = method->method_data();
2520 if (md->is_empty()) {
2521 // Assume the trap has not occurred, or that it occurred only
2522 // because of a transient condition during start-up in the interpreter.
2523 return false;
2524 }
2525 if (md->has_trap_at(bci, reason) != 0) {
2526 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2527 // Also, if there are multiple reasons, or if there is no per-BCI record,
2528 // assume the worst.
2529 if (log())
2530 log()->elem("observe trap='%s' count='%d'",
2531 Deoptimization::trap_reason_name(reason),
2532 md->trap_count(reason));
2533 return true;
2534 } else {
2535 // Ignore method/bci and see if there have been too many globally.
2536 return too_many_traps(reason, md);
2537 }
2538 }
2539
2540 // Less-accurate variant which does not require a method and bci.
2541 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2542 ciMethodData* logmd) {
2543 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2544 // Too many traps globally.
2545 // Note that we use cumulative trap_count, not just md->trap_count.
2546 if (log()) {
2547 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2548 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2549 Deoptimization::trap_reason_name(reason),
2550 mcount, trap_count(reason));
2551 }
2552 return true;
2553 } else {
2554 // The coast is clear.
2555 return false;
2556 }
2557 }
2558
2559 //--------------------------too_many_recompiles--------------------------------
2560 // Report if there are too many recompiles at the current method and bci.
2561 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2562 // Is not eager to return true, since this will cause the compiler to use
2563 // Action_none for a trap point, to avoid too many recompilations.
2564 bool Compile::too_many_recompiles(ciMethod* method,
2565 int bci,
2566 Deoptimization::DeoptReason reason) {
2567 ciMethodData* md = method->method_data();
2568 if (md->is_empty()) {
2569 // Assume the trap has not occurred, or that it occurred only
2570 // because of a transient condition during start-up in the interpreter.
2571 return false;
2572 }
2573 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2574 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2575 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2576 Deoptimization::DeoptReason per_bc_reason
2577 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2578 if ((per_bc_reason == Deoptimization::Reason_none
2579 || md->has_trap_at(bci, reason) != 0)
2580 // The trap frequency measure we care about is the recompile count:
2581 && md->trap_recompiled_at(bci)
2582 && md->overflow_recompile_count() >= bc_cutoff) {
2583 // Do not emit a trap here if it has already caused recompilations.
2584 // Also, if there are multiple reasons, or if there is no per-BCI record,
2585 // assume the worst.
2586 if (log())
2587 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2588 Deoptimization::trap_reason_name(reason),
2589 md->trap_count(reason),
2590 md->overflow_recompile_count());
2591 return true;
2592 } else if (trap_count(reason) != 0
2593 && decompile_count() >= m_cutoff) {
2594 // Too many recompiles globally, and we have seen this sort of trap.
2595 // Use cumulative decompile_count, not just md->decompile_count.
2596 if (log())
2597 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2598 Deoptimization::trap_reason_name(reason),
2599 md->trap_count(reason), trap_count(reason),
2600 md->decompile_count(), decompile_count());
2601 return true;
2602 } else {
2603 // The coast is clear.
2604 return false;
2605 }
2606 }
2607
2608
2609 #ifndef PRODUCT
2610 //------------------------------verify_graph_edges---------------------------
2611 // Walk the Graph and verify that there is a one-to-one correspondence
2612 // between Use-Def edges and Def-Use edges in the graph.
2613 void Compile::verify_graph_edges(bool no_dead_code) {
2614 if (VerifyGraphEdges) {
2615 ResourceArea *area = Thread::current()->resource_area();
2616 Unique_Node_List visited(area);
2617 // Call recursive graph walk to check edges
2618 _root->verify_edges(visited);
2619 if (no_dead_code) {
2620 // Now make sure that no visited node is used by an unvisited node.
2621 bool dead_nodes = 0;
2622 Unique_Node_List checked(area);
2623 while (visited.size() > 0) {
2624 Node* n = visited.pop();
2625 checked.push(n);
2626 for (uint i = 0; i < n->outcnt(); i++) {
2627 Node* use = n->raw_out(i);
2628 if (checked.member(use)) continue; // already checked
2629 if (visited.member(use)) continue; // already in the graph
2630 if (use->is_Con()) continue; // a dead ConNode is OK
2631 // At this point, we have found a dead node which is DU-reachable.
2632 if (dead_nodes++ == 0)
2633 tty->print_cr("*** Dead nodes reachable via DU edges:");
2634 use->dump(2);
2635 tty->print_cr("---");
2636 checked.push(use); // No repeats; pretend it is now checked.
2637 }
2638 }
2639 assert(dead_nodes == 0, "using nodes must be reachable from root");
2640 }
2641 }
2642 }
2643 #endif
2644
2645 // The Compile object keeps track of failure reasons separately from the ciEnv.
2646 // This is required because there is not quite a 1-1 relation between the
2647 // ciEnv and its compilation task and the Compile object. Note that one
2648 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2649 // to backtrack and retry without subsuming loads. Other than this backtracking
2650 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2651 // by the logic in C2Compiler.
2652 void Compile::record_failure(const char* reason) {
2653 if (log() != NULL) {
2654 log()->elem("failure reason='%s' phase='compile'", reason);
2655 }
2656 if (_failure_reason == NULL) {
2657 // Record the first failure reason.
2658 _failure_reason = reason;
2659 }
2660 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2661 C->print_method(_failure_reason);
2662 }
2663 _root = NULL; // flush the graph, too
2664 }
2665
2666 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2667 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2668 {
2669 if (dolog) {
2670 C = Compile::current();
2671 _log = C->log();
2672 } else {
2673 C = NULL;
2674 _log = NULL;
2675 }
2676 if (_log != NULL) {
2677 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2678 _log->stamp();
2679 _log->end_head();
2680 }
2681 }
2682
2683 Compile::TracePhase::~TracePhase() {
2684 if (_log != NULL) {
2685 _log->done("phase nodes='%d'", C->unique());
2686 }
2687 }