1 #ifdef USE_PRAGMA_IDENT_SRC
2 #pragma ident "@(#)compile.cpp 1.633 07/09/28 10:23:11 JVM"
3 #endif
4 /*
5 * Copyright 1997-2008 Sun Microsystems, Inc. All Rights Reserved.
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7 *
8 * This code is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License version 2 only, as
10 * published by the Free Software Foundation.
11 *
12 * This code is distributed in the hope that it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 * version 2 for more details (a copy is included in the LICENSE file that
16 * accompanied this code).
17 *
18 * You should have received a copy of the GNU General Public License version
19 * 2 along with this work; if not, write to the Free Software Foundation,
20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
21 *
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
24 * have any questions.
25 *
26 */
27
28 #include "incls/_precompiled.incl"
29 #include "incls/_compile.cpp.incl"
30
31 /// Support for intrinsics.
32
33 // Return the index at which m must be inserted (or already exists).
34 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
35 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
36 #ifdef ASSERT
37 for (int i = 1; i < _intrinsics->length(); i++) {
38 CallGenerator* cg1 = _intrinsics->at(i-1);
39 CallGenerator* cg2 = _intrinsics->at(i);
40 assert(cg1->method() != cg2->method()
41 ? cg1->method() < cg2->method()
42 : cg1->is_virtual() < cg2->is_virtual(),
43 "compiler intrinsics list must stay sorted");
44 }
45 #endif
46 // Binary search sorted list, in decreasing intervals [lo, hi].
47 int lo = 0, hi = _intrinsics->length()-1;
48 while (lo <= hi) {
49 int mid = (uint)(hi + lo) / 2;
50 ciMethod* mid_m = _intrinsics->at(mid)->method();
51 if (m < mid_m) {
52 hi = mid-1;
53 } else if (m > mid_m) {
54 lo = mid+1;
55 } else {
56 // look at minor sort key
57 bool mid_virt = _intrinsics->at(mid)->is_virtual();
58 if (is_virtual < mid_virt) {
59 hi = mid-1;
60 } else if (is_virtual > mid_virt) {
61 lo = mid+1;
62 } else {
63 return mid; // exact match
64 }
65 }
66 }
67 return lo; // inexact match
68 }
69
70 void Compile::register_intrinsic(CallGenerator* cg) {
71 if (_intrinsics == NULL) {
72 _intrinsics = new GrowableArray<CallGenerator*>(60);
73 }
74 // This code is stolen from ciObjectFactory::insert.
75 // Really, GrowableArray should have methods for
76 // insert_at, remove_at, and binary_search.
77 int len = _intrinsics->length();
78 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
79 if (index == len) {
80 _intrinsics->append(cg);
81 } else {
82 #ifdef ASSERT
83 CallGenerator* oldcg = _intrinsics->at(index);
84 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
85 #endif
86 _intrinsics->append(_intrinsics->at(len-1));
87 int pos;
88 for (pos = len-2; pos >= index; pos--) {
89 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
90 }
91 _intrinsics->at_put(index, cg);
92 }
93 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
94 }
95
96 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
97 assert(m->is_loaded(), "don't try this on unloaded methods");
98 if (_intrinsics != NULL) {
99 int index = intrinsic_insertion_index(m, is_virtual);
100 if (index < _intrinsics->length()
101 && _intrinsics->at(index)->method() == m
102 && _intrinsics->at(index)->is_virtual() == is_virtual) {
103 return _intrinsics->at(index);
104 }
105 }
106 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
107 if (m->intrinsic_id() != vmIntrinsics::_none) {
108 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
109 if (cg != NULL) {
110 // Save it for next time:
111 register_intrinsic(cg);
112 return cg;
113 } else {
114 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
115 }
116 }
117 return NULL;
118 }
119
120 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
121 // in library_call.cpp.
122
123
124 #ifndef PRODUCT
125 // statistics gathering...
126
127 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
128 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
129
130 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
131 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
132 int oflags = _intrinsic_hist_flags[id];
133 assert(flags != 0, "what happened?");
134 if (is_virtual) {
135 flags |= _intrinsic_virtual;
136 }
137 bool changed = (flags != oflags);
138 if ((flags & _intrinsic_worked) != 0) {
139 juint count = (_intrinsic_hist_count[id] += 1);
140 if (count == 1) {
141 changed = true; // first time
142 }
143 // increment the overall count also:
144 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
145 }
146 if (changed) {
147 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
148 // Something changed about the intrinsic's virtuality.
149 if ((flags & _intrinsic_virtual) != 0) {
150 // This is the first use of this intrinsic as a virtual call.
151 if (oflags != 0) {
152 // We already saw it as a non-virtual, so note both cases.
153 flags |= _intrinsic_both;
154 }
155 } else if ((oflags & _intrinsic_both) == 0) {
156 // This is the first use of this intrinsic as a non-virtual
157 flags |= _intrinsic_both;
158 }
159 }
160 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
161 }
162 // update the overall flags also:
163 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
164 return changed;
165 }
166
167 static char* format_flags(int flags, char* buf) {
168 buf[0] = 0;
169 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
170 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
171 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
172 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
173 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
174 if (buf[0] == 0) strcat(buf, ",");
175 assert(buf[0] == ',', "must be");
176 return &buf[1];
177 }
178
179 void Compile::print_intrinsic_statistics() {
180 char flagsbuf[100];
181 ttyLocker ttyl;
182 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
183 tty->print_cr("Compiler intrinsic usage:");
184 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
185 if (total == 0) total = 1; // avoid div0 in case of no successes
186 #define PRINT_STAT_LINE(name, c, f) \
187 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
188 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
189 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
190 int flags = _intrinsic_hist_flags[id];
191 juint count = _intrinsic_hist_count[id];
192 if ((flags | count) != 0) {
193 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
194 }
195 }
196 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
197 if (xtty != NULL) xtty->tail("statistics");
198 }
199
200 void Compile::print_statistics() {
201 { ttyLocker ttyl;
202 if (xtty != NULL) xtty->head("statistics type='opto'");
203 Parse::print_statistics();
204 PhaseCCP::print_statistics();
205 PhaseRegAlloc::print_statistics();
206 Scheduling::print_statistics();
207 PhasePeephole::print_statistics();
208 PhaseIdealLoop::print_statistics();
209 if (xtty != NULL) xtty->tail("statistics");
210 }
211 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
212 // put this under its own <statistics> element.
213 print_intrinsic_statistics();
214 }
215 }
216 #endif //PRODUCT
217
218 // Support for bundling info
219 Bundle* Compile::node_bundling(const Node *n) {
220 assert(valid_bundle_info(n), "oob");
221 return &_node_bundling_base[n->_idx];
222 }
223
224 bool Compile::valid_bundle_info(const Node *n) {
225 return (_node_bundling_limit > n->_idx);
226 }
227
228
229 // Identify all nodes that are reachable from below, useful.
230 // Use breadth-first pass that records state in a Unique_Node_List,
231 // recursive traversal is slower.
232 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
233 int estimated_worklist_size = unique();
234 useful.map( estimated_worklist_size, NULL ); // preallocate space
235
236 // Initialize worklist
237 if (root() != NULL) { useful.push(root()); }
238 // If 'top' is cached, declare it useful to preserve cached node
239 if( cached_top_node() ) { useful.push(cached_top_node()); }
240
241 // Push all useful nodes onto the list, breadthfirst
242 for( uint next = 0; next < useful.size(); ++next ) {
243 assert( next < unique(), "Unique useful nodes < total nodes");
244 Node *n = useful.at(next);
245 uint max = n->len();
246 for( uint i = 0; i < max; ++i ) {
247 Node *m = n->in(i);
248 if( m == NULL ) continue;
249 useful.push(m);
250 }
251 }
252 }
253
254 // Disconnect all useless nodes by disconnecting those at the boundary.
255 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
256 uint next = 0;
257 while( next < useful.size() ) {
258 Node *n = useful.at(next++);
259 // Use raw traversal of out edges since this code removes out edges
260 int max = n->outcnt();
261 for (int j = 0; j < max; ++j ) {
262 Node* child = n->raw_out(j);
263 if( ! useful.member(child) ) {
264 assert( !child->is_top() || child != top(),
265 "If top is cached in Compile object it is in useful list");
266 // Only need to remove this out-edge to the useless node
267 n->raw_del_out(j);
268 --j;
269 --max;
270 }
271 }
272 if (n->outcnt() == 1 && n->has_special_unique_user()) {
273 record_for_igvn( n->unique_out() );
274 }
275 }
276 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
277 }
278
279 //------------------------------frame_size_in_words-----------------------------
280 // frame_slots in units of words
281 int Compile::frame_size_in_words() const {
282 // shift is 0 in LP32 and 1 in LP64
283 const int shift = (LogBytesPerWord - LogBytesPerInt);
284 int words = _frame_slots >> shift;
285 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
286 return words;
287 }
288
289 // ============================================================================
290 //------------------------------CompileWrapper---------------------------------
291 class CompileWrapper : public StackObj {
292 Compile *const _compile;
293 public:
294 CompileWrapper(Compile* compile);
295
296 ~CompileWrapper();
297 };
298
299 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
300 // the Compile* pointer is stored in the current ciEnv:
301 ciEnv* env = compile->env();
302 assert(env == ciEnv::current(), "must already be a ciEnv active");
303 assert(env->compiler_data() == NULL, "compile already active?");
304 env->set_compiler_data(compile);
305 assert(compile == Compile::current(), "sanity");
306
307 compile->set_type_dict(NULL);
308 compile->set_type_hwm(NULL);
309 compile->set_type_last_size(0);
310 compile->set_last_tf(NULL, NULL);
311 compile->set_indexSet_arena(NULL);
312 compile->set_indexSet_free_block_list(NULL);
313 compile->init_type_arena();
314 Type::Initialize(compile);
315 _compile->set_scratch_buffer_blob(NULL);
316 _compile->begin_method();
317 }
318 CompileWrapper::~CompileWrapper() {
319 _compile->end_method();
320 if (_compile->scratch_buffer_blob() != NULL)
321 BufferBlob::free(_compile->scratch_buffer_blob());
322 _compile->env()->set_compiler_data(NULL);
323 }
324
325
326 //----------------------------print_compile_messages---------------------------
327 void Compile::print_compile_messages() {
328 #ifndef PRODUCT
329 // Check if recompiling
330 if (_subsume_loads == false && PrintOpto) {
331 // Recompiling without allowing machine instructions to subsume loads
332 tty->print_cr("*********************************************************");
333 tty->print_cr("** Bailout: Recompile without subsuming loads **");
334 tty->print_cr("*********************************************************");
335 }
336 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
337 // Recompiling without escape analysis
338 tty->print_cr("*********************************************************");
339 tty->print_cr("** Bailout: Recompile without escape analysis **");
340 tty->print_cr("*********************************************************");
341 }
342 if (env()->break_at_compile()) {
343 // Open the debugger when compiing this method.
344 tty->print("### Breaking when compiling: ");
345 method()->print_short_name();
346 tty->cr();
347 BREAKPOINT;
348 }
349
350 if( PrintOpto ) {
351 if (is_osr_compilation()) {
352 tty->print("[OSR]%3d", _compile_id);
353 } else {
354 tty->print("%3d", _compile_id);
355 }
356 }
357 #endif
358 }
359
360
361 void Compile::init_scratch_buffer_blob() {
362 if( scratch_buffer_blob() != NULL ) return;
363
364 // Construct a temporary CodeBuffer to have it construct a BufferBlob
365 // Cache this BufferBlob for this compile.
366 ResourceMark rm;
367 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size);
368 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size);
369 // Record the buffer blob for next time.
370 set_scratch_buffer_blob(blob);
371 // Have we run out of code space?
372 if (scratch_buffer_blob() == NULL) {
373 // Let CompilerBroker disable further compilations.
374 record_failure("Not enough space for scratch buffer in CodeCache");
375 return;
376 }
377
378 // Initialize the relocation buffers
379 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size;
380 set_scratch_locs_memory(locs_buf);
381 }
382
383
384 //-----------------------scratch_emit_size-------------------------------------
385 // Helper function that computes size by emitting code
386 uint Compile::scratch_emit_size(const Node* n) {
387 // Emit into a trash buffer and count bytes emitted.
388 // This is a pretty expensive way to compute a size,
389 // but it works well enough if seldom used.
390 // All common fixed-size instructions are given a size
391 // method by the AD file.
392 // Note that the scratch buffer blob and locs memory are
393 // allocated at the beginning of the compile task, and
394 // may be shared by several calls to scratch_emit_size.
395 // The allocation of the scratch buffer blob is particularly
396 // expensive, since it has to grab the code cache lock.
397 BufferBlob* blob = this->scratch_buffer_blob();
398 assert(blob != NULL, "Initialize BufferBlob at start");
399 assert(blob->size() > MAX_inst_size, "sanity");
400 relocInfo* locs_buf = scratch_locs_memory();
401 address blob_begin = blob->instructions_begin();
402 address blob_end = (address)locs_buf;
403 assert(blob->instructions_contains(blob_end), "sanity");
404 CodeBuffer buf(blob_begin, blob_end - blob_begin);
405 buf.initialize_consts_size(MAX_const_size);
406 buf.initialize_stubs_size(MAX_stubs_size);
407 assert(locs_buf != NULL, "sanity");
408 int lsize = MAX_locs_size / 2;
409 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize);
410 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize);
411 n->emit(buf, this->regalloc());
412 return buf.code_size();
413 }
414
415
416 // ============================================================================
417 //------------------------------Compile standard-------------------------------
418 debug_only( int Compile::_debug_idx = 100000; )
419
420 // Compile a method. entry_bci is -1 for normal compilations and indicates
421 // the continuation bci for on stack replacement.
422
423
424 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
425 : Phase(Compiler),
426 _env(ci_env),
427 _log(ci_env->log()),
428 _compile_id(ci_env->compile_id()),
429 _save_argument_registers(false),
430 _stub_name(NULL),
431 _stub_function(NULL),
432 _stub_entry_point(NULL),
433 _method(target),
434 _entry_bci(osr_bci),
435 _initial_gvn(NULL),
436 _for_igvn(NULL),
437 _warm_calls(NULL),
438 _subsume_loads(subsume_loads),
439 _do_escape_analysis(do_escape_analysis),
440 _failure_reason(NULL),
441 _code_buffer("Compile::Fill_buffer"),
442 _orig_pc_slot(0),
443 _orig_pc_slot_offset_in_bytes(0),
444 _node_bundling_limit(0),
445 _node_bundling_base(NULL),
446 #ifndef PRODUCT
447 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
448 _printer(IdealGraphPrinter::printer()),
449 #endif
450 _congraph(NULL) {
451 C = this;
452
453 CompileWrapper cw(this);
454 #ifndef PRODUCT
455 if (TimeCompiler2) {
456 tty->print(" ");
457 target->holder()->name()->print();
458 tty->print(".");
459 target->print_short_name();
460 tty->print(" ");
461 }
462 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
463 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
464 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
465 if (!print_opto_assembly) {
466 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
467 if (print_assembly && !Disassembler::can_decode()) {
468 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
469 print_opto_assembly = true;
470 }
471 }
472 set_print_assembly(print_opto_assembly);
473 set_parsed_irreducible_loop(false);
474 #endif
475
476 if (ProfileTraps) {
477 // Make sure the method being compiled gets its own MDO,
478 // so we can at least track the decompile_count().
479 method()->build_method_data();
480 }
481
482 Init(::AliasLevel);
483
484
485 print_compile_messages();
486
487 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
488 _ilt = InlineTree::build_inline_tree_root();
489 else
490 _ilt = NULL;
491
492 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
493 assert(num_alias_types() >= AliasIdxRaw, "");
494
495 #define MINIMUM_NODE_HASH 1023
496 // Node list that Iterative GVN will start with
497 Unique_Node_List for_igvn(comp_arena());
498 set_for_igvn(&for_igvn);
499
500 // GVN that will be run immediately on new nodes
501 uint estimated_size = method()->code_size()*4+64;
502 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
503 PhaseGVN gvn(node_arena(), estimated_size);
504 set_initial_gvn(&gvn);
505
506 { // Scope for timing the parser
507 TracePhase t3("parse", &_t_parser, true);
508
509 // Put top into the hash table ASAP.
510 initial_gvn()->transform_no_reclaim(top());
511
512 // Set up tf(), start(), and find a CallGenerator.
513 CallGenerator* cg;
514 if (is_osr_compilation()) {
515 const TypeTuple *domain = StartOSRNode::osr_domain();
516 const TypeTuple *range = TypeTuple::make_range(method()->signature());
517 init_tf(TypeFunc::make(domain, range));
518 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
519 initial_gvn()->set_type_bottom(s);
520 init_start(s);
521 cg = CallGenerator::for_osr(method(), entry_bci());
522 } else {
523 // Normal case.
524 init_tf(TypeFunc::make(method()));
525 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
526 initial_gvn()->set_type_bottom(s);
527 init_start(s);
528 float past_uses = method()->interpreter_invocation_count();
529 float expected_uses = past_uses;
530 cg = CallGenerator::for_inline(method(), expected_uses);
531 }
532 if (failing()) return;
533 if (cg == NULL) {
534 record_method_not_compilable_all_tiers("cannot parse method");
535 return;
536 }
537 JVMState* jvms = build_start_state(start(), tf());
538 if ((jvms = cg->generate(jvms)) == NULL) {
539 record_method_not_compilable("method parse failed");
540 return;
541 }
542 GraphKit kit(jvms);
543
544 if (!kit.stopped()) {
545 // Accept return values, and transfer control we know not where.
546 // This is done by a special, unique ReturnNode bound to root.
547 return_values(kit.jvms());
548 }
549
550 if (kit.has_exceptions()) {
551 // Any exceptions that escape from this call must be rethrown
552 // to whatever caller is dynamically above us on the stack.
553 // This is done by a special, unique RethrowNode bound to root.
554 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
555 }
556
557 print_method("Before RemoveUseless", 3);
558
559 // Remove clutter produced by parsing.
560 if (!failing()) {
561 ResourceMark rm;
562 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
563 }
564 }
565
566 // Note: Large methods are capped off in do_one_bytecode().
567 if (failing()) return;
568
569 // After parsing, node notes are no longer automagic.
570 // They must be propagated by register_new_node_with_optimizer(),
571 // clone(), or the like.
572 set_default_node_notes(NULL);
573
574 for (;;) {
575 int successes = Inline_Warm();
576 if (failing()) return;
577 if (successes == 0) break;
578 }
579
580 // Drain the list.
581 Finish_Warm();
582 #ifndef PRODUCT
583 if (_printer) {
584 _printer->print_inlining(this);
585 }
586 #endif
587
588 if (failing()) return;
589 NOT_PRODUCT( verify_graph_edges(); )
590
591 // Perform escape analysis
592 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
593 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
594 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction.
595 PhaseGVN* igvn = initial_gvn();
596 Node* oop_null = igvn->zerocon(T_OBJECT);
597 Node* noop_null = igvn->zerocon(T_NARROWOOP);
598
599 _congraph = new(comp_arena()) ConnectionGraph(this);
600 bool has_non_escaping_obj = _congraph->compute_escape();
601
602 #ifndef PRODUCT
603 if (PrintEscapeAnalysis) {
604 _congraph->dump();
605 }
606 #endif
607 // Cleanup.
608 if (oop_null->outcnt() == 0)
609 igvn->hash_delete(oop_null);
610 if (noop_null->outcnt() == 0)
611 igvn->hash_delete(noop_null);
612
613 if (!has_non_escaping_obj) {
614 _congraph = NULL;
615 }
616
617 if (failing()) return;
618 }
619 // Now optimize
620 Optimize();
621 if (failing()) return;
622 NOT_PRODUCT( verify_graph_edges(); )
623
624 #ifndef PRODUCT
625 if (PrintIdeal) {
626 ttyLocker ttyl; // keep the following output all in one block
627 // This output goes directly to the tty, not the compiler log.
628 // To enable tools to match it up with the compilation activity,
629 // be sure to tag this tty output with the compile ID.
630 if (xtty != NULL) {
631 xtty->head("ideal compile_id='%d'%s", compile_id(),
632 is_osr_compilation() ? " compile_kind='osr'" :
633 "");
634 }
635 root()->dump(9999);
636 if (xtty != NULL) {
637 xtty->tail("ideal");
638 }
639 }
640 #endif
641
642 // Now that we know the size of all the monitors we can add a fixed slot
643 // for the original deopt pc.
644
645 _orig_pc_slot = fixed_slots();
646 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
647 set_fixed_slots(next_slot);
648
649 // Now generate code
650 Code_Gen();
651 if (failing()) return;
652
653 // Check if we want to skip execution of all compiled code.
654 {
655 #ifndef PRODUCT
656 if (OptoNoExecute) {
657 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
658 return;
659 }
660 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
661 #endif
662
663 if (is_osr_compilation()) {
664 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
665 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
666 } else {
667 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
668 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
669 }
670
671 env()->register_method(_method, _entry_bci,
672 &_code_offsets,
673 _orig_pc_slot_offset_in_bytes,
674 code_buffer(),
675 frame_size_in_words(), _oop_map_set,
676 &_handler_table, &_inc_table,
677 compiler,
678 env()->comp_level(),
679 true, /*has_debug_info*/
680 has_unsafe_access()
681 );
682 }
683 }
684
685 //------------------------------Compile----------------------------------------
686 // Compile a runtime stub
687 Compile::Compile( ciEnv* ci_env,
688 TypeFunc_generator generator,
689 address stub_function,
690 const char *stub_name,
691 int is_fancy_jump,
692 bool pass_tls,
693 bool save_arg_registers,
694 bool return_pc )
695 : Phase(Compiler),
696 _env(ci_env),
697 _log(ci_env->log()),
698 _compile_id(-1),
699 _save_argument_registers(save_arg_registers),
700 _method(NULL),
701 _stub_name(stub_name),
702 _stub_function(stub_function),
703 _stub_entry_point(NULL),
704 _entry_bci(InvocationEntryBci),
705 _initial_gvn(NULL),
706 _for_igvn(NULL),
707 _warm_calls(NULL),
708 _orig_pc_slot(0),
709 _orig_pc_slot_offset_in_bytes(0),
710 _subsume_loads(true),
711 _do_escape_analysis(false),
712 _failure_reason(NULL),
713 _code_buffer("Compile::Fill_buffer"),
714 _node_bundling_limit(0),
715 _node_bundling_base(NULL),
716 #ifndef PRODUCT
717 _trace_opto_output(TraceOptoOutput),
718 _printer(NULL),
719 #endif
720 _congraph(NULL) {
721 C = this;
722
723 #ifndef PRODUCT
724 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
725 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
726 set_print_assembly(PrintFrameConverterAssembly);
727 set_parsed_irreducible_loop(false);
728 #endif
729 CompileWrapper cw(this);
730 Init(/*AliasLevel=*/ 0);
731 init_tf((*generator)());
732
733 {
734 // The following is a dummy for the sake of GraphKit::gen_stub
735 Unique_Node_List for_igvn(comp_arena());
736 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
737 PhaseGVN gvn(Thread::current()->resource_area(),255);
738 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
739 gvn.transform_no_reclaim(top());
740
741 GraphKit kit;
742 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
743 }
744
745 NOT_PRODUCT( verify_graph_edges(); )
746 Code_Gen();
747 if (failing()) return;
748
749
750 // Entry point will be accessed using compile->stub_entry_point();
751 if (code_buffer() == NULL) {
752 Matcher::soft_match_failure();
753 } else {
754 if (PrintAssembly && (WizardMode || Verbose))
755 tty->print_cr("### Stub::%s", stub_name);
756
757 if (!failing()) {
758 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
759
760 // Make the NMethod
761 // For now we mark the frame as never safe for profile stackwalking
762 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
763 code_buffer(),
764 CodeOffsets::frame_never_safe,
765 // _code_offsets.value(CodeOffsets::Frame_Complete),
766 frame_size_in_words(),
767 _oop_map_set,
768 save_arg_registers);
769 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
770
771 _stub_entry_point = rs->entry_point();
772 }
773 }
774 }
775
776 #ifndef PRODUCT
777 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
778 if(PrintOpto && Verbose) {
779 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
780 }
781 }
782 #endif
783
784 void Compile::print_codes() {
785 }
786
787 //------------------------------Init-------------------------------------------
788 // Prepare for a single compilation
789 void Compile::Init(int aliaslevel) {
790 _unique = 0;
791 _regalloc = NULL;
792
793 _tf = NULL; // filled in later
794 _top = NULL; // cached later
795 _matcher = NULL; // filled in later
796 _cfg = NULL; // filled in later
797
798 set_24_bit_selection_and_mode(Use24BitFP, false);
799
800 _node_note_array = NULL;
801 _default_node_notes = NULL;
802
803 _immutable_memory = NULL; // filled in at first inquiry
804
805 // Globally visible Nodes
806 // First set TOP to NULL to give safe behavior during creation of RootNode
807 set_cached_top_node(NULL);
808 set_root(new (this, 3) RootNode());
809 // Now that you have a Root to point to, create the real TOP
810 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
811 set_recent_alloc(NULL, NULL);
812
813 // Create Debug Information Recorder to record scopes, oopmaps, etc.
814 env()->set_oop_recorder(new OopRecorder(comp_arena()));
815 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
816 env()->set_dependencies(new Dependencies(env()));
817
818 _fixed_slots = 0;
819 set_has_split_ifs(false);
820 set_has_loops(has_method() && method()->has_loops()); // first approximation
821 _deopt_happens = true; // start out assuming the worst
822 _trap_can_recompile = false; // no traps emitted yet
823 _major_progress = true; // start out assuming good things will happen
824 set_has_unsafe_access(false);
825 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
826 set_decompile_count(0);
827
828 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
829 // Compilation level related initialization
830 if (env()->comp_level() == CompLevel_fast_compile) {
831 set_num_loop_opts(Tier1LoopOptsCount);
832 set_do_inlining(Tier1Inline != 0);
833 set_max_inline_size(Tier1MaxInlineSize);
834 set_freq_inline_size(Tier1FreqInlineSize);
835 set_do_scheduling(false);
836 set_do_count_invocations(Tier1CountInvocations);
837 set_do_method_data_update(Tier1UpdateMethodData);
838 } else {
839 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
840 set_num_loop_opts(LoopOptsCount);
841 set_do_inlining(Inline);
842 set_max_inline_size(MaxInlineSize);
843 set_freq_inline_size(FreqInlineSize);
844 set_do_scheduling(OptoScheduling);
845 set_do_count_invocations(false);
846 set_do_method_data_update(false);
847 }
848
849 if (debug_info()->recording_non_safepoints()) {
850 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
851 (comp_arena(), 8, 0, NULL));
852 set_default_node_notes(Node_Notes::make(this));
853 }
854
855 // // -- Initialize types before each compile --
856 // // Update cached type information
857 // if( _method && _method->constants() )
858 // Type::update_loaded_types(_method, _method->constants());
859
860 // Init alias_type map.
861 if (!_do_escape_analysis && aliaslevel == 3)
862 aliaslevel = 2; // No unique types without escape analysis
863 _AliasLevel = aliaslevel;
864 const int grow_ats = 16;
865 _max_alias_types = grow_ats;
866 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
867 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
868 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
869 {
870 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
871 }
872 // Initialize the first few types.
873 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
874 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
875 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
876 _num_alias_types = AliasIdxRaw+1;
877 // Zero out the alias type cache.
878 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
879 // A NULL adr_type hits in the cache right away. Preload the right answer.
880 probe_alias_cache(NULL)->_index = AliasIdxTop;
881
882 _intrinsics = NULL;
883 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
884 register_library_intrinsics();
885 }
886
887 //---------------------------init_start----------------------------------------
888 // Install the StartNode on this compile object.
889 void Compile::init_start(StartNode* s) {
890 if (failing())
891 return; // already failing
892 assert(s == start(), "");
893 }
894
895 StartNode* Compile::start() const {
896 assert(!failing(), "");
897 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
898 Node* start = root()->fast_out(i);
899 if( start->is_Start() )
900 return start->as_Start();
901 }
902 ShouldNotReachHere();
903 return NULL;
904 }
905
906 //-------------------------------immutable_memory-------------------------------------
907 // Access immutable memory
908 Node* Compile::immutable_memory() {
909 if (_immutable_memory != NULL) {
910 return _immutable_memory;
911 }
912 StartNode* s = start();
913 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
914 Node *p = s->fast_out(i);
915 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
916 _immutable_memory = p;
917 return _immutable_memory;
918 }
919 }
920 ShouldNotReachHere();
921 return NULL;
922 }
923
924 //----------------------set_cached_top_node------------------------------------
925 // Install the cached top node, and make sure Node::is_top works correctly.
926 void Compile::set_cached_top_node(Node* tn) {
927 if (tn != NULL) verify_top(tn);
928 Node* old_top = _top;
929 _top = tn;
930 // Calling Node::setup_is_top allows the nodes the chance to adjust
931 // their _out arrays.
932 if (_top != NULL) _top->setup_is_top();
933 if (old_top != NULL) old_top->setup_is_top();
934 assert(_top == NULL || top()->is_top(), "");
935 }
936
937 #ifndef PRODUCT
938 void Compile::verify_top(Node* tn) const {
939 if (tn != NULL) {
940 assert(tn->is_Con(), "top node must be a constant");
941 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
942 assert(tn->in(0) != NULL, "must have live top node");
943 }
944 }
945 #endif
946
947
948 ///-------------------Managing Per-Node Debug & Profile Info-------------------
949
950 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
951 guarantee(arr != NULL, "");
952 int num_blocks = arr->length();
953 if (grow_by < num_blocks) grow_by = num_blocks;
954 int num_notes = grow_by * _node_notes_block_size;
955 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
956 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
957 while (num_notes > 0) {
958 arr->append(notes);
959 notes += _node_notes_block_size;
960 num_notes -= _node_notes_block_size;
961 }
962 assert(num_notes == 0, "exact multiple, please");
963 }
964
965 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
966 if (source == NULL || dest == NULL) return false;
967
968 if (dest->is_Con())
969 return false; // Do not push debug info onto constants.
970
971 #ifdef ASSERT
972 // Leave a bread crumb trail pointing to the original node:
973 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
974 dest->set_debug_orig(source);
975 }
976 #endif
977
978 if (node_note_array() == NULL)
979 return false; // Not collecting any notes now.
980
981 // This is a copy onto a pre-existing node, which may already have notes.
982 // If both nodes have notes, do not overwrite any pre-existing notes.
983 Node_Notes* source_notes = node_notes_at(source->_idx);
984 if (source_notes == NULL || source_notes->is_clear()) return false;
985 Node_Notes* dest_notes = node_notes_at(dest->_idx);
986 if (dest_notes == NULL || dest_notes->is_clear()) {
987 return set_node_notes_at(dest->_idx, source_notes);
988 }
989
990 Node_Notes merged_notes = (*source_notes);
991 // The order of operations here ensures that dest notes will win...
992 merged_notes.update_from(dest_notes);
993 return set_node_notes_at(dest->_idx, &merged_notes);
994 }
995
996
997 //--------------------------allow_range_check_smearing-------------------------
998 // Gating condition for coalescing similar range checks.
999 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1000 // single covering check that is at least as strong as any of them.
1001 // If the optimization succeeds, the simplified (strengthened) range check
1002 // will always succeed. If it fails, we will deopt, and then give up
1003 // on the optimization.
1004 bool Compile::allow_range_check_smearing() const {
1005 // If this method has already thrown a range-check,
1006 // assume it was because we already tried range smearing
1007 // and it failed.
1008 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1009 return !already_trapped;
1010 }
1011
1012
1013 //------------------------------flatten_alias_type-----------------------------
1014 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1015 int offset = tj->offset();
1016 TypePtr::PTR ptr = tj->ptr();
1017
1018 // Known instance (scalarizable allocation) alias only with itself.
1019 bool is_known_inst = tj->isa_oopptr() != NULL &&
1020 tj->is_oopptr()->is_known_instance();
1021
1022 // Process weird unsafe references.
1023 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1024 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1025 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1026 tj = TypeOopPtr::BOTTOM;
1027 ptr = tj->ptr();
1028 offset = tj->offset();
1029 }
1030
1031 // Array pointers need some flattening
1032 const TypeAryPtr *ta = tj->isa_aryptr();
1033 if( ta && is_known_inst ) {
1034 if ( offset != Type::OffsetBot &&
1035 offset > arrayOopDesc::length_offset_in_bytes() ) {
1036 offset = Type::OffsetBot; // Flatten constant access into array body only
1037 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1038 }
1039 } else if( ta && _AliasLevel >= 2 ) {
1040 // For arrays indexed by constant indices, we flatten the alias
1041 // space to include all of the array body. Only the header, klass
1042 // and array length can be accessed un-aliased.
1043 if( offset != Type::OffsetBot ) {
1044 if( ta->const_oop() ) { // methodDataOop or methodOop
1045 offset = Type::OffsetBot; // Flatten constant access into array body
1046 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1047 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1048 // range is OK as-is.
1049 tj = ta = TypeAryPtr::RANGE;
1050 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1051 tj = TypeInstPtr::KLASS; // all klass loads look alike
1052 ta = TypeAryPtr::RANGE; // generic ignored junk
1053 ptr = TypePtr::BotPTR;
1054 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1055 tj = TypeInstPtr::MARK;
1056 ta = TypeAryPtr::RANGE; // generic ignored junk
1057 ptr = TypePtr::BotPTR;
1058 } else { // Random constant offset into array body
1059 offset = Type::OffsetBot; // Flatten constant access into array body
1060 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1061 }
1062 }
1063 // Arrays of fixed size alias with arrays of unknown size.
1064 if (ta->size() != TypeInt::POS) {
1065 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1066 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1067 }
1068 // Arrays of known objects become arrays of unknown objects.
1069 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1070 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1071 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1072 }
1073 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1074 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1075 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1076 }
1077 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1078 // cannot be distinguished by bytecode alone.
1079 if (ta->elem() == TypeInt::BOOL) {
1080 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1081 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1082 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1083 }
1084 // During the 2nd round of IterGVN, NotNull castings are removed.
1085 // Make sure the Bottom and NotNull variants alias the same.
1086 // Also, make sure exact and non-exact variants alias the same.
1087 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1088 if (ta->const_oop()) {
1089 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1090 } else {
1091 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1092 }
1093 }
1094 }
1095
1096 // Oop pointers need some flattening
1097 const TypeInstPtr *to = tj->isa_instptr();
1098 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1099 if( ptr == TypePtr::Constant ) {
1100 // No constant oop pointers (such as Strings); they alias with
1101 // unknown strings.
1102 assert(!is_known_inst, "not scalarizable allocation");
1103 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1104 } else if( is_known_inst ) {
1105 tj = to; // Keep NotNull and klass_is_exact for instance type
1106 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1107 // During the 2nd round of IterGVN, NotNull castings are removed.
1108 // Make sure the Bottom and NotNull variants alias the same.
1109 // Also, make sure exact and non-exact variants alias the same.
1110 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1111 }
1112 // Canonicalize the holder of this field
1113 ciInstanceKlass *k = to->klass()->as_instance_klass();
1114 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1115 // First handle header references such as a LoadKlassNode, even if the
1116 // object's klass is unloaded at compile time (4965979).
1117 if (!is_known_inst) { // Do it only for non-instance types
1118 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1119 }
1120 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1121 to = NULL;
1122 tj = TypeOopPtr::BOTTOM;
1123 offset = tj->offset();
1124 } else {
1125 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1126 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1127 if( is_known_inst ) {
1128 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1129 } else {
1130 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1131 }
1132 }
1133 }
1134 }
1135
1136 // Klass pointers to object array klasses need some flattening
1137 const TypeKlassPtr *tk = tj->isa_klassptr();
1138 if( tk ) {
1139 // If we are referencing a field within a Klass, we need
1140 // to assume the worst case of an Object. Both exact and
1141 // inexact types must flatten to the same alias class.
1142 // Since the flattened result for a klass is defined to be
1143 // precisely java.lang.Object, use a constant ptr.
1144 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1145
1146 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1147 TypeKlassPtr::OBJECT->klass(),
1148 offset);
1149 }
1150
1151 ciKlass* klass = tk->klass();
1152 if( klass->is_obj_array_klass() ) {
1153 ciKlass* k = TypeAryPtr::OOPS->klass();
1154 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1155 k = TypeInstPtr::BOTTOM->klass();
1156 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1157 }
1158
1159 // Check for precise loads from the primary supertype array and force them
1160 // to the supertype cache alias index. Check for generic array loads from
1161 // the primary supertype array and also force them to the supertype cache
1162 // alias index. Since the same load can reach both, we need to merge
1163 // these 2 disparate memories into the same alias class. Since the
1164 // primary supertype array is read-only, there's no chance of confusion
1165 // where we bypass an array load and an array store.
1166 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1167 if( offset == Type::OffsetBot ||
1168 off2 < Klass::primary_super_limit()*wordSize ) {
1169 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1170 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1171 }
1172 }
1173
1174 // Flatten all Raw pointers together.
1175 if (tj->base() == Type::RawPtr)
1176 tj = TypeRawPtr::BOTTOM;
1177
1178 if (tj->base() == Type::AnyPtr)
1179 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1180
1181 // Flatten all to bottom for now
1182 switch( _AliasLevel ) {
1183 case 0:
1184 tj = TypePtr::BOTTOM;
1185 break;
1186 case 1: // Flatten to: oop, static, field or array
1187 switch (tj->base()) {
1188 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1189 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1190 case Type::AryPtr: // do not distinguish arrays at all
1191 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1192 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1193 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1194 default: ShouldNotReachHere();
1195 }
1196 break;
1197 case 2: // No collasping at level 2; keep all splits
1198 case 3: // No collasping at level 3; keep all splits
1199 break;
1200 default:
1201 Unimplemented();
1202 }
1203
1204 offset = tj->offset();
1205 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1206
1207 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1208 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1209 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1210 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1211 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1212 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1213 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1214 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1215 assert( tj->ptr() != TypePtr::TopPTR &&
1216 tj->ptr() != TypePtr::AnyNull &&
1217 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1218 // assert( tj->ptr() != TypePtr::Constant ||
1219 // tj->base() == Type::RawPtr ||
1220 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1221
1222 return tj;
1223 }
1224
1225 void Compile::AliasType::Init(int i, const TypePtr* at) {
1226 _index = i;
1227 _adr_type = at;
1228 _field = NULL;
1229 _is_rewritable = true; // default
1230 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1231 if (atoop != NULL && atoop->is_known_instance()) {
1232 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1233 _general_index = Compile::current()->get_alias_index(gt);
1234 } else {
1235 _general_index = 0;
1236 }
1237 }
1238
1239 //---------------------------------print_on------------------------------------
1240 #ifndef PRODUCT
1241 void Compile::AliasType::print_on(outputStream* st) {
1242 if (index() < 10)
1243 st->print("@ <%d> ", index());
1244 else st->print("@ <%d>", index());
1245 st->print(is_rewritable() ? " " : " RO");
1246 int offset = adr_type()->offset();
1247 if (offset == Type::OffsetBot)
1248 st->print(" +any");
1249 else st->print(" +%-3d", offset);
1250 st->print(" in ");
1251 adr_type()->dump_on(st);
1252 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1253 if (field() != NULL && tjp) {
1254 if (tjp->klass() != field()->holder() ||
1255 tjp->offset() != field()->offset_in_bytes()) {
1256 st->print(" != ");
1257 field()->print();
1258 st->print(" ***");
1259 }
1260 }
1261 }
1262
1263 void print_alias_types() {
1264 Compile* C = Compile::current();
1265 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1266 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1267 C->alias_type(idx)->print_on(tty);
1268 tty->cr();
1269 }
1270 }
1271 #endif
1272
1273
1274 //----------------------------probe_alias_cache--------------------------------
1275 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1276 intptr_t key = (intptr_t) adr_type;
1277 key ^= key >> logAliasCacheSize;
1278 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1279 }
1280
1281
1282 //-----------------------------grow_alias_types--------------------------------
1283 void Compile::grow_alias_types() {
1284 const int old_ats = _max_alias_types; // how many before?
1285 const int new_ats = old_ats; // how many more?
1286 const int grow_ats = old_ats+new_ats; // how many now?
1287 _max_alias_types = grow_ats;
1288 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1289 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1290 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1291 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1292 }
1293
1294
1295 //--------------------------------find_alias_type------------------------------
1296 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1297 if (_AliasLevel == 0)
1298 return alias_type(AliasIdxBot);
1299
1300 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1301 if (ace->_adr_type == adr_type) {
1302 return alias_type(ace->_index);
1303 }
1304
1305 // Handle special cases.
1306 if (adr_type == NULL) return alias_type(AliasIdxTop);
1307 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1308
1309 // Do it the slow way.
1310 const TypePtr* flat = flatten_alias_type(adr_type);
1311
1312 #ifdef ASSERT
1313 assert(flat == flatten_alias_type(flat), "idempotent");
1314 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1315 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1316 const TypeOopPtr* foop = flat->is_oopptr();
1317 // Scalarizable allocations have exact klass always.
1318 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1319 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1320 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1321 }
1322 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1323 #endif
1324
1325 int idx = AliasIdxTop;
1326 for (int i = 0; i < num_alias_types(); i++) {
1327 if (alias_type(i)->adr_type() == flat) {
1328 idx = i;
1329 break;
1330 }
1331 }
1332
1333 if (idx == AliasIdxTop) {
1334 if (no_create) return NULL;
1335 // Grow the array if necessary.
1336 if (_num_alias_types == _max_alias_types) grow_alias_types();
1337 // Add a new alias type.
1338 idx = _num_alias_types++;
1339 _alias_types[idx]->Init(idx, flat);
1340 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1341 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1342 if (flat->isa_instptr()) {
1343 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1344 && flat->is_instptr()->klass() == env()->Class_klass())
1345 alias_type(idx)->set_rewritable(false);
1346 }
1347 if (flat->isa_klassptr()) {
1348 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1349 alias_type(idx)->set_rewritable(false);
1350 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1351 alias_type(idx)->set_rewritable(false);
1352 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1353 alias_type(idx)->set_rewritable(false);
1354 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1355 alias_type(idx)->set_rewritable(false);
1356 }
1357 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1358 // but the base pointer type is not distinctive enough to identify
1359 // references into JavaThread.)
1360
1361 // Check for final instance fields.
1362 const TypeInstPtr* tinst = flat->isa_instptr();
1363 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1364 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1365 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1366 // Set field() and is_rewritable() attributes.
1367 if (field != NULL) alias_type(idx)->set_field(field);
1368 }
1369 const TypeKlassPtr* tklass = flat->isa_klassptr();
1370 // Check for final static fields.
1371 if (tklass && tklass->klass()->is_instance_klass()) {
1372 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1373 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1374 // Set field() and is_rewritable() attributes.
1375 if (field != NULL) alias_type(idx)->set_field(field);
1376 }
1377 }
1378
1379 // Fill the cache for next time.
1380 ace->_adr_type = adr_type;
1381 ace->_index = idx;
1382 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1383
1384 // Might as well try to fill the cache for the flattened version, too.
1385 AliasCacheEntry* face = probe_alias_cache(flat);
1386 if (face->_adr_type == NULL) {
1387 face->_adr_type = flat;
1388 face->_index = idx;
1389 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1390 }
1391
1392 return alias_type(idx);
1393 }
1394
1395
1396 Compile::AliasType* Compile::alias_type(ciField* field) {
1397 const TypeOopPtr* t;
1398 if (field->is_static())
1399 t = TypeKlassPtr::make(field->holder());
1400 else
1401 t = TypeOopPtr::make_from_klass_raw(field->holder());
1402 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1403 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1404 return atp;
1405 }
1406
1407
1408 //------------------------------have_alias_type--------------------------------
1409 bool Compile::have_alias_type(const TypePtr* adr_type) {
1410 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1411 if (ace->_adr_type == adr_type) {
1412 return true;
1413 }
1414
1415 // Handle special cases.
1416 if (adr_type == NULL) return true;
1417 if (adr_type == TypePtr::BOTTOM) return true;
1418
1419 return find_alias_type(adr_type, true) != NULL;
1420 }
1421
1422 //-----------------------------must_alias--------------------------------------
1423 // True if all values of the given address type are in the given alias category.
1424 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1425 if (alias_idx == AliasIdxBot) return true; // the universal category
1426 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1427 if (alias_idx == AliasIdxTop) return false; // the empty category
1428 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1429
1430 // the only remaining possible overlap is identity
1431 int adr_idx = get_alias_index(adr_type);
1432 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1433 assert(adr_idx == alias_idx ||
1434 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1435 && adr_type != TypeOopPtr::BOTTOM),
1436 "should not be testing for overlap with an unsafe pointer");
1437 return adr_idx == alias_idx;
1438 }
1439
1440 //------------------------------can_alias--------------------------------------
1441 // True if any values of the given address type are in the given alias category.
1442 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1443 if (alias_idx == AliasIdxTop) return false; // the empty category
1444 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1445 if (alias_idx == AliasIdxBot) return true; // the universal category
1446 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1447
1448 // the only remaining possible overlap is identity
1449 int adr_idx = get_alias_index(adr_type);
1450 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1451 return adr_idx == alias_idx;
1452 }
1453
1454
1455
1456 //---------------------------pop_warm_call-------------------------------------
1457 WarmCallInfo* Compile::pop_warm_call() {
1458 WarmCallInfo* wci = _warm_calls;
1459 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1460 return wci;
1461 }
1462
1463 //----------------------------Inline_Warm--------------------------------------
1464 int Compile::Inline_Warm() {
1465 // If there is room, try to inline some more warm call sites.
1466 // %%% Do a graph index compaction pass when we think we're out of space?
1467 if (!InlineWarmCalls) return 0;
1468
1469 int calls_made_hot = 0;
1470 int room_to_grow = NodeCountInliningCutoff - unique();
1471 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1472 int amount_grown = 0;
1473 WarmCallInfo* call;
1474 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1475 int est_size = (int)call->size();
1476 if (est_size > (room_to_grow - amount_grown)) {
1477 // This one won't fit anyway. Get rid of it.
1478 call->make_cold();
1479 continue;
1480 }
1481 call->make_hot();
1482 calls_made_hot++;
1483 amount_grown += est_size;
1484 amount_to_grow -= est_size;
1485 }
1486
1487 if (calls_made_hot > 0) set_major_progress();
1488 return calls_made_hot;
1489 }
1490
1491
1492 //----------------------------Finish_Warm--------------------------------------
1493 void Compile::Finish_Warm() {
1494 if (!InlineWarmCalls) return;
1495 if (failing()) return;
1496 if (warm_calls() == NULL) return;
1497
1498 // Clean up loose ends, if we are out of space for inlining.
1499 WarmCallInfo* call;
1500 while ((call = pop_warm_call()) != NULL) {
1501 call->make_cold();
1502 }
1503 }
1504
1505
1506 //------------------------------Optimize---------------------------------------
1507 // Given a graph, optimize it.
1508 void Compile::Optimize() {
1509 TracePhase t1("optimizer", &_t_optimizer, true);
1510
1511 #ifndef PRODUCT
1512 if (env()->break_at_compile()) {
1513 BREAKPOINT;
1514 }
1515
1516 #endif
1517
1518 ResourceMark rm;
1519 int loop_opts_cnt;
1520
1521 NOT_PRODUCT( verify_graph_edges(); )
1522
1523 print_method("After Parsing");
1524
1525 {
1526 // Iterative Global Value Numbering, including ideal transforms
1527 // Initialize IterGVN with types and values from parse-time GVN
1528 PhaseIterGVN igvn(initial_gvn());
1529 {
1530 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1531 igvn.optimize();
1532 }
1533
1534 print_method("Iter GVN 1", 2);
1535
1536 if (failing()) return;
1537
1538 // Loop transforms on the ideal graph. Range Check Elimination,
1539 // peeling, unrolling, etc.
1540
1541 // Set loop opts counter
1542 loop_opts_cnt = num_loop_opts();
1543 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1544 {
1545 TracePhase t2("idealLoop", &_t_idealLoop, true);
1546 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1547 loop_opts_cnt--;
1548 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1549 if (failing()) return;
1550 }
1551 // Loop opts pass if partial peeling occurred in previous pass
1552 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1553 TracePhase t3("idealLoop", &_t_idealLoop, true);
1554 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1555 loop_opts_cnt--;
1556 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1557 if (failing()) return;
1558 }
1559 // Loop opts pass for loop-unrolling before CCP
1560 if(major_progress() && (loop_opts_cnt > 0)) {
1561 TracePhase t4("idealLoop", &_t_idealLoop, true);
1562 PhaseIdealLoop ideal_loop( igvn, NULL, false );
1563 loop_opts_cnt--;
1564 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1565 }
1566 }
1567 if (failing()) return;
1568
1569 // Conditional Constant Propagation;
1570 PhaseCCP ccp( &igvn );
1571 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1572 {
1573 TracePhase t2("ccp", &_t_ccp, true);
1574 ccp.do_transform();
1575 }
1576 print_method("PhaseCPP 1", 2);
1577
1578 assert( true, "Break here to ccp.dump_old2new_map()");
1579
1580 // Iterative Global Value Numbering, including ideal transforms
1581 {
1582 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1583 igvn = ccp;
1584 igvn.optimize();
1585 }
1586
1587 print_method("Iter GVN 2", 2);
1588
1589 if (failing()) return;
1590
1591 // Loop transforms on the ideal graph. Range Check Elimination,
1592 // peeling, unrolling, etc.
1593 if(loop_opts_cnt > 0) {
1594 debug_only( int cnt = 0; );
1595 while(major_progress() && (loop_opts_cnt > 0)) {
1596 TracePhase t2("idealLoop", &_t_idealLoop, true);
1597 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1598 PhaseIdealLoop ideal_loop( igvn, NULL, true );
1599 loop_opts_cnt--;
1600 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1601 if (failing()) return;
1602 }
1603 }
1604 {
1605 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1606 PhaseMacroExpand mex(igvn);
1607 if (mex.expand_macro_nodes()) {
1608 assert(failing(), "must bail out w/ explicit message");
1609 return;
1610 }
1611 }
1612
1613 } // (End scope of igvn; run destructor if necessary for asserts.)
1614
1615 // A method with only infinite loops has no edges entering loops from root
1616 {
1617 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1618 if (final_graph_reshaping()) {
1619 assert(failing(), "must bail out w/ explicit message");
1620 return;
1621 }
1622 }
1623
1624 print_method("Optimize finished", 2);
1625 }
1626
1627
1628 //------------------------------Code_Gen---------------------------------------
1629 // Given a graph, generate code for it
1630 void Compile::Code_Gen() {
1631 if (failing()) return;
1632
1633 // Perform instruction selection. You might think we could reclaim Matcher
1634 // memory PDQ, but actually the Matcher is used in generating spill code.
1635 // Internals of the Matcher (including some VectorSets) must remain live
1636 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1637 // set a bit in reclaimed memory.
1638
1639 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1640 // nodes. Mapping is only valid at the root of each matched subtree.
1641 NOT_PRODUCT( verify_graph_edges(); )
1642
1643 Node_List proj_list;
1644 Matcher m(proj_list);
1645 _matcher = &m;
1646 {
1647 TracePhase t2("matcher", &_t_matcher, true);
1648 m.match();
1649 }
1650 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1651 // nodes. Mapping is only valid at the root of each matched subtree.
1652 NOT_PRODUCT( verify_graph_edges(); )
1653
1654 // If you have too many nodes, or if matching has failed, bail out
1655 check_node_count(0, "out of nodes matching instructions");
1656 if (failing()) return;
1657
1658 // Build a proper-looking CFG
1659 PhaseCFG cfg(node_arena(), root(), m);
1660 _cfg = &cfg;
1661 {
1662 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1663 cfg.Dominators();
1664 if (failing()) return;
1665
1666 NOT_PRODUCT( verify_graph_edges(); )
1667
1668 cfg.Estimate_Block_Frequency();
1669 cfg.GlobalCodeMotion(m,unique(),proj_list);
1670
1671 print_method("Global code motion", 2);
1672
1673 if (failing()) return;
1674 NOT_PRODUCT( verify_graph_edges(); )
1675
1676 debug_only( cfg.verify(); )
1677 }
1678 NOT_PRODUCT( verify_graph_edges(); )
1679
1680 PhaseChaitin regalloc(unique(),cfg,m);
1681 _regalloc = ®alloc;
1682 {
1683 TracePhase t2("regalloc", &_t_registerAllocation, true);
1684 // Perform any platform dependent preallocation actions. This is used,
1685 // for example, to avoid taking an implicit null pointer exception
1686 // using the frame pointer on win95.
1687 _regalloc->pd_preallocate_hook();
1688
1689 // Perform register allocation. After Chaitin, use-def chains are
1690 // no longer accurate (at spill code) and so must be ignored.
1691 // Node->LRG->reg mappings are still accurate.
1692 _regalloc->Register_Allocate();
1693
1694 // Bail out if the allocator builds too many nodes
1695 if (failing()) return;
1696 }
1697
1698 // Prior to register allocation we kept empty basic blocks in case the
1699 // the allocator needed a place to spill. After register allocation we
1700 // are not adding any new instructions. If any basic block is empty, we
1701 // can now safely remove it.
1702 {
1703 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1704 cfg.remove_empty();
1705 if (do_freq_based_layout()) {
1706 PhaseBlockLayout layout(cfg);
1707 } else {
1708 cfg.set_loop_alignment();
1709 }
1710 cfg.fixup_flow();
1711 }
1712
1713 // Perform any platform dependent postallocation verifications.
1714 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1715
1716 // Apply peephole optimizations
1717 if( OptoPeephole ) {
1718 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1719 PhasePeephole peep( _regalloc, cfg);
1720 peep.do_transform();
1721 }
1722
1723 // Convert Nodes to instruction bits in a buffer
1724 {
1725 // %%%% workspace merge brought two timers together for one job
1726 TracePhase t2a("output", &_t_output, true);
1727 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1728 Output();
1729 }
1730
1731 print_method("Final Code");
1732
1733 // He's dead, Jim.
1734 _cfg = (PhaseCFG*)0xdeadbeef;
1735 _regalloc = (PhaseChaitin*)0xdeadbeef;
1736 }
1737
1738
1739 //------------------------------dump_asm---------------------------------------
1740 // Dump formatted assembly
1741 #ifndef PRODUCT
1742 void Compile::dump_asm(int *pcs, uint pc_limit) {
1743 bool cut_short = false;
1744 tty->print_cr("#");
1745 tty->print("# "); _tf->dump(); tty->cr();
1746 tty->print_cr("#");
1747
1748 // For all blocks
1749 int pc = 0x0; // Program counter
1750 char starts_bundle = ' ';
1751 _regalloc->dump_frame();
1752
1753 Node *n = NULL;
1754 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1755 if (VMThread::should_terminate()) { cut_short = true; break; }
1756 Block *b = _cfg->_blocks[i];
1757 if (b->is_connector() && !Verbose) continue;
1758 n = b->_nodes[0];
1759 if (pcs && n->_idx < pc_limit)
1760 tty->print("%3.3x ", pcs[n->_idx]);
1761 else
1762 tty->print(" ");
1763 b->dump_head( &_cfg->_bbs );
1764 if (b->is_connector()) {
1765 tty->print_cr(" # Empty connector block");
1766 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1767 tty->print_cr(" # Block is sole successor of call");
1768 }
1769
1770 // For all instructions
1771 Node *delay = NULL;
1772 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1773 if (VMThread::should_terminate()) { cut_short = true; break; }
1774 n = b->_nodes[j];
1775 if (valid_bundle_info(n)) {
1776 Bundle *bundle = node_bundling(n);
1777 if (bundle->used_in_unconditional_delay()) {
1778 delay = n;
1779 continue;
1780 }
1781 if (bundle->starts_bundle())
1782 starts_bundle = '+';
1783 }
1784
1785 if (WizardMode) n->dump();
1786
1787 if( !n->is_Region() && // Dont print in the Assembly
1788 !n->is_Phi() && // a few noisely useless nodes
1789 !n->is_Proj() &&
1790 !n->is_MachTemp() &&
1791 !n->is_Catch() && // Would be nice to print exception table targets
1792 !n->is_MergeMem() && // Not very interesting
1793 !n->is_top() && // Debug info table constants
1794 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1795 ) {
1796 if (pcs && n->_idx < pc_limit)
1797 tty->print("%3.3x", pcs[n->_idx]);
1798 else
1799 tty->print(" ");
1800 tty->print(" %c ", starts_bundle);
1801 starts_bundle = ' ';
1802 tty->print("\t");
1803 n->format(_regalloc, tty);
1804 tty->cr();
1805 }
1806
1807 // If we have an instruction with a delay slot, and have seen a delay,
1808 // then back up and print it
1809 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1810 assert(delay != NULL, "no unconditional delay instruction");
1811 if (WizardMode) delay->dump();
1812
1813 if (node_bundling(delay)->starts_bundle())
1814 starts_bundle = '+';
1815 if (pcs && n->_idx < pc_limit)
1816 tty->print("%3.3x", pcs[n->_idx]);
1817 else
1818 tty->print(" ");
1819 tty->print(" %c ", starts_bundle);
1820 starts_bundle = ' ';
1821 tty->print("\t");
1822 delay->format(_regalloc, tty);
1823 tty->print_cr("");
1824 delay = NULL;
1825 }
1826
1827 // Dump the exception table as well
1828 if( n->is_Catch() && (Verbose || WizardMode) ) {
1829 // Print the exception table for this offset
1830 _handler_table.print_subtable_for(pc);
1831 }
1832 }
1833
1834 if (pcs && n->_idx < pc_limit)
1835 tty->print_cr("%3.3x", pcs[n->_idx]);
1836 else
1837 tty->print_cr("");
1838
1839 assert(cut_short || delay == NULL, "no unconditional delay branch");
1840
1841 } // End of per-block dump
1842 tty->print_cr("");
1843
1844 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1845 }
1846 #endif
1847
1848 //------------------------------Final_Reshape_Counts---------------------------
1849 // This class defines counters to help identify when a method
1850 // may/must be executed using hardware with only 24-bit precision.
1851 struct Final_Reshape_Counts : public StackObj {
1852 int _call_count; // count non-inlined 'common' calls
1853 int _float_count; // count float ops requiring 24-bit precision
1854 int _double_count; // count double ops requiring more precision
1855 int _java_call_count; // count non-inlined 'java' calls
1856 VectorSet _visited; // Visitation flags
1857 Node_List _tests; // Set of IfNodes & PCTableNodes
1858
1859 Final_Reshape_Counts() :
1860 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0),
1861 _visited( Thread::current()->resource_area() ) { }
1862
1863 void inc_call_count () { _call_count ++; }
1864 void inc_float_count () { _float_count ++; }
1865 void inc_double_count() { _double_count++; }
1866 void inc_java_call_count() { _java_call_count++; }
1867
1868 int get_call_count () const { return _call_count ; }
1869 int get_float_count () const { return _float_count ; }
1870 int get_double_count() const { return _double_count; }
1871 int get_java_call_count() const { return _java_call_count; }
1872 };
1873
1874 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1875 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1876 // Make sure the offset goes inside the instance layout.
1877 return k->contains_field_offset(tp->offset());
1878 // Note that OffsetBot and OffsetTop are very negative.
1879 }
1880
1881 //------------------------------final_graph_reshaping_impl----------------------
1882 // Implement items 1-5 from final_graph_reshaping below.
1883 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
1884
1885 if ( n->outcnt() == 0 ) return; // dead node
1886 uint nop = n->Opcode();
1887
1888 // Check for 2-input instruction with "last use" on right input.
1889 // Swap to left input. Implements item (2).
1890 if( n->req() == 3 && // two-input instruction
1891 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1892 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1893 n->in(2)->outcnt() == 1 &&// right use IS a last use
1894 !n->in(2)->is_Con() ) { // right use is not a constant
1895 // Check for commutative opcode
1896 switch( nop ) {
1897 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1898 case Op_MaxI: case Op_MinI:
1899 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1900 case Op_AndL: case Op_XorL: case Op_OrL:
1901 case Op_AndI: case Op_XorI: case Op_OrI: {
1902 // Move "last use" input to left by swapping inputs
1903 n->swap_edges(1, 2);
1904 break;
1905 }
1906 default:
1907 break;
1908 }
1909 }
1910
1911 // Count FPU ops and common calls, implements item (3)
1912 switch( nop ) {
1913 // Count all float operations that may use FPU
1914 case Op_AddF:
1915 case Op_SubF:
1916 case Op_MulF:
1917 case Op_DivF:
1918 case Op_NegF:
1919 case Op_ModF:
1920 case Op_ConvI2F:
1921 case Op_ConF:
1922 case Op_CmpF:
1923 case Op_CmpF3:
1924 // case Op_ConvL2F: // longs are split into 32-bit halves
1925 fpu.inc_float_count();
1926 break;
1927
1928 case Op_ConvF2D:
1929 case Op_ConvD2F:
1930 fpu.inc_float_count();
1931 fpu.inc_double_count();
1932 break;
1933
1934 // Count all double operations that may use FPU
1935 case Op_AddD:
1936 case Op_SubD:
1937 case Op_MulD:
1938 case Op_DivD:
1939 case Op_NegD:
1940 case Op_ModD:
1941 case Op_ConvI2D:
1942 case Op_ConvD2I:
1943 // case Op_ConvL2D: // handled by leaf call
1944 // case Op_ConvD2L: // handled by leaf call
1945 case Op_ConD:
1946 case Op_CmpD:
1947 case Op_CmpD3:
1948 fpu.inc_double_count();
1949 break;
1950 case Op_Opaque1: // Remove Opaque Nodes before matching
1951 case Op_Opaque2: // Remove Opaque Nodes before matching
1952 n->subsume_by(n->in(1));
1953 break;
1954 case Op_CallStaticJava:
1955 case Op_CallJava:
1956 case Op_CallDynamicJava:
1957 fpu.inc_java_call_count(); // Count java call site;
1958 case Op_CallRuntime:
1959 case Op_CallLeaf:
1960 case Op_CallLeafNoFP: {
1961 assert( n->is_Call(), "" );
1962 CallNode *call = n->as_Call();
1963 // Count call sites where the FP mode bit would have to be flipped.
1964 // Do not count uncommon runtime calls:
1965 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
1966 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
1967 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
1968 fpu.inc_call_count(); // Count the call site
1969 } else { // See if uncommon argument is shared
1970 Node *n = call->in(TypeFunc::Parms);
1971 int nop = n->Opcode();
1972 // Clone shared simple arguments to uncommon calls, item (1).
1973 if( n->outcnt() > 1 &&
1974 !n->is_Proj() &&
1975 nop != Op_CreateEx &&
1976 nop != Op_CheckCastPP &&
1977 nop != Op_DecodeN &&
1978 !n->is_Mem() ) {
1979 Node *x = n->clone();
1980 call->set_req( TypeFunc::Parms, x );
1981 }
1982 }
1983 break;
1984 }
1985
1986 case Op_StoreD:
1987 case Op_LoadD:
1988 case Op_LoadD_unaligned:
1989 fpu.inc_double_count();
1990 goto handle_mem;
1991 case Op_StoreF:
1992 case Op_LoadF:
1993 fpu.inc_float_count();
1994 goto handle_mem;
1995
1996 case Op_StoreB:
1997 case Op_StoreC:
1998 case Op_StoreCM:
1999 case Op_StorePConditional:
2000 case Op_StoreI:
2001 case Op_StoreL:
2002 case Op_StoreIConditional:
2003 case Op_StoreLConditional:
2004 case Op_CompareAndSwapI:
2005 case Op_CompareAndSwapL:
2006 case Op_CompareAndSwapP:
2007 case Op_CompareAndSwapN:
2008 case Op_StoreP:
2009 case Op_StoreN:
2010 case Op_LoadB:
2011 case Op_LoadC:
2012 case Op_LoadI:
2013 case Op_LoadKlass:
2014 case Op_LoadNKlass:
2015 case Op_LoadL:
2016 case Op_LoadL_unaligned:
2017 case Op_LoadPLocked:
2018 case Op_LoadLLocked:
2019 case Op_LoadP:
2020 case Op_LoadN:
2021 case Op_LoadRange:
2022 case Op_LoadS: {
2023 handle_mem:
2024 #ifdef ASSERT
2025 if( VerifyOptoOopOffsets ) {
2026 assert( n->is_Mem(), "" );
2027 MemNode *mem = (MemNode*)n;
2028 // Check to see if address types have grounded out somehow.
2029 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2030 assert( !tp || oop_offset_is_sane(tp), "" );
2031 }
2032 #endif
2033 break;
2034 }
2035
2036 case Op_AddP: { // Assert sane base pointers
2037 Node *addp = n->in(AddPNode::Address);
2038 assert( !addp->is_AddP() ||
2039 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2040 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2041 "Base pointers must match" );
2042 #ifdef _LP64
2043 if (UseCompressedOops &&
2044 addp->Opcode() == Op_ConP &&
2045 addp == n->in(AddPNode::Base) &&
2046 n->in(AddPNode::Offset)->is_Con()) {
2047 // Use addressing with narrow klass to load with offset on x86.
2048 // On sparc loading 32-bits constant and decoding it have less
2049 // instructions (4) then load 64-bits constant (7).
2050 // Do this transformation here since IGVN will convert ConN back to ConP.
2051 const Type* t = addp->bottom_type();
2052 if (t->isa_oopptr()) {
2053 Node* nn = NULL;
2054
2055 // Look for existing ConN node of the same exact type.
2056 Compile* C = Compile::current();
2057 Node* r = C->root();
2058 uint cnt = r->outcnt();
2059 for (uint i = 0; i < cnt; i++) {
2060 Node* m = r->raw_out(i);
2061 if (m!= NULL && m->Opcode() == Op_ConN &&
2062 m->bottom_type()->make_ptr() == t) {
2063 nn = m;
2064 break;
2065 }
2066 }
2067 if (nn != NULL) {
2068 // Decode a narrow oop to match address
2069 // [R12 + narrow_oop_reg<<3 + offset]
2070 nn = new (C, 2) DecodeNNode(nn, t);
2071 n->set_req(AddPNode::Base, nn);
2072 n->set_req(AddPNode::Address, nn);
2073 if (addp->outcnt() == 0) {
2074 addp->disconnect_inputs(NULL);
2075 }
2076 }
2077 }
2078 }
2079 #endif
2080 break;
2081 }
2082
2083 #ifdef _LP64
2084 case Op_CastPP:
2085 if (n->in(1)->is_DecodeN() && UseImplicitNullCheckForNarrowOop) {
2086 Compile* C = Compile::current();
2087 Node* in1 = n->in(1);
2088 const Type* t = n->bottom_type();
2089 Node* new_in1 = in1->clone();
2090 new_in1->as_DecodeN()->set_type(t);
2091
2092 if (!Matcher::clone_shift_expressions) {
2093 //
2094 // x86, ARM and friends can handle 2 adds in addressing mode
2095 // and Matcher can fold a DecodeN node into address by using
2096 // a narrow oop directly and do implicit NULL check in address:
2097 //
2098 // [R12 + narrow_oop_reg<<3 + offset]
2099 // NullCheck narrow_oop_reg
2100 //
2101 // On other platforms (Sparc) we have to keep new DecodeN node and
2102 // use it to do implicit NULL check in address:
2103 //
2104 // decode_not_null narrow_oop_reg, base_reg
2105 // [base_reg + offset]
2106 // NullCheck base_reg
2107 //
2108 // Pin the new DecodeN node to non-null path on these patforms (Sparc)
2109 // to keep the information to which NULL check the new DecodeN node
2110 // corresponds to use it as value in implicit_null_check().
2111 //
2112 new_in1->set_req(0, n->in(0));
2113 }
2114
2115 n->subsume_by(new_in1);
2116 if (in1->outcnt() == 0) {
2117 in1->disconnect_inputs(NULL);
2118 }
2119 }
2120 break;
2121
2122 case Op_CmpP:
2123 // Do this transformation here to preserve CmpPNode::sub() and
2124 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2125 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2126 Node* in1 = n->in(1);
2127 Node* in2 = n->in(2);
2128 if (!in1->is_DecodeN()) {
2129 in2 = in1;
2130 in1 = n->in(2);
2131 }
2132 assert(in1->is_DecodeN(), "sanity");
2133
2134 Compile* C = Compile::current();
2135 Node* new_in2 = NULL;
2136 if (in2->is_DecodeN()) {
2137 new_in2 = in2->in(1);
2138 } else if (in2->Opcode() == Op_ConP) {
2139 const Type* t = in2->bottom_type();
2140 if (t == TypePtr::NULL_PTR && UseImplicitNullCheckForNarrowOop) {
2141 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2142 //
2143 // This transformation together with CastPP transformation above
2144 // will generated code for implicit NULL checks for compressed oops.
2145 //
2146 // The original code after Optimize()
2147 //
2148 // LoadN memory, narrow_oop_reg
2149 // decode narrow_oop_reg, base_reg
2150 // CmpP base_reg, NULL
2151 // CastPP base_reg // NotNull
2152 // Load [base_reg + offset], val_reg
2153 //
2154 // after these transformations will be
2155 //
2156 // LoadN memory, narrow_oop_reg
2157 // CmpN narrow_oop_reg, NULL
2158 // decode_not_null narrow_oop_reg, base_reg
2159 // Load [base_reg + offset], val_reg
2160 //
2161 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2162 // since narrow oops can be used in debug info now (see the code in
2163 // final_graph_reshaping_walk()).
2164 //
2165 // At the end the code will be matched to
2166 // on x86:
2167 //
2168 // Load_narrow_oop memory, narrow_oop_reg
2169 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2170 // NullCheck narrow_oop_reg
2171 //
2172 // and on sparc:
2173 //
2174 // Load_narrow_oop memory, narrow_oop_reg
2175 // decode_not_null narrow_oop_reg, base_reg
2176 // Load [base_reg + offset], val_reg
2177 // NullCheck base_reg
2178 //
2179 } else if (t->isa_oopptr()) {
2180 new_in2 = ConNode::make(C, t->make_narrowoop());
2181 }
2182 }
2183 if (new_in2 != NULL) {
2184 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2185 n->subsume_by( cmpN );
2186 if (in1->outcnt() == 0) {
2187 in1->disconnect_inputs(NULL);
2188 }
2189 if (in2->outcnt() == 0) {
2190 in2->disconnect_inputs(NULL);
2191 }
2192 }
2193 }
2194 break;
2195
2196 case Op_DecodeN:
2197 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2198 // DecodeN could be pinned on Sparc where it can't be fold into
2199 // an address expression, see the code for Op_CastPP above.
2200 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc");
2201 break;
2202
2203 case Op_EncodeP: {
2204 Node* in1 = n->in(1);
2205 if (in1->is_DecodeN()) {
2206 n->subsume_by(in1->in(1));
2207 } else if (in1->Opcode() == Op_ConP) {
2208 Compile* C = Compile::current();
2209 const Type* t = in1->bottom_type();
2210 if (t == TypePtr::NULL_PTR) {
2211 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2212 } else if (t->isa_oopptr()) {
2213 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2214 }
2215 }
2216 if (in1->outcnt() == 0) {
2217 in1->disconnect_inputs(NULL);
2218 }
2219 break;
2220 }
2221
2222 case Op_Phi:
2223 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2224 // The EncodeP optimization may create Phi with the same edges
2225 // for all paths. It is not handled well by Register Allocator.
2226 Node* unique_in = n->in(1);
2227 assert(unique_in != NULL, "");
2228 uint cnt = n->req();
2229 for (uint i = 2; i < cnt; i++) {
2230 Node* m = n->in(i);
2231 assert(m != NULL, "");
2232 if (unique_in != m)
2233 unique_in = NULL;
2234 }
2235 if (unique_in != NULL) {
2236 n->subsume_by(unique_in);
2237 }
2238 }
2239 break;
2240
2241 #endif
2242
2243 case Op_ModI:
2244 if (UseDivMod) {
2245 // Check if a%b and a/b both exist
2246 Node* d = n->find_similar(Op_DivI);
2247 if (d) {
2248 // Replace them with a fused divmod if supported
2249 Compile* C = Compile::current();
2250 if (Matcher::has_match_rule(Op_DivModI)) {
2251 DivModINode* divmod = DivModINode::make(C, n);
2252 d->subsume_by(divmod->div_proj());
2253 n->subsume_by(divmod->mod_proj());
2254 } else {
2255 // replace a%b with a-((a/b)*b)
2256 Node* mult = new (C, 3) MulINode(d, d->in(2));
2257 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2258 n->subsume_by( sub );
2259 }
2260 }
2261 }
2262 break;
2263
2264 case Op_ModL:
2265 if (UseDivMod) {
2266 // Check if a%b and a/b both exist
2267 Node* d = n->find_similar(Op_DivL);
2268 if (d) {
2269 // Replace them with a fused divmod if supported
2270 Compile* C = Compile::current();
2271 if (Matcher::has_match_rule(Op_DivModL)) {
2272 DivModLNode* divmod = DivModLNode::make(C, n);
2273 d->subsume_by(divmod->div_proj());
2274 n->subsume_by(divmod->mod_proj());
2275 } else {
2276 // replace a%b with a-((a/b)*b)
2277 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2278 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2279 n->subsume_by( sub );
2280 }
2281 }
2282 }
2283 break;
2284
2285 case Op_Load16B:
2286 case Op_Load8B:
2287 case Op_Load4B:
2288 case Op_Load8S:
2289 case Op_Load4S:
2290 case Op_Load2S:
2291 case Op_Load8C:
2292 case Op_Load4C:
2293 case Op_Load2C:
2294 case Op_Load4I:
2295 case Op_Load2I:
2296 case Op_Load2L:
2297 case Op_Load4F:
2298 case Op_Load2F:
2299 case Op_Load2D:
2300 case Op_Store16B:
2301 case Op_Store8B:
2302 case Op_Store4B:
2303 case Op_Store8C:
2304 case Op_Store4C:
2305 case Op_Store2C:
2306 case Op_Store4I:
2307 case Op_Store2I:
2308 case Op_Store2L:
2309 case Op_Store4F:
2310 case Op_Store2F:
2311 case Op_Store2D:
2312 break;
2313
2314 case Op_PackB:
2315 case Op_PackS:
2316 case Op_PackC:
2317 case Op_PackI:
2318 case Op_PackF:
2319 case Op_PackL:
2320 case Op_PackD:
2321 if (n->req()-1 > 2) {
2322 // Replace many operand PackNodes with a binary tree for matching
2323 PackNode* p = (PackNode*) n;
2324 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2325 n->subsume_by(btp);
2326 }
2327 break;
2328 default:
2329 assert( !n->is_Call(), "" );
2330 assert( !n->is_Mem(), "" );
2331 break;
2332 }
2333
2334 // Collect CFG split points
2335 if (n->is_MultiBranch())
2336 fpu._tests.push(n);
2337 }
2338
2339 //------------------------------final_graph_reshaping_walk---------------------
2340 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2341 // requires that the walk visits a node's inputs before visiting the node.
2342 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) {
2343 ResourceArea *area = Thread::current()->resource_area();
2344 Unique_Node_List sfpt(area);
2345
2346 fpu._visited.set(root->_idx); // first, mark node as visited
2347 uint cnt = root->req();
2348 Node *n = root;
2349 uint i = 0;
2350 while (true) {
2351 if (i < cnt) {
2352 // Place all non-visited non-null inputs onto stack
2353 Node* m = n->in(i);
2354 ++i;
2355 if (m != NULL && !fpu._visited.test_set(m->_idx)) {
2356 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2357 sfpt.push(m);
2358 cnt = m->req();
2359 nstack.push(n, i); // put on stack parent and next input's index
2360 n = m;
2361 i = 0;
2362 }
2363 } else {
2364 // Now do post-visit work
2365 final_graph_reshaping_impl( n, fpu );
2366 if (nstack.is_empty())
2367 break; // finished
2368 n = nstack.node(); // Get node from stack
2369 cnt = n->req();
2370 i = nstack.index();
2371 nstack.pop(); // Shift to the next node on stack
2372 }
2373 }
2374
2375 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2376 // It could be done for an uncommon traps or any safepoints/calls
2377 // if the DecodeN node is referenced only in a debug info.
2378 while (sfpt.size() > 0) {
2379 n = sfpt.pop();
2380 JVMState *jvms = n->as_SafePoint()->jvms();
2381 assert(jvms != NULL, "sanity");
2382 int start = jvms->debug_start();
2383 int end = n->req();
2384 bool is_uncommon = (n->is_CallStaticJava() &&
2385 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2386 for (int j = start; j < end; j++) {
2387 Node* in = n->in(j);
2388 if (in->is_DecodeN()) {
2389 bool safe_to_skip = true;
2390 if (!is_uncommon ) {
2391 // Is it safe to skip?
2392 for (uint i = 0; i < in->outcnt(); i++) {
2393 Node* u = in->raw_out(i);
2394 if (!u->is_SafePoint() ||
2395 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2396 safe_to_skip = false;
2397 }
2398 }
2399 }
2400 if (safe_to_skip) {
2401 n->set_req(j, in->in(1));
2402 }
2403 if (in->outcnt() == 0) {
2404 in->disconnect_inputs(NULL);
2405 }
2406 }
2407 }
2408 }
2409 }
2410
2411 //------------------------------final_graph_reshaping--------------------------
2412 // Final Graph Reshaping.
2413 //
2414 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2415 // and not commoned up and forced early. Must come after regular
2416 // optimizations to avoid GVN undoing the cloning. Clone constant
2417 // inputs to Loop Phis; these will be split by the allocator anyways.
2418 // Remove Opaque nodes.
2419 // (2) Move last-uses by commutative operations to the left input to encourage
2420 // Intel update-in-place two-address operations and better register usage
2421 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2422 // calls canonicalizing them back.
2423 // (3) Count the number of double-precision FP ops, single-precision FP ops
2424 // and call sites. On Intel, we can get correct rounding either by
2425 // forcing singles to memory (requires extra stores and loads after each
2426 // FP bytecode) or we can set a rounding mode bit (requires setting and
2427 // clearing the mode bit around call sites). The mode bit is only used
2428 // if the relative frequency of single FP ops to calls is low enough.
2429 // This is a key transform for SPEC mpeg_audio.
2430 // (4) Detect infinite loops; blobs of code reachable from above but not
2431 // below. Several of the Code_Gen algorithms fail on such code shapes,
2432 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2433 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2434 // Detection is by looking for IfNodes where only 1 projection is
2435 // reachable from below or CatchNodes missing some targets.
2436 // (5) Assert for insane oop offsets in debug mode.
2437
2438 bool Compile::final_graph_reshaping() {
2439 // an infinite loop may have been eliminated by the optimizer,
2440 // in which case the graph will be empty.
2441 if (root()->req() == 1) {
2442 record_method_not_compilable("trivial infinite loop");
2443 return true;
2444 }
2445
2446 Final_Reshape_Counts fpu;
2447
2448 // Visit everybody reachable!
2449 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2450 Node_Stack nstack(unique() >> 1);
2451 final_graph_reshaping_walk(nstack, root(), fpu);
2452
2453 // Check for unreachable (from below) code (i.e., infinite loops).
2454 for( uint i = 0; i < fpu._tests.size(); i++ ) {
2455 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch();
2456 // Get number of CFG targets.
2457 // Note that PCTables include exception targets after calls.
2458 uint required_outcnt = n->required_outcnt();
2459 if (n->outcnt() != required_outcnt) {
2460 // Check for a few special cases. Rethrow Nodes never take the
2461 // 'fall-thru' path, so expected kids is 1 less.
2462 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2463 if (n->in(0)->in(0)->is_Call()) {
2464 CallNode *call = n->in(0)->in(0)->as_Call();
2465 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2466 required_outcnt--; // Rethrow always has 1 less kid
2467 } else if (call->req() > TypeFunc::Parms &&
2468 call->is_CallDynamicJava()) {
2469 // Check for null receiver. In such case, the optimizer has
2470 // detected that the virtual call will always result in a null
2471 // pointer exception. The fall-through projection of this CatchNode
2472 // will not be populated.
2473 Node *arg0 = call->in(TypeFunc::Parms);
2474 if (arg0->is_Type() &&
2475 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2476 required_outcnt--;
2477 }
2478 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2479 call->req() > TypeFunc::Parms+1 &&
2480 call->is_CallStaticJava()) {
2481 // Check for negative array length. In such case, the optimizer has
2482 // detected that the allocation attempt will always result in an
2483 // exception. There is no fall-through projection of this CatchNode .
2484 Node *arg1 = call->in(TypeFunc::Parms+1);
2485 if (arg1->is_Type() &&
2486 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2487 required_outcnt--;
2488 }
2489 }
2490 }
2491 }
2492 // Recheck with a better notion of 'required_outcnt'
2493 if (n->outcnt() != required_outcnt) {
2494 record_method_not_compilable("malformed control flow");
2495 return true; // Not all targets reachable!
2496 }
2497 }
2498 // Check that I actually visited all kids. Unreached kids
2499 // must be infinite loops.
2500 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2501 if (!fpu._visited.test(n->fast_out(j)->_idx)) {
2502 record_method_not_compilable("infinite loop");
2503 return true; // Found unvisited kid; must be unreach
2504 }
2505 }
2506
2507 // If original bytecodes contained a mixture of floats and doubles
2508 // check if the optimizer has made it homogenous, item (3).
2509 if( Use24BitFPMode && Use24BitFP &&
2510 fpu.get_float_count() > 32 &&
2511 fpu.get_double_count() == 0 &&
2512 (10 * fpu.get_call_count() < fpu.get_float_count()) ) {
2513 set_24_bit_selection_and_mode( false, true );
2514 }
2515
2516 set_has_java_calls(fpu.get_java_call_count() > 0);
2517
2518 // No infinite loops, no reason to bail out.
2519 return false;
2520 }
2521
2522 //-----------------------------too_many_traps----------------------------------
2523 // Report if there are too many traps at the current method and bci.
2524 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2525 bool Compile::too_many_traps(ciMethod* method,
2526 int bci,
2527 Deoptimization::DeoptReason reason) {
2528 ciMethodData* md = method->method_data();
2529 if (md->is_empty()) {
2530 // Assume the trap has not occurred, or that it occurred only
2531 // because of a transient condition during start-up in the interpreter.
2532 return false;
2533 }
2534 if (md->has_trap_at(bci, reason) != 0) {
2535 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2536 // Also, if there are multiple reasons, or if there is no per-BCI record,
2537 // assume the worst.
2538 if (log())
2539 log()->elem("observe trap='%s' count='%d'",
2540 Deoptimization::trap_reason_name(reason),
2541 md->trap_count(reason));
2542 return true;
2543 } else {
2544 // Ignore method/bci and see if there have been too many globally.
2545 return too_many_traps(reason, md);
2546 }
2547 }
2548
2549 // Less-accurate variant which does not require a method and bci.
2550 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2551 ciMethodData* logmd) {
2552 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2553 // Too many traps globally.
2554 // Note that we use cumulative trap_count, not just md->trap_count.
2555 if (log()) {
2556 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2557 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2558 Deoptimization::trap_reason_name(reason),
2559 mcount, trap_count(reason));
2560 }
2561 return true;
2562 } else {
2563 // The coast is clear.
2564 return false;
2565 }
2566 }
2567
2568 //--------------------------too_many_recompiles--------------------------------
2569 // Report if there are too many recompiles at the current method and bci.
2570 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2571 // Is not eager to return true, since this will cause the compiler to use
2572 // Action_none for a trap point, to avoid too many recompilations.
2573 bool Compile::too_many_recompiles(ciMethod* method,
2574 int bci,
2575 Deoptimization::DeoptReason reason) {
2576 ciMethodData* md = method->method_data();
2577 if (md->is_empty()) {
2578 // Assume the trap has not occurred, or that it occurred only
2579 // because of a transient condition during start-up in the interpreter.
2580 return false;
2581 }
2582 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2583 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2584 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2585 Deoptimization::DeoptReason per_bc_reason
2586 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2587 if ((per_bc_reason == Deoptimization::Reason_none
2588 || md->has_trap_at(bci, reason) != 0)
2589 // The trap frequency measure we care about is the recompile count:
2590 && md->trap_recompiled_at(bci)
2591 && md->overflow_recompile_count() >= bc_cutoff) {
2592 // Do not emit a trap here if it has already caused recompilations.
2593 // Also, if there are multiple reasons, or if there is no per-BCI record,
2594 // assume the worst.
2595 if (log())
2596 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2597 Deoptimization::trap_reason_name(reason),
2598 md->trap_count(reason),
2599 md->overflow_recompile_count());
2600 return true;
2601 } else if (trap_count(reason) != 0
2602 && decompile_count() >= m_cutoff) {
2603 // Too many recompiles globally, and we have seen this sort of trap.
2604 // Use cumulative decompile_count, not just md->decompile_count.
2605 if (log())
2606 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2607 Deoptimization::trap_reason_name(reason),
2608 md->trap_count(reason), trap_count(reason),
2609 md->decompile_count(), decompile_count());
2610 return true;
2611 } else {
2612 // The coast is clear.
2613 return false;
2614 }
2615 }
2616
2617
2618 #ifndef PRODUCT
2619 //------------------------------verify_graph_edges---------------------------
2620 // Walk the Graph and verify that there is a one-to-one correspondence
2621 // between Use-Def edges and Def-Use edges in the graph.
2622 void Compile::verify_graph_edges(bool no_dead_code) {
2623 if (VerifyGraphEdges) {
2624 ResourceArea *area = Thread::current()->resource_area();
2625 Unique_Node_List visited(area);
2626 // Call recursive graph walk to check edges
2627 _root->verify_edges(visited);
2628 if (no_dead_code) {
2629 // Now make sure that no visited node is used by an unvisited node.
2630 bool dead_nodes = 0;
2631 Unique_Node_List checked(area);
2632 while (visited.size() > 0) {
2633 Node* n = visited.pop();
2634 checked.push(n);
2635 for (uint i = 0; i < n->outcnt(); i++) {
2636 Node* use = n->raw_out(i);
2637 if (checked.member(use)) continue; // already checked
2638 if (visited.member(use)) continue; // already in the graph
2639 if (use->is_Con()) continue; // a dead ConNode is OK
2640 // At this point, we have found a dead node which is DU-reachable.
2641 if (dead_nodes++ == 0)
2642 tty->print_cr("*** Dead nodes reachable via DU edges:");
2643 use->dump(2);
2644 tty->print_cr("---");
2645 checked.push(use); // No repeats; pretend it is now checked.
2646 }
2647 }
2648 assert(dead_nodes == 0, "using nodes must be reachable from root");
2649 }
2650 }
2651 }
2652 #endif
2653
2654 // The Compile object keeps track of failure reasons separately from the ciEnv.
2655 // This is required because there is not quite a 1-1 relation between the
2656 // ciEnv and its compilation task and the Compile object. Note that one
2657 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2658 // to backtrack and retry without subsuming loads. Other than this backtracking
2659 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2660 // by the logic in C2Compiler.
2661 void Compile::record_failure(const char* reason) {
2662 if (log() != NULL) {
2663 log()->elem("failure reason='%s' phase='compile'", reason);
2664 }
2665 if (_failure_reason == NULL) {
2666 // Record the first failure reason.
2667 _failure_reason = reason;
2668 }
2669 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2670 C->print_method(_failure_reason);
2671 }
2672 _root = NULL; // flush the graph, too
2673 }
2674
2675 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2676 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2677 {
2678 if (dolog) {
2679 C = Compile::current();
2680 _log = C->log();
2681 } else {
2682 C = NULL;
2683 _log = NULL;
2684 }
2685 if (_log != NULL) {
2686 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2687 _log->stamp();
2688 _log->end_head();
2689 }
2690 }
2691
2692 Compile::TracePhase::~TracePhase() {
2693 if (_log != NULL) {
2694 _log->done("phase nodes='%d'", C->unique());
2695 }
2696 }