Thu, 12 Nov 2009 09:24:21 -0800
6892658: C2 should optimize some stringbuilder patterns
Reviewed-by: kvn, twisti
1 /*
2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_compile.cpp.incl"
28 /// Support for intrinsics.
30 // Return the index at which m must be inserted (or already exists).
31 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
33 #ifdef ASSERT
34 for (int i = 1; i < _intrinsics->length(); i++) {
35 CallGenerator* cg1 = _intrinsics->at(i-1);
36 CallGenerator* cg2 = _intrinsics->at(i);
37 assert(cg1->method() != cg2->method()
38 ? cg1->method() < cg2->method()
39 : cg1->is_virtual() < cg2->is_virtual(),
40 "compiler intrinsics list must stay sorted");
41 }
42 #endif
43 // Binary search sorted list, in decreasing intervals [lo, hi].
44 int lo = 0, hi = _intrinsics->length()-1;
45 while (lo <= hi) {
46 int mid = (uint)(hi + lo) / 2;
47 ciMethod* mid_m = _intrinsics->at(mid)->method();
48 if (m < mid_m) {
49 hi = mid-1;
50 } else if (m > mid_m) {
51 lo = mid+1;
52 } else {
53 // look at minor sort key
54 bool mid_virt = _intrinsics->at(mid)->is_virtual();
55 if (is_virtual < mid_virt) {
56 hi = mid-1;
57 } else if (is_virtual > mid_virt) {
58 lo = mid+1;
59 } else {
60 return mid; // exact match
61 }
62 }
63 }
64 return lo; // inexact match
65 }
67 void Compile::register_intrinsic(CallGenerator* cg) {
68 if (_intrinsics == NULL) {
69 _intrinsics = new GrowableArray<CallGenerator*>(60);
70 }
71 // This code is stolen from ciObjectFactory::insert.
72 // Really, GrowableArray should have methods for
73 // insert_at, remove_at, and binary_search.
74 int len = _intrinsics->length();
75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
76 if (index == len) {
77 _intrinsics->append(cg);
78 } else {
79 #ifdef ASSERT
80 CallGenerator* oldcg = _intrinsics->at(index);
81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
82 #endif
83 _intrinsics->append(_intrinsics->at(len-1));
84 int pos;
85 for (pos = len-2; pos >= index; pos--) {
86 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
87 }
88 _intrinsics->at_put(index, cg);
89 }
90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
91 }
93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
94 assert(m->is_loaded(), "don't try this on unloaded methods");
95 if (_intrinsics != NULL) {
96 int index = intrinsic_insertion_index(m, is_virtual);
97 if (index < _intrinsics->length()
98 && _intrinsics->at(index)->method() == m
99 && _intrinsics->at(index)->is_virtual() == is_virtual) {
100 return _intrinsics->at(index);
101 }
102 }
103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
104 if (m->intrinsic_id() != vmIntrinsics::_none &&
105 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
106 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
107 if (cg != NULL) {
108 // Save it for next time:
109 register_intrinsic(cg);
110 return cg;
111 } else {
112 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
113 }
114 }
115 return NULL;
116 }
118 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
119 // in library_call.cpp.
122 #ifndef PRODUCT
123 // statistics gathering...
125 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
126 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
128 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
129 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
130 int oflags = _intrinsic_hist_flags[id];
131 assert(flags != 0, "what happened?");
132 if (is_virtual) {
133 flags |= _intrinsic_virtual;
134 }
135 bool changed = (flags != oflags);
136 if ((flags & _intrinsic_worked) != 0) {
137 juint count = (_intrinsic_hist_count[id] += 1);
138 if (count == 1) {
139 changed = true; // first time
140 }
141 // increment the overall count also:
142 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
143 }
144 if (changed) {
145 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
146 // Something changed about the intrinsic's virtuality.
147 if ((flags & _intrinsic_virtual) != 0) {
148 // This is the first use of this intrinsic as a virtual call.
149 if (oflags != 0) {
150 // We already saw it as a non-virtual, so note both cases.
151 flags |= _intrinsic_both;
152 }
153 } else if ((oflags & _intrinsic_both) == 0) {
154 // This is the first use of this intrinsic as a non-virtual
155 flags |= _intrinsic_both;
156 }
157 }
158 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
159 }
160 // update the overall flags also:
161 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
162 return changed;
163 }
165 static char* format_flags(int flags, char* buf) {
166 buf[0] = 0;
167 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
168 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
169 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
170 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
171 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
172 if (buf[0] == 0) strcat(buf, ",");
173 assert(buf[0] == ',', "must be");
174 return &buf[1];
175 }
177 void Compile::print_intrinsic_statistics() {
178 char flagsbuf[100];
179 ttyLocker ttyl;
180 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
181 tty->print_cr("Compiler intrinsic usage:");
182 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
183 if (total == 0) total = 1; // avoid div0 in case of no successes
184 #define PRINT_STAT_LINE(name, c, f) \
185 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
186 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
187 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
188 int flags = _intrinsic_hist_flags[id];
189 juint count = _intrinsic_hist_count[id];
190 if ((flags | count) != 0) {
191 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
192 }
193 }
194 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
195 if (xtty != NULL) xtty->tail("statistics");
196 }
198 void Compile::print_statistics() {
199 { ttyLocker ttyl;
200 if (xtty != NULL) xtty->head("statistics type='opto'");
201 Parse::print_statistics();
202 PhaseCCP::print_statistics();
203 PhaseRegAlloc::print_statistics();
204 Scheduling::print_statistics();
205 PhasePeephole::print_statistics();
206 PhaseIdealLoop::print_statistics();
207 if (xtty != NULL) xtty->tail("statistics");
208 }
209 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
210 // put this under its own <statistics> element.
211 print_intrinsic_statistics();
212 }
213 }
214 #endif //PRODUCT
216 // Support for bundling info
217 Bundle* Compile::node_bundling(const Node *n) {
218 assert(valid_bundle_info(n), "oob");
219 return &_node_bundling_base[n->_idx];
220 }
222 bool Compile::valid_bundle_info(const Node *n) {
223 return (_node_bundling_limit > n->_idx);
224 }
227 void Compile::gvn_replace_by(Node* n, Node* nn) {
228 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
229 Node* use = n->last_out(i);
230 bool is_in_table = initial_gvn()->hash_delete(use);
231 uint uses_found = 0;
232 for (uint j = 0; j < use->len(); j++) {
233 if (use->in(j) == n) {
234 if (j < use->req())
235 use->set_req(j, nn);
236 else
237 use->set_prec(j, nn);
238 uses_found++;
239 }
240 }
241 if (is_in_table) {
242 // reinsert into table
243 initial_gvn()->hash_find_insert(use);
244 }
245 record_for_igvn(use);
246 i -= uses_found; // we deleted 1 or more copies of this edge
247 }
248 }
253 // Identify all nodes that are reachable from below, useful.
254 // Use breadth-first pass that records state in a Unique_Node_List,
255 // recursive traversal is slower.
256 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
257 int estimated_worklist_size = unique();
258 useful.map( estimated_worklist_size, NULL ); // preallocate space
260 // Initialize worklist
261 if (root() != NULL) { useful.push(root()); }
262 // If 'top' is cached, declare it useful to preserve cached node
263 if( cached_top_node() ) { useful.push(cached_top_node()); }
265 // Push all useful nodes onto the list, breadthfirst
266 for( uint next = 0; next < useful.size(); ++next ) {
267 assert( next < unique(), "Unique useful nodes < total nodes");
268 Node *n = useful.at(next);
269 uint max = n->len();
270 for( uint i = 0; i < max; ++i ) {
271 Node *m = n->in(i);
272 if( m == NULL ) continue;
273 useful.push(m);
274 }
275 }
276 }
278 // Disconnect all useless nodes by disconnecting those at the boundary.
279 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
280 uint next = 0;
281 while( next < useful.size() ) {
282 Node *n = useful.at(next++);
283 // Use raw traversal of out edges since this code removes out edges
284 int max = n->outcnt();
285 for (int j = 0; j < max; ++j ) {
286 Node* child = n->raw_out(j);
287 if( ! useful.member(child) ) {
288 assert( !child->is_top() || child != top(),
289 "If top is cached in Compile object it is in useful list");
290 // Only need to remove this out-edge to the useless node
291 n->raw_del_out(j);
292 --j;
293 --max;
294 }
295 }
296 if (n->outcnt() == 1 && n->has_special_unique_user()) {
297 record_for_igvn( n->unique_out() );
298 }
299 }
300 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
301 }
303 //------------------------------frame_size_in_words-----------------------------
304 // frame_slots in units of words
305 int Compile::frame_size_in_words() const {
306 // shift is 0 in LP32 and 1 in LP64
307 const int shift = (LogBytesPerWord - LogBytesPerInt);
308 int words = _frame_slots >> shift;
309 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
310 return words;
311 }
313 // ============================================================================
314 //------------------------------CompileWrapper---------------------------------
315 class CompileWrapper : public StackObj {
316 Compile *const _compile;
317 public:
318 CompileWrapper(Compile* compile);
320 ~CompileWrapper();
321 };
323 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
324 // the Compile* pointer is stored in the current ciEnv:
325 ciEnv* env = compile->env();
326 assert(env == ciEnv::current(), "must already be a ciEnv active");
327 assert(env->compiler_data() == NULL, "compile already active?");
328 env->set_compiler_data(compile);
329 assert(compile == Compile::current(), "sanity");
331 compile->set_type_dict(NULL);
332 compile->set_type_hwm(NULL);
333 compile->set_type_last_size(0);
334 compile->set_last_tf(NULL, NULL);
335 compile->set_indexSet_arena(NULL);
336 compile->set_indexSet_free_block_list(NULL);
337 compile->init_type_arena();
338 Type::Initialize(compile);
339 _compile->set_scratch_buffer_blob(NULL);
340 _compile->begin_method();
341 }
342 CompileWrapper::~CompileWrapper() {
343 _compile->end_method();
344 if (_compile->scratch_buffer_blob() != NULL)
345 BufferBlob::free(_compile->scratch_buffer_blob());
346 _compile->env()->set_compiler_data(NULL);
347 }
350 //----------------------------print_compile_messages---------------------------
351 void Compile::print_compile_messages() {
352 #ifndef PRODUCT
353 // Check if recompiling
354 if (_subsume_loads == false && PrintOpto) {
355 // Recompiling without allowing machine instructions to subsume loads
356 tty->print_cr("*********************************************************");
357 tty->print_cr("** Bailout: Recompile without subsuming loads **");
358 tty->print_cr("*********************************************************");
359 }
360 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
361 // Recompiling without escape analysis
362 tty->print_cr("*********************************************************");
363 tty->print_cr("** Bailout: Recompile without escape analysis **");
364 tty->print_cr("*********************************************************");
365 }
366 if (env()->break_at_compile()) {
367 // Open the debugger when compiling this method.
368 tty->print("### Breaking when compiling: ");
369 method()->print_short_name();
370 tty->cr();
371 BREAKPOINT;
372 }
374 if( PrintOpto ) {
375 if (is_osr_compilation()) {
376 tty->print("[OSR]%3d", _compile_id);
377 } else {
378 tty->print("%3d", _compile_id);
379 }
380 }
381 #endif
382 }
385 void Compile::init_scratch_buffer_blob() {
386 if( scratch_buffer_blob() != NULL ) return;
388 // Construct a temporary CodeBuffer to have it construct a BufferBlob
389 // Cache this BufferBlob for this compile.
390 ResourceMark rm;
391 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size);
392 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size);
393 // Record the buffer blob for next time.
394 set_scratch_buffer_blob(blob);
395 // Have we run out of code space?
396 if (scratch_buffer_blob() == NULL) {
397 // Let CompilerBroker disable further compilations.
398 record_failure("Not enough space for scratch buffer in CodeCache");
399 return;
400 }
402 // Initialize the relocation buffers
403 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size;
404 set_scratch_locs_memory(locs_buf);
405 }
408 //-----------------------scratch_emit_size-------------------------------------
409 // Helper function that computes size by emitting code
410 uint Compile::scratch_emit_size(const Node* n) {
411 // Emit into a trash buffer and count bytes emitted.
412 // This is a pretty expensive way to compute a size,
413 // but it works well enough if seldom used.
414 // All common fixed-size instructions are given a size
415 // method by the AD file.
416 // Note that the scratch buffer blob and locs memory are
417 // allocated at the beginning of the compile task, and
418 // may be shared by several calls to scratch_emit_size.
419 // The allocation of the scratch buffer blob is particularly
420 // expensive, since it has to grab the code cache lock.
421 BufferBlob* blob = this->scratch_buffer_blob();
422 assert(blob != NULL, "Initialize BufferBlob at start");
423 assert(blob->size() > MAX_inst_size, "sanity");
424 relocInfo* locs_buf = scratch_locs_memory();
425 address blob_begin = blob->instructions_begin();
426 address blob_end = (address)locs_buf;
427 assert(blob->instructions_contains(blob_end), "sanity");
428 CodeBuffer buf(blob_begin, blob_end - blob_begin);
429 buf.initialize_consts_size(MAX_const_size);
430 buf.initialize_stubs_size(MAX_stubs_size);
431 assert(locs_buf != NULL, "sanity");
432 int lsize = MAX_locs_size / 2;
433 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize);
434 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize);
435 n->emit(buf, this->regalloc());
436 return buf.code_size();
437 }
440 // ============================================================================
441 //------------------------------Compile standard-------------------------------
442 debug_only( int Compile::_debug_idx = 100000; )
444 // Compile a method. entry_bci is -1 for normal compilations and indicates
445 // the continuation bci for on stack replacement.
448 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
449 : Phase(Compiler),
450 _env(ci_env),
451 _log(ci_env->log()),
452 _compile_id(ci_env->compile_id()),
453 _save_argument_registers(false),
454 _stub_name(NULL),
455 _stub_function(NULL),
456 _stub_entry_point(NULL),
457 _method(target),
458 _entry_bci(osr_bci),
459 _initial_gvn(NULL),
460 _for_igvn(NULL),
461 _warm_calls(NULL),
462 _subsume_loads(subsume_loads),
463 _do_escape_analysis(do_escape_analysis),
464 _failure_reason(NULL),
465 _code_buffer("Compile::Fill_buffer"),
466 _orig_pc_slot(0),
467 _orig_pc_slot_offset_in_bytes(0),
468 _node_bundling_limit(0),
469 _node_bundling_base(NULL),
470 _java_calls(0),
471 _inner_loops(0),
472 #ifndef PRODUCT
473 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
474 _printer(IdealGraphPrinter::printer()),
475 #endif
476 _congraph(NULL) {
477 C = this;
479 CompileWrapper cw(this);
480 #ifndef PRODUCT
481 if (TimeCompiler2) {
482 tty->print(" ");
483 target->holder()->name()->print();
484 tty->print(".");
485 target->print_short_name();
486 tty->print(" ");
487 }
488 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
489 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
490 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
491 if (!print_opto_assembly) {
492 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
493 if (print_assembly && !Disassembler::can_decode()) {
494 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
495 print_opto_assembly = true;
496 }
497 }
498 set_print_assembly(print_opto_assembly);
499 set_parsed_irreducible_loop(false);
500 #endif
502 if (ProfileTraps) {
503 // Make sure the method being compiled gets its own MDO,
504 // so we can at least track the decompile_count().
505 method()->build_method_data();
506 }
508 Init(::AliasLevel);
511 print_compile_messages();
513 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
514 _ilt = InlineTree::build_inline_tree_root();
515 else
516 _ilt = NULL;
518 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
519 assert(num_alias_types() >= AliasIdxRaw, "");
521 #define MINIMUM_NODE_HASH 1023
522 // Node list that Iterative GVN will start with
523 Unique_Node_List for_igvn(comp_arena());
524 set_for_igvn(&for_igvn);
526 // GVN that will be run immediately on new nodes
527 uint estimated_size = method()->code_size()*4+64;
528 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
529 PhaseGVN gvn(node_arena(), estimated_size);
530 set_initial_gvn(&gvn);
532 { // Scope for timing the parser
533 TracePhase t3("parse", &_t_parser, true);
535 // Put top into the hash table ASAP.
536 initial_gvn()->transform_no_reclaim(top());
538 // Set up tf(), start(), and find a CallGenerator.
539 CallGenerator* cg;
540 if (is_osr_compilation()) {
541 const TypeTuple *domain = StartOSRNode::osr_domain();
542 const TypeTuple *range = TypeTuple::make_range(method()->signature());
543 init_tf(TypeFunc::make(domain, range));
544 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
545 initial_gvn()->set_type_bottom(s);
546 init_start(s);
547 cg = CallGenerator::for_osr(method(), entry_bci());
548 } else {
549 // Normal case.
550 init_tf(TypeFunc::make(method()));
551 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
552 initial_gvn()->set_type_bottom(s);
553 init_start(s);
554 float past_uses = method()->interpreter_invocation_count();
555 float expected_uses = past_uses;
556 cg = CallGenerator::for_inline(method(), expected_uses);
557 }
558 if (failing()) return;
559 if (cg == NULL) {
560 record_method_not_compilable_all_tiers("cannot parse method");
561 return;
562 }
563 JVMState* jvms = build_start_state(start(), tf());
564 if ((jvms = cg->generate(jvms)) == NULL) {
565 record_method_not_compilable("method parse failed");
566 return;
567 }
568 GraphKit kit(jvms);
570 if (!kit.stopped()) {
571 // Accept return values, and transfer control we know not where.
572 // This is done by a special, unique ReturnNode bound to root.
573 return_values(kit.jvms());
574 }
576 if (kit.has_exceptions()) {
577 // Any exceptions that escape from this call must be rethrown
578 // to whatever caller is dynamically above us on the stack.
579 // This is done by a special, unique RethrowNode bound to root.
580 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
581 }
583 if (!failing() && has_stringbuilder()) {
584 {
585 // remove useless nodes to make the usage analysis simpler
586 ResourceMark rm;
587 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
588 }
590 {
591 ResourceMark rm;
592 print_method("Before StringOpts", 3);
593 PhaseStringOpts pso(initial_gvn(), &for_igvn);
594 print_method("After StringOpts", 3);
595 }
597 // now inline anything that we skipped the first time around
598 while (_late_inlines.length() > 0) {
599 CallGenerator* cg = _late_inlines.pop();
600 cg->do_late_inline();
601 }
602 }
603 assert(_late_inlines.length() == 0, "should have been processed");
605 print_method("Before RemoveUseless", 3);
607 // Remove clutter produced by parsing.
608 if (!failing()) {
609 ResourceMark rm;
610 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
611 }
612 }
614 // Note: Large methods are capped off in do_one_bytecode().
615 if (failing()) return;
617 // After parsing, node notes are no longer automagic.
618 // They must be propagated by register_new_node_with_optimizer(),
619 // clone(), or the like.
620 set_default_node_notes(NULL);
622 for (;;) {
623 int successes = Inline_Warm();
624 if (failing()) return;
625 if (successes == 0) break;
626 }
628 // Drain the list.
629 Finish_Warm();
630 #ifndef PRODUCT
631 if (_printer) {
632 _printer->print_inlining(this);
633 }
634 #endif
636 if (failing()) return;
637 NOT_PRODUCT( verify_graph_edges(); )
639 // Perform escape analysis
640 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
641 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
642 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction.
643 PhaseGVN* igvn = initial_gvn();
644 Node* oop_null = igvn->zerocon(T_OBJECT);
645 Node* noop_null = igvn->zerocon(T_NARROWOOP);
647 _congraph = new(comp_arena()) ConnectionGraph(this);
648 bool has_non_escaping_obj = _congraph->compute_escape();
650 #ifndef PRODUCT
651 if (PrintEscapeAnalysis) {
652 _congraph->dump();
653 }
654 #endif
655 // Cleanup.
656 if (oop_null->outcnt() == 0)
657 igvn->hash_delete(oop_null);
658 if (noop_null->outcnt() == 0)
659 igvn->hash_delete(noop_null);
661 if (!has_non_escaping_obj) {
662 _congraph = NULL;
663 }
665 if (failing()) return;
666 }
667 // Now optimize
668 Optimize();
669 if (failing()) return;
670 NOT_PRODUCT( verify_graph_edges(); )
672 #ifndef PRODUCT
673 if (PrintIdeal) {
674 ttyLocker ttyl; // keep the following output all in one block
675 // This output goes directly to the tty, not the compiler log.
676 // To enable tools to match it up with the compilation activity,
677 // be sure to tag this tty output with the compile ID.
678 if (xtty != NULL) {
679 xtty->head("ideal compile_id='%d'%s", compile_id(),
680 is_osr_compilation() ? " compile_kind='osr'" :
681 "");
682 }
683 root()->dump(9999);
684 if (xtty != NULL) {
685 xtty->tail("ideal");
686 }
687 }
688 #endif
690 // Now that we know the size of all the monitors we can add a fixed slot
691 // for the original deopt pc.
693 _orig_pc_slot = fixed_slots();
694 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
695 set_fixed_slots(next_slot);
697 // Now generate code
698 Code_Gen();
699 if (failing()) return;
701 // Check if we want to skip execution of all compiled code.
702 {
703 #ifndef PRODUCT
704 if (OptoNoExecute) {
705 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
706 return;
707 }
708 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
709 #endif
711 if (is_osr_compilation()) {
712 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
713 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
714 } else {
715 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
716 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
717 }
719 env()->register_method(_method, _entry_bci,
720 &_code_offsets,
721 _orig_pc_slot_offset_in_bytes,
722 code_buffer(),
723 frame_size_in_words(), _oop_map_set,
724 &_handler_table, &_inc_table,
725 compiler,
726 env()->comp_level(),
727 true, /*has_debug_info*/
728 has_unsafe_access()
729 );
730 }
731 }
733 //------------------------------Compile----------------------------------------
734 // Compile a runtime stub
735 Compile::Compile( ciEnv* ci_env,
736 TypeFunc_generator generator,
737 address stub_function,
738 const char *stub_name,
739 int is_fancy_jump,
740 bool pass_tls,
741 bool save_arg_registers,
742 bool return_pc )
743 : Phase(Compiler),
744 _env(ci_env),
745 _log(ci_env->log()),
746 _compile_id(-1),
747 _save_argument_registers(save_arg_registers),
748 _method(NULL),
749 _stub_name(stub_name),
750 _stub_function(stub_function),
751 _stub_entry_point(NULL),
752 _entry_bci(InvocationEntryBci),
753 _initial_gvn(NULL),
754 _for_igvn(NULL),
755 _warm_calls(NULL),
756 _orig_pc_slot(0),
757 _orig_pc_slot_offset_in_bytes(0),
758 _subsume_loads(true),
759 _do_escape_analysis(false),
760 _failure_reason(NULL),
761 _code_buffer("Compile::Fill_buffer"),
762 _node_bundling_limit(0),
763 _node_bundling_base(NULL),
764 _java_calls(0),
765 _inner_loops(0),
766 #ifndef PRODUCT
767 _trace_opto_output(TraceOptoOutput),
768 _printer(NULL),
769 #endif
770 _congraph(NULL) {
771 C = this;
773 #ifndef PRODUCT
774 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
775 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
776 set_print_assembly(PrintFrameConverterAssembly);
777 set_parsed_irreducible_loop(false);
778 #endif
779 CompileWrapper cw(this);
780 Init(/*AliasLevel=*/ 0);
781 init_tf((*generator)());
783 {
784 // The following is a dummy for the sake of GraphKit::gen_stub
785 Unique_Node_List for_igvn(comp_arena());
786 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
787 PhaseGVN gvn(Thread::current()->resource_area(),255);
788 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
789 gvn.transform_no_reclaim(top());
791 GraphKit kit;
792 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
793 }
795 NOT_PRODUCT( verify_graph_edges(); )
796 Code_Gen();
797 if (failing()) return;
800 // Entry point will be accessed using compile->stub_entry_point();
801 if (code_buffer() == NULL) {
802 Matcher::soft_match_failure();
803 } else {
804 if (PrintAssembly && (WizardMode || Verbose))
805 tty->print_cr("### Stub::%s", stub_name);
807 if (!failing()) {
808 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
810 // Make the NMethod
811 // For now we mark the frame as never safe for profile stackwalking
812 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
813 code_buffer(),
814 CodeOffsets::frame_never_safe,
815 // _code_offsets.value(CodeOffsets::Frame_Complete),
816 frame_size_in_words(),
817 _oop_map_set,
818 save_arg_registers);
819 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
821 _stub_entry_point = rs->entry_point();
822 }
823 }
824 }
826 #ifndef PRODUCT
827 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
828 if(PrintOpto && Verbose) {
829 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
830 }
831 }
832 #endif
834 void Compile::print_codes() {
835 }
837 //------------------------------Init-------------------------------------------
838 // Prepare for a single compilation
839 void Compile::Init(int aliaslevel) {
840 _unique = 0;
841 _regalloc = NULL;
843 _tf = NULL; // filled in later
844 _top = NULL; // cached later
845 _matcher = NULL; // filled in later
846 _cfg = NULL; // filled in later
848 set_24_bit_selection_and_mode(Use24BitFP, false);
850 _node_note_array = NULL;
851 _default_node_notes = NULL;
853 _immutable_memory = NULL; // filled in at first inquiry
855 // Globally visible Nodes
856 // First set TOP to NULL to give safe behavior during creation of RootNode
857 set_cached_top_node(NULL);
858 set_root(new (this, 3) RootNode());
859 // Now that you have a Root to point to, create the real TOP
860 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
861 set_recent_alloc(NULL, NULL);
863 // Create Debug Information Recorder to record scopes, oopmaps, etc.
864 env()->set_oop_recorder(new OopRecorder(comp_arena()));
865 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
866 env()->set_dependencies(new Dependencies(env()));
868 _fixed_slots = 0;
869 set_has_split_ifs(false);
870 set_has_loops(has_method() && method()->has_loops()); // first approximation
871 set_has_stringbuilder(false);
872 _deopt_happens = true; // start out assuming the worst
873 _trap_can_recompile = false; // no traps emitted yet
874 _major_progress = true; // start out assuming good things will happen
875 set_has_unsafe_access(false);
876 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
877 set_decompile_count(0);
879 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
880 // Compilation level related initialization
881 if (env()->comp_level() == CompLevel_fast_compile) {
882 set_num_loop_opts(Tier1LoopOptsCount);
883 set_do_inlining(Tier1Inline != 0);
884 set_max_inline_size(Tier1MaxInlineSize);
885 set_freq_inline_size(Tier1FreqInlineSize);
886 set_do_scheduling(false);
887 set_do_count_invocations(Tier1CountInvocations);
888 set_do_method_data_update(Tier1UpdateMethodData);
889 } else {
890 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
891 set_num_loop_opts(LoopOptsCount);
892 set_do_inlining(Inline);
893 set_max_inline_size(MaxInlineSize);
894 set_freq_inline_size(FreqInlineSize);
895 set_do_scheduling(OptoScheduling);
896 set_do_count_invocations(false);
897 set_do_method_data_update(false);
898 }
900 if (debug_info()->recording_non_safepoints()) {
901 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
902 (comp_arena(), 8, 0, NULL));
903 set_default_node_notes(Node_Notes::make(this));
904 }
906 // // -- Initialize types before each compile --
907 // // Update cached type information
908 // if( _method && _method->constants() )
909 // Type::update_loaded_types(_method, _method->constants());
911 // Init alias_type map.
912 if (!_do_escape_analysis && aliaslevel == 3)
913 aliaslevel = 2; // No unique types without escape analysis
914 _AliasLevel = aliaslevel;
915 const int grow_ats = 16;
916 _max_alias_types = grow_ats;
917 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
918 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
919 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
920 {
921 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
922 }
923 // Initialize the first few types.
924 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
925 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
926 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
927 _num_alias_types = AliasIdxRaw+1;
928 // Zero out the alias type cache.
929 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
930 // A NULL adr_type hits in the cache right away. Preload the right answer.
931 probe_alias_cache(NULL)->_index = AliasIdxTop;
933 _intrinsics = NULL;
934 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
935 register_library_intrinsics();
936 }
938 //---------------------------init_start----------------------------------------
939 // Install the StartNode on this compile object.
940 void Compile::init_start(StartNode* s) {
941 if (failing())
942 return; // already failing
943 assert(s == start(), "");
944 }
946 StartNode* Compile::start() const {
947 assert(!failing(), "");
948 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
949 Node* start = root()->fast_out(i);
950 if( start->is_Start() )
951 return start->as_Start();
952 }
953 ShouldNotReachHere();
954 return NULL;
955 }
957 //-------------------------------immutable_memory-------------------------------------
958 // Access immutable memory
959 Node* Compile::immutable_memory() {
960 if (_immutable_memory != NULL) {
961 return _immutable_memory;
962 }
963 StartNode* s = start();
964 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
965 Node *p = s->fast_out(i);
966 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
967 _immutable_memory = p;
968 return _immutable_memory;
969 }
970 }
971 ShouldNotReachHere();
972 return NULL;
973 }
975 //----------------------set_cached_top_node------------------------------------
976 // Install the cached top node, and make sure Node::is_top works correctly.
977 void Compile::set_cached_top_node(Node* tn) {
978 if (tn != NULL) verify_top(tn);
979 Node* old_top = _top;
980 _top = tn;
981 // Calling Node::setup_is_top allows the nodes the chance to adjust
982 // their _out arrays.
983 if (_top != NULL) _top->setup_is_top();
984 if (old_top != NULL) old_top->setup_is_top();
985 assert(_top == NULL || top()->is_top(), "");
986 }
988 #ifndef PRODUCT
989 void Compile::verify_top(Node* tn) const {
990 if (tn != NULL) {
991 assert(tn->is_Con(), "top node must be a constant");
992 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
993 assert(tn->in(0) != NULL, "must have live top node");
994 }
995 }
996 #endif
999 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1001 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1002 guarantee(arr != NULL, "");
1003 int num_blocks = arr->length();
1004 if (grow_by < num_blocks) grow_by = num_blocks;
1005 int num_notes = grow_by * _node_notes_block_size;
1006 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1007 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1008 while (num_notes > 0) {
1009 arr->append(notes);
1010 notes += _node_notes_block_size;
1011 num_notes -= _node_notes_block_size;
1012 }
1013 assert(num_notes == 0, "exact multiple, please");
1014 }
1016 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1017 if (source == NULL || dest == NULL) return false;
1019 if (dest->is_Con())
1020 return false; // Do not push debug info onto constants.
1022 #ifdef ASSERT
1023 // Leave a bread crumb trail pointing to the original node:
1024 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1025 dest->set_debug_orig(source);
1026 }
1027 #endif
1029 if (node_note_array() == NULL)
1030 return false; // Not collecting any notes now.
1032 // This is a copy onto a pre-existing node, which may already have notes.
1033 // If both nodes have notes, do not overwrite any pre-existing notes.
1034 Node_Notes* source_notes = node_notes_at(source->_idx);
1035 if (source_notes == NULL || source_notes->is_clear()) return false;
1036 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1037 if (dest_notes == NULL || dest_notes->is_clear()) {
1038 return set_node_notes_at(dest->_idx, source_notes);
1039 }
1041 Node_Notes merged_notes = (*source_notes);
1042 // The order of operations here ensures that dest notes will win...
1043 merged_notes.update_from(dest_notes);
1044 return set_node_notes_at(dest->_idx, &merged_notes);
1045 }
1048 //--------------------------allow_range_check_smearing-------------------------
1049 // Gating condition for coalescing similar range checks.
1050 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1051 // single covering check that is at least as strong as any of them.
1052 // If the optimization succeeds, the simplified (strengthened) range check
1053 // will always succeed. If it fails, we will deopt, and then give up
1054 // on the optimization.
1055 bool Compile::allow_range_check_smearing() const {
1056 // If this method has already thrown a range-check,
1057 // assume it was because we already tried range smearing
1058 // and it failed.
1059 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1060 return !already_trapped;
1061 }
1064 //------------------------------flatten_alias_type-----------------------------
1065 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1066 int offset = tj->offset();
1067 TypePtr::PTR ptr = tj->ptr();
1069 // Known instance (scalarizable allocation) alias only with itself.
1070 bool is_known_inst = tj->isa_oopptr() != NULL &&
1071 tj->is_oopptr()->is_known_instance();
1073 // Process weird unsafe references.
1074 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1075 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1076 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1077 tj = TypeOopPtr::BOTTOM;
1078 ptr = tj->ptr();
1079 offset = tj->offset();
1080 }
1082 // Array pointers need some flattening
1083 const TypeAryPtr *ta = tj->isa_aryptr();
1084 if( ta && is_known_inst ) {
1085 if ( offset != Type::OffsetBot &&
1086 offset > arrayOopDesc::length_offset_in_bytes() ) {
1087 offset = Type::OffsetBot; // Flatten constant access into array body only
1088 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1089 }
1090 } else if( ta && _AliasLevel >= 2 ) {
1091 // For arrays indexed by constant indices, we flatten the alias
1092 // space to include all of the array body. Only the header, klass
1093 // and array length can be accessed un-aliased.
1094 if( offset != Type::OffsetBot ) {
1095 if( ta->const_oop() ) { // methodDataOop or methodOop
1096 offset = Type::OffsetBot; // Flatten constant access into array body
1097 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1098 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1099 // range is OK as-is.
1100 tj = ta = TypeAryPtr::RANGE;
1101 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1102 tj = TypeInstPtr::KLASS; // all klass loads look alike
1103 ta = TypeAryPtr::RANGE; // generic ignored junk
1104 ptr = TypePtr::BotPTR;
1105 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1106 tj = TypeInstPtr::MARK;
1107 ta = TypeAryPtr::RANGE; // generic ignored junk
1108 ptr = TypePtr::BotPTR;
1109 } else { // Random constant offset into array body
1110 offset = Type::OffsetBot; // Flatten constant access into array body
1111 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1112 }
1113 }
1114 // Arrays of fixed size alias with arrays of unknown size.
1115 if (ta->size() != TypeInt::POS) {
1116 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1117 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1118 }
1119 // Arrays of known objects become arrays of unknown objects.
1120 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1121 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1122 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1123 }
1124 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1125 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1126 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1127 }
1128 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1129 // cannot be distinguished by bytecode alone.
1130 if (ta->elem() == TypeInt::BOOL) {
1131 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1132 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1133 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1134 }
1135 // During the 2nd round of IterGVN, NotNull castings are removed.
1136 // Make sure the Bottom and NotNull variants alias the same.
1137 // Also, make sure exact and non-exact variants alias the same.
1138 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1139 if (ta->const_oop()) {
1140 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1141 } else {
1142 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1143 }
1144 }
1145 }
1147 // Oop pointers need some flattening
1148 const TypeInstPtr *to = tj->isa_instptr();
1149 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1150 if( ptr == TypePtr::Constant ) {
1151 // No constant oop pointers (such as Strings); they alias with
1152 // unknown strings.
1153 assert(!is_known_inst, "not scalarizable allocation");
1154 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1155 } else if( is_known_inst ) {
1156 tj = to; // Keep NotNull and klass_is_exact for instance type
1157 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1158 // During the 2nd round of IterGVN, NotNull castings are removed.
1159 // Make sure the Bottom and NotNull variants alias the same.
1160 // Also, make sure exact and non-exact variants alias the same.
1161 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1162 }
1163 // Canonicalize the holder of this field
1164 ciInstanceKlass *k = to->klass()->as_instance_klass();
1165 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1166 // First handle header references such as a LoadKlassNode, even if the
1167 // object's klass is unloaded at compile time (4965979).
1168 if (!is_known_inst) { // Do it only for non-instance types
1169 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1170 }
1171 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1172 to = NULL;
1173 tj = TypeOopPtr::BOTTOM;
1174 offset = tj->offset();
1175 } else {
1176 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1177 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1178 if( is_known_inst ) {
1179 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1180 } else {
1181 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1182 }
1183 }
1184 }
1185 }
1187 // Klass pointers to object array klasses need some flattening
1188 const TypeKlassPtr *tk = tj->isa_klassptr();
1189 if( tk ) {
1190 // If we are referencing a field within a Klass, we need
1191 // to assume the worst case of an Object. Both exact and
1192 // inexact types must flatten to the same alias class.
1193 // Since the flattened result for a klass is defined to be
1194 // precisely java.lang.Object, use a constant ptr.
1195 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1197 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1198 TypeKlassPtr::OBJECT->klass(),
1199 offset);
1200 }
1202 ciKlass* klass = tk->klass();
1203 if( klass->is_obj_array_klass() ) {
1204 ciKlass* k = TypeAryPtr::OOPS->klass();
1205 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1206 k = TypeInstPtr::BOTTOM->klass();
1207 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1208 }
1210 // Check for precise loads from the primary supertype array and force them
1211 // to the supertype cache alias index. Check for generic array loads from
1212 // the primary supertype array and also force them to the supertype cache
1213 // alias index. Since the same load can reach both, we need to merge
1214 // these 2 disparate memories into the same alias class. Since the
1215 // primary supertype array is read-only, there's no chance of confusion
1216 // where we bypass an array load and an array store.
1217 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1218 if( offset == Type::OffsetBot ||
1219 off2 < Klass::primary_super_limit()*wordSize ) {
1220 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1221 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1222 }
1223 }
1225 // Flatten all Raw pointers together.
1226 if (tj->base() == Type::RawPtr)
1227 tj = TypeRawPtr::BOTTOM;
1229 if (tj->base() == Type::AnyPtr)
1230 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1232 // Flatten all to bottom for now
1233 switch( _AliasLevel ) {
1234 case 0:
1235 tj = TypePtr::BOTTOM;
1236 break;
1237 case 1: // Flatten to: oop, static, field or array
1238 switch (tj->base()) {
1239 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1240 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1241 case Type::AryPtr: // do not distinguish arrays at all
1242 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1243 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1244 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1245 default: ShouldNotReachHere();
1246 }
1247 break;
1248 case 2: // No collapsing at level 2; keep all splits
1249 case 3: // No collapsing at level 3; keep all splits
1250 break;
1251 default:
1252 Unimplemented();
1253 }
1255 offset = tj->offset();
1256 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1258 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1259 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1260 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1261 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1262 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1263 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1264 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1265 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1266 assert( tj->ptr() != TypePtr::TopPTR &&
1267 tj->ptr() != TypePtr::AnyNull &&
1268 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1269 // assert( tj->ptr() != TypePtr::Constant ||
1270 // tj->base() == Type::RawPtr ||
1271 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1273 return tj;
1274 }
1276 void Compile::AliasType::Init(int i, const TypePtr* at) {
1277 _index = i;
1278 _adr_type = at;
1279 _field = NULL;
1280 _is_rewritable = true; // default
1281 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1282 if (atoop != NULL && atoop->is_known_instance()) {
1283 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1284 _general_index = Compile::current()->get_alias_index(gt);
1285 } else {
1286 _general_index = 0;
1287 }
1288 }
1290 //---------------------------------print_on------------------------------------
1291 #ifndef PRODUCT
1292 void Compile::AliasType::print_on(outputStream* st) {
1293 if (index() < 10)
1294 st->print("@ <%d> ", index());
1295 else st->print("@ <%d>", index());
1296 st->print(is_rewritable() ? " " : " RO");
1297 int offset = adr_type()->offset();
1298 if (offset == Type::OffsetBot)
1299 st->print(" +any");
1300 else st->print(" +%-3d", offset);
1301 st->print(" in ");
1302 adr_type()->dump_on(st);
1303 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1304 if (field() != NULL && tjp) {
1305 if (tjp->klass() != field()->holder() ||
1306 tjp->offset() != field()->offset_in_bytes()) {
1307 st->print(" != ");
1308 field()->print();
1309 st->print(" ***");
1310 }
1311 }
1312 }
1314 void print_alias_types() {
1315 Compile* C = Compile::current();
1316 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1317 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1318 C->alias_type(idx)->print_on(tty);
1319 tty->cr();
1320 }
1321 }
1322 #endif
1325 //----------------------------probe_alias_cache--------------------------------
1326 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1327 intptr_t key = (intptr_t) adr_type;
1328 key ^= key >> logAliasCacheSize;
1329 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1330 }
1333 //-----------------------------grow_alias_types--------------------------------
1334 void Compile::grow_alias_types() {
1335 const int old_ats = _max_alias_types; // how many before?
1336 const int new_ats = old_ats; // how many more?
1337 const int grow_ats = old_ats+new_ats; // how many now?
1338 _max_alias_types = grow_ats;
1339 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1340 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1341 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1342 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1343 }
1346 //--------------------------------find_alias_type------------------------------
1347 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1348 if (_AliasLevel == 0)
1349 return alias_type(AliasIdxBot);
1351 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1352 if (ace->_adr_type == adr_type) {
1353 return alias_type(ace->_index);
1354 }
1356 // Handle special cases.
1357 if (adr_type == NULL) return alias_type(AliasIdxTop);
1358 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1360 // Do it the slow way.
1361 const TypePtr* flat = flatten_alias_type(adr_type);
1363 #ifdef ASSERT
1364 assert(flat == flatten_alias_type(flat), "idempotent");
1365 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1366 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1367 const TypeOopPtr* foop = flat->is_oopptr();
1368 // Scalarizable allocations have exact klass always.
1369 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1370 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1371 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1372 }
1373 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1374 #endif
1376 int idx = AliasIdxTop;
1377 for (int i = 0; i < num_alias_types(); i++) {
1378 if (alias_type(i)->adr_type() == flat) {
1379 idx = i;
1380 break;
1381 }
1382 }
1384 if (idx == AliasIdxTop) {
1385 if (no_create) return NULL;
1386 // Grow the array if necessary.
1387 if (_num_alias_types == _max_alias_types) grow_alias_types();
1388 // Add a new alias type.
1389 idx = _num_alias_types++;
1390 _alias_types[idx]->Init(idx, flat);
1391 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1392 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1393 if (flat->isa_instptr()) {
1394 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1395 && flat->is_instptr()->klass() == env()->Class_klass())
1396 alias_type(idx)->set_rewritable(false);
1397 }
1398 if (flat->isa_klassptr()) {
1399 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1400 alias_type(idx)->set_rewritable(false);
1401 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1402 alias_type(idx)->set_rewritable(false);
1403 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1404 alias_type(idx)->set_rewritable(false);
1405 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1406 alias_type(idx)->set_rewritable(false);
1407 }
1408 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1409 // but the base pointer type is not distinctive enough to identify
1410 // references into JavaThread.)
1412 // Check for final instance fields.
1413 const TypeInstPtr* tinst = flat->isa_instptr();
1414 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1415 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1416 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1417 // Set field() and is_rewritable() attributes.
1418 if (field != NULL) alias_type(idx)->set_field(field);
1419 }
1420 const TypeKlassPtr* tklass = flat->isa_klassptr();
1421 // Check for final static fields.
1422 if (tklass && tklass->klass()->is_instance_klass()) {
1423 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1424 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1425 // Set field() and is_rewritable() attributes.
1426 if (field != NULL) alias_type(idx)->set_field(field);
1427 }
1428 }
1430 // Fill the cache for next time.
1431 ace->_adr_type = adr_type;
1432 ace->_index = idx;
1433 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1435 // Might as well try to fill the cache for the flattened version, too.
1436 AliasCacheEntry* face = probe_alias_cache(flat);
1437 if (face->_adr_type == NULL) {
1438 face->_adr_type = flat;
1439 face->_index = idx;
1440 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1441 }
1443 return alias_type(idx);
1444 }
1447 Compile::AliasType* Compile::alias_type(ciField* field) {
1448 const TypeOopPtr* t;
1449 if (field->is_static())
1450 t = TypeKlassPtr::make(field->holder());
1451 else
1452 t = TypeOopPtr::make_from_klass_raw(field->holder());
1453 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1454 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1455 return atp;
1456 }
1459 //------------------------------have_alias_type--------------------------------
1460 bool Compile::have_alias_type(const TypePtr* adr_type) {
1461 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1462 if (ace->_adr_type == adr_type) {
1463 return true;
1464 }
1466 // Handle special cases.
1467 if (adr_type == NULL) return true;
1468 if (adr_type == TypePtr::BOTTOM) return true;
1470 return find_alias_type(adr_type, true) != NULL;
1471 }
1473 //-----------------------------must_alias--------------------------------------
1474 // True if all values of the given address type are in the given alias category.
1475 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1476 if (alias_idx == AliasIdxBot) return true; // the universal category
1477 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1478 if (alias_idx == AliasIdxTop) return false; // the empty category
1479 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1481 // the only remaining possible overlap is identity
1482 int adr_idx = get_alias_index(adr_type);
1483 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1484 assert(adr_idx == alias_idx ||
1485 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1486 && adr_type != TypeOopPtr::BOTTOM),
1487 "should not be testing for overlap with an unsafe pointer");
1488 return adr_idx == alias_idx;
1489 }
1491 //------------------------------can_alias--------------------------------------
1492 // True if any values of the given address type are in the given alias category.
1493 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1494 if (alias_idx == AliasIdxTop) return false; // the empty category
1495 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1496 if (alias_idx == AliasIdxBot) return true; // the universal category
1497 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1499 // the only remaining possible overlap is identity
1500 int adr_idx = get_alias_index(adr_type);
1501 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1502 return adr_idx == alias_idx;
1503 }
1507 //---------------------------pop_warm_call-------------------------------------
1508 WarmCallInfo* Compile::pop_warm_call() {
1509 WarmCallInfo* wci = _warm_calls;
1510 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1511 return wci;
1512 }
1514 //----------------------------Inline_Warm--------------------------------------
1515 int Compile::Inline_Warm() {
1516 // If there is room, try to inline some more warm call sites.
1517 // %%% Do a graph index compaction pass when we think we're out of space?
1518 if (!InlineWarmCalls) return 0;
1520 int calls_made_hot = 0;
1521 int room_to_grow = NodeCountInliningCutoff - unique();
1522 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1523 int amount_grown = 0;
1524 WarmCallInfo* call;
1525 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1526 int est_size = (int)call->size();
1527 if (est_size > (room_to_grow - amount_grown)) {
1528 // This one won't fit anyway. Get rid of it.
1529 call->make_cold();
1530 continue;
1531 }
1532 call->make_hot();
1533 calls_made_hot++;
1534 amount_grown += est_size;
1535 amount_to_grow -= est_size;
1536 }
1538 if (calls_made_hot > 0) set_major_progress();
1539 return calls_made_hot;
1540 }
1543 //----------------------------Finish_Warm--------------------------------------
1544 void Compile::Finish_Warm() {
1545 if (!InlineWarmCalls) return;
1546 if (failing()) return;
1547 if (warm_calls() == NULL) return;
1549 // Clean up loose ends, if we are out of space for inlining.
1550 WarmCallInfo* call;
1551 while ((call = pop_warm_call()) != NULL) {
1552 call->make_cold();
1553 }
1554 }
1557 //------------------------------Optimize---------------------------------------
1558 // Given a graph, optimize it.
1559 void Compile::Optimize() {
1560 TracePhase t1("optimizer", &_t_optimizer, true);
1562 #ifndef PRODUCT
1563 if (env()->break_at_compile()) {
1564 BREAKPOINT;
1565 }
1567 #endif
1569 ResourceMark rm;
1570 int loop_opts_cnt;
1572 NOT_PRODUCT( verify_graph_edges(); )
1574 print_method("After Parsing");
1576 {
1577 // Iterative Global Value Numbering, including ideal transforms
1578 // Initialize IterGVN with types and values from parse-time GVN
1579 PhaseIterGVN igvn(initial_gvn());
1580 {
1581 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1582 igvn.optimize();
1583 }
1585 print_method("Iter GVN 1", 2);
1587 if (failing()) return;
1589 // Loop transforms on the ideal graph. Range Check Elimination,
1590 // peeling, unrolling, etc.
1592 // Set loop opts counter
1593 loop_opts_cnt = num_loop_opts();
1594 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1595 {
1596 TracePhase t2("idealLoop", &_t_idealLoop, true);
1597 PhaseIdealLoop ideal_loop( igvn, true );
1598 loop_opts_cnt--;
1599 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1600 if (failing()) return;
1601 }
1602 // Loop opts pass if partial peeling occurred in previous pass
1603 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1604 TracePhase t3("idealLoop", &_t_idealLoop, true);
1605 PhaseIdealLoop ideal_loop( igvn, false );
1606 loop_opts_cnt--;
1607 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1608 if (failing()) return;
1609 }
1610 // Loop opts pass for loop-unrolling before CCP
1611 if(major_progress() && (loop_opts_cnt > 0)) {
1612 TracePhase t4("idealLoop", &_t_idealLoop, true);
1613 PhaseIdealLoop ideal_loop( igvn, false );
1614 loop_opts_cnt--;
1615 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1616 }
1617 if (!failing()) {
1618 // Verify that last round of loop opts produced a valid graph
1619 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1620 PhaseIdealLoop::verify(igvn);
1621 }
1622 }
1623 if (failing()) return;
1625 // Conditional Constant Propagation;
1626 PhaseCCP ccp( &igvn );
1627 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1628 {
1629 TracePhase t2("ccp", &_t_ccp, true);
1630 ccp.do_transform();
1631 }
1632 print_method("PhaseCPP 1", 2);
1634 assert( true, "Break here to ccp.dump_old2new_map()");
1636 // Iterative Global Value Numbering, including ideal transforms
1637 {
1638 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1639 igvn = ccp;
1640 igvn.optimize();
1641 }
1643 print_method("Iter GVN 2", 2);
1645 if (failing()) return;
1647 // Loop transforms on the ideal graph. Range Check Elimination,
1648 // peeling, unrolling, etc.
1649 if(loop_opts_cnt > 0) {
1650 debug_only( int cnt = 0; );
1651 while(major_progress() && (loop_opts_cnt > 0)) {
1652 TracePhase t2("idealLoop", &_t_idealLoop, true);
1653 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1654 PhaseIdealLoop ideal_loop( igvn, true );
1655 loop_opts_cnt--;
1656 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1657 if (failing()) return;
1658 }
1659 }
1661 {
1662 // Verify that all previous optimizations produced a valid graph
1663 // at least to this point, even if no loop optimizations were done.
1664 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1665 PhaseIdealLoop::verify(igvn);
1666 }
1668 {
1669 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1670 PhaseMacroExpand mex(igvn);
1671 if (mex.expand_macro_nodes()) {
1672 assert(failing(), "must bail out w/ explicit message");
1673 return;
1674 }
1675 }
1677 } // (End scope of igvn; run destructor if necessary for asserts.)
1679 // A method with only infinite loops has no edges entering loops from root
1680 {
1681 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1682 if (final_graph_reshaping()) {
1683 assert(failing(), "must bail out w/ explicit message");
1684 return;
1685 }
1686 }
1688 print_method("Optimize finished", 2);
1689 }
1692 //------------------------------Code_Gen---------------------------------------
1693 // Given a graph, generate code for it
1694 void Compile::Code_Gen() {
1695 if (failing()) return;
1697 // Perform instruction selection. You might think we could reclaim Matcher
1698 // memory PDQ, but actually the Matcher is used in generating spill code.
1699 // Internals of the Matcher (including some VectorSets) must remain live
1700 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1701 // set a bit in reclaimed memory.
1703 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1704 // nodes. Mapping is only valid at the root of each matched subtree.
1705 NOT_PRODUCT( verify_graph_edges(); )
1707 Node_List proj_list;
1708 Matcher m(proj_list);
1709 _matcher = &m;
1710 {
1711 TracePhase t2("matcher", &_t_matcher, true);
1712 m.match();
1713 }
1714 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1715 // nodes. Mapping is only valid at the root of each matched subtree.
1716 NOT_PRODUCT( verify_graph_edges(); )
1718 // If you have too many nodes, or if matching has failed, bail out
1719 check_node_count(0, "out of nodes matching instructions");
1720 if (failing()) return;
1722 // Build a proper-looking CFG
1723 PhaseCFG cfg(node_arena(), root(), m);
1724 _cfg = &cfg;
1725 {
1726 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1727 cfg.Dominators();
1728 if (failing()) return;
1730 NOT_PRODUCT( verify_graph_edges(); )
1732 cfg.Estimate_Block_Frequency();
1733 cfg.GlobalCodeMotion(m,unique(),proj_list);
1735 print_method("Global code motion", 2);
1737 if (failing()) return;
1738 NOT_PRODUCT( verify_graph_edges(); )
1740 debug_only( cfg.verify(); )
1741 }
1742 NOT_PRODUCT( verify_graph_edges(); )
1744 PhaseChaitin regalloc(unique(),cfg,m);
1745 _regalloc = ®alloc;
1746 {
1747 TracePhase t2("regalloc", &_t_registerAllocation, true);
1748 // Perform any platform dependent preallocation actions. This is used,
1749 // for example, to avoid taking an implicit null pointer exception
1750 // using the frame pointer on win95.
1751 _regalloc->pd_preallocate_hook();
1753 // Perform register allocation. After Chaitin, use-def chains are
1754 // no longer accurate (at spill code) and so must be ignored.
1755 // Node->LRG->reg mappings are still accurate.
1756 _regalloc->Register_Allocate();
1758 // Bail out if the allocator builds too many nodes
1759 if (failing()) return;
1760 }
1762 // Prior to register allocation we kept empty basic blocks in case the
1763 // the allocator needed a place to spill. After register allocation we
1764 // are not adding any new instructions. If any basic block is empty, we
1765 // can now safely remove it.
1766 {
1767 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1768 cfg.remove_empty();
1769 if (do_freq_based_layout()) {
1770 PhaseBlockLayout layout(cfg);
1771 } else {
1772 cfg.set_loop_alignment();
1773 }
1774 cfg.fixup_flow();
1775 }
1777 // Perform any platform dependent postallocation verifications.
1778 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1780 // Apply peephole optimizations
1781 if( OptoPeephole ) {
1782 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1783 PhasePeephole peep( _regalloc, cfg);
1784 peep.do_transform();
1785 }
1787 // Convert Nodes to instruction bits in a buffer
1788 {
1789 // %%%% workspace merge brought two timers together for one job
1790 TracePhase t2a("output", &_t_output, true);
1791 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1792 Output();
1793 }
1795 print_method("Final Code");
1797 // He's dead, Jim.
1798 _cfg = (PhaseCFG*)0xdeadbeef;
1799 _regalloc = (PhaseChaitin*)0xdeadbeef;
1800 }
1803 //------------------------------dump_asm---------------------------------------
1804 // Dump formatted assembly
1805 #ifndef PRODUCT
1806 void Compile::dump_asm(int *pcs, uint pc_limit) {
1807 bool cut_short = false;
1808 tty->print_cr("#");
1809 tty->print("# "); _tf->dump(); tty->cr();
1810 tty->print_cr("#");
1812 // For all blocks
1813 int pc = 0x0; // Program counter
1814 char starts_bundle = ' ';
1815 _regalloc->dump_frame();
1817 Node *n = NULL;
1818 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1819 if (VMThread::should_terminate()) { cut_short = true; break; }
1820 Block *b = _cfg->_blocks[i];
1821 if (b->is_connector() && !Verbose) continue;
1822 n = b->_nodes[0];
1823 if (pcs && n->_idx < pc_limit)
1824 tty->print("%3.3x ", pcs[n->_idx]);
1825 else
1826 tty->print(" ");
1827 b->dump_head( &_cfg->_bbs );
1828 if (b->is_connector()) {
1829 tty->print_cr(" # Empty connector block");
1830 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1831 tty->print_cr(" # Block is sole successor of call");
1832 }
1834 // For all instructions
1835 Node *delay = NULL;
1836 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1837 if (VMThread::should_terminate()) { cut_short = true; break; }
1838 n = b->_nodes[j];
1839 if (valid_bundle_info(n)) {
1840 Bundle *bundle = node_bundling(n);
1841 if (bundle->used_in_unconditional_delay()) {
1842 delay = n;
1843 continue;
1844 }
1845 if (bundle->starts_bundle())
1846 starts_bundle = '+';
1847 }
1849 if (WizardMode) n->dump();
1851 if( !n->is_Region() && // Dont print in the Assembly
1852 !n->is_Phi() && // a few noisely useless nodes
1853 !n->is_Proj() &&
1854 !n->is_MachTemp() &&
1855 !n->is_Catch() && // Would be nice to print exception table targets
1856 !n->is_MergeMem() && // Not very interesting
1857 !n->is_top() && // Debug info table constants
1858 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1859 ) {
1860 if (pcs && n->_idx < pc_limit)
1861 tty->print("%3.3x", pcs[n->_idx]);
1862 else
1863 tty->print(" ");
1864 tty->print(" %c ", starts_bundle);
1865 starts_bundle = ' ';
1866 tty->print("\t");
1867 n->format(_regalloc, tty);
1868 tty->cr();
1869 }
1871 // If we have an instruction with a delay slot, and have seen a delay,
1872 // then back up and print it
1873 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1874 assert(delay != NULL, "no unconditional delay instruction");
1875 if (WizardMode) delay->dump();
1877 if (node_bundling(delay)->starts_bundle())
1878 starts_bundle = '+';
1879 if (pcs && n->_idx < pc_limit)
1880 tty->print("%3.3x", pcs[n->_idx]);
1881 else
1882 tty->print(" ");
1883 tty->print(" %c ", starts_bundle);
1884 starts_bundle = ' ';
1885 tty->print("\t");
1886 delay->format(_regalloc, tty);
1887 tty->print_cr("");
1888 delay = NULL;
1889 }
1891 // Dump the exception table as well
1892 if( n->is_Catch() && (Verbose || WizardMode) ) {
1893 // Print the exception table for this offset
1894 _handler_table.print_subtable_for(pc);
1895 }
1896 }
1898 if (pcs && n->_idx < pc_limit)
1899 tty->print_cr("%3.3x", pcs[n->_idx]);
1900 else
1901 tty->print_cr("");
1903 assert(cut_short || delay == NULL, "no unconditional delay branch");
1905 } // End of per-block dump
1906 tty->print_cr("");
1908 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1909 }
1910 #endif
1912 //------------------------------Final_Reshape_Counts---------------------------
1913 // This class defines counters to help identify when a method
1914 // may/must be executed using hardware with only 24-bit precision.
1915 struct Final_Reshape_Counts : public StackObj {
1916 int _call_count; // count non-inlined 'common' calls
1917 int _float_count; // count float ops requiring 24-bit precision
1918 int _double_count; // count double ops requiring more precision
1919 int _java_call_count; // count non-inlined 'java' calls
1920 int _inner_loop_count; // count loops which need alignment
1921 VectorSet _visited; // Visitation flags
1922 Node_List _tests; // Set of IfNodes & PCTableNodes
1924 Final_Reshape_Counts() :
1925 _call_count(0), _float_count(0), _double_count(0),
1926 _java_call_count(0), _inner_loop_count(0),
1927 _visited( Thread::current()->resource_area() ) { }
1929 void inc_call_count () { _call_count ++; }
1930 void inc_float_count () { _float_count ++; }
1931 void inc_double_count() { _double_count++; }
1932 void inc_java_call_count() { _java_call_count++; }
1933 void inc_inner_loop_count() { _inner_loop_count++; }
1935 int get_call_count () const { return _call_count ; }
1936 int get_float_count () const { return _float_count ; }
1937 int get_double_count() const { return _double_count; }
1938 int get_java_call_count() const { return _java_call_count; }
1939 int get_inner_loop_count() const { return _inner_loop_count; }
1940 };
1942 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1943 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1944 // Make sure the offset goes inside the instance layout.
1945 return k->contains_field_offset(tp->offset());
1946 // Note that OffsetBot and OffsetTop are very negative.
1947 }
1949 //------------------------------final_graph_reshaping_impl----------------------
1950 // Implement items 1-5 from final_graph_reshaping below.
1951 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
1953 if ( n->outcnt() == 0 ) return; // dead node
1954 uint nop = n->Opcode();
1956 // Check for 2-input instruction with "last use" on right input.
1957 // Swap to left input. Implements item (2).
1958 if( n->req() == 3 && // two-input instruction
1959 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1960 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1961 n->in(2)->outcnt() == 1 &&// right use IS a last use
1962 !n->in(2)->is_Con() ) { // right use is not a constant
1963 // Check for commutative opcode
1964 switch( nop ) {
1965 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1966 case Op_MaxI: case Op_MinI:
1967 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1968 case Op_AndL: case Op_XorL: case Op_OrL:
1969 case Op_AndI: case Op_XorI: case Op_OrI: {
1970 // Move "last use" input to left by swapping inputs
1971 n->swap_edges(1, 2);
1972 break;
1973 }
1974 default:
1975 break;
1976 }
1977 }
1979 // Count FPU ops and common calls, implements item (3)
1980 switch( nop ) {
1981 // Count all float operations that may use FPU
1982 case Op_AddF:
1983 case Op_SubF:
1984 case Op_MulF:
1985 case Op_DivF:
1986 case Op_NegF:
1987 case Op_ModF:
1988 case Op_ConvI2F:
1989 case Op_ConF:
1990 case Op_CmpF:
1991 case Op_CmpF3:
1992 // case Op_ConvL2F: // longs are split into 32-bit halves
1993 frc.inc_float_count();
1994 break;
1996 case Op_ConvF2D:
1997 case Op_ConvD2F:
1998 frc.inc_float_count();
1999 frc.inc_double_count();
2000 break;
2002 // Count all double operations that may use FPU
2003 case Op_AddD:
2004 case Op_SubD:
2005 case Op_MulD:
2006 case Op_DivD:
2007 case Op_NegD:
2008 case Op_ModD:
2009 case Op_ConvI2D:
2010 case Op_ConvD2I:
2011 // case Op_ConvL2D: // handled by leaf call
2012 // case Op_ConvD2L: // handled by leaf call
2013 case Op_ConD:
2014 case Op_CmpD:
2015 case Op_CmpD3:
2016 frc.inc_double_count();
2017 break;
2018 case Op_Opaque1: // Remove Opaque Nodes before matching
2019 case Op_Opaque2: // Remove Opaque Nodes before matching
2020 n->subsume_by(n->in(1));
2021 break;
2022 case Op_CallStaticJava:
2023 case Op_CallJava:
2024 case Op_CallDynamicJava:
2025 frc.inc_java_call_count(); // Count java call site;
2026 case Op_CallRuntime:
2027 case Op_CallLeaf:
2028 case Op_CallLeafNoFP: {
2029 assert( n->is_Call(), "" );
2030 CallNode *call = n->as_Call();
2031 // Count call sites where the FP mode bit would have to be flipped.
2032 // Do not count uncommon runtime calls:
2033 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2034 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2035 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2036 frc.inc_call_count(); // Count the call site
2037 } else { // See if uncommon argument is shared
2038 Node *n = call->in(TypeFunc::Parms);
2039 int nop = n->Opcode();
2040 // Clone shared simple arguments to uncommon calls, item (1).
2041 if( n->outcnt() > 1 &&
2042 !n->is_Proj() &&
2043 nop != Op_CreateEx &&
2044 nop != Op_CheckCastPP &&
2045 nop != Op_DecodeN &&
2046 !n->is_Mem() ) {
2047 Node *x = n->clone();
2048 call->set_req( TypeFunc::Parms, x );
2049 }
2050 }
2051 break;
2052 }
2054 case Op_StoreD:
2055 case Op_LoadD:
2056 case Op_LoadD_unaligned:
2057 frc.inc_double_count();
2058 goto handle_mem;
2059 case Op_StoreF:
2060 case Op_LoadF:
2061 frc.inc_float_count();
2062 goto handle_mem;
2064 case Op_StoreB:
2065 case Op_StoreC:
2066 case Op_StoreCM:
2067 case Op_StorePConditional:
2068 case Op_StoreI:
2069 case Op_StoreL:
2070 case Op_StoreIConditional:
2071 case Op_StoreLConditional:
2072 case Op_CompareAndSwapI:
2073 case Op_CompareAndSwapL:
2074 case Op_CompareAndSwapP:
2075 case Op_CompareAndSwapN:
2076 case Op_StoreP:
2077 case Op_StoreN:
2078 case Op_LoadB:
2079 case Op_LoadUB:
2080 case Op_LoadUS:
2081 case Op_LoadI:
2082 case Op_LoadUI2L:
2083 case Op_LoadKlass:
2084 case Op_LoadNKlass:
2085 case Op_LoadL:
2086 case Op_LoadL_unaligned:
2087 case Op_LoadPLocked:
2088 case Op_LoadLLocked:
2089 case Op_LoadP:
2090 case Op_LoadN:
2091 case Op_LoadRange:
2092 case Op_LoadS: {
2093 handle_mem:
2094 #ifdef ASSERT
2095 if( VerifyOptoOopOffsets ) {
2096 assert( n->is_Mem(), "" );
2097 MemNode *mem = (MemNode*)n;
2098 // Check to see if address types have grounded out somehow.
2099 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2100 assert( !tp || oop_offset_is_sane(tp), "" );
2101 }
2102 #endif
2103 break;
2104 }
2106 case Op_AddP: { // Assert sane base pointers
2107 Node *addp = n->in(AddPNode::Address);
2108 assert( !addp->is_AddP() ||
2109 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2110 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2111 "Base pointers must match" );
2112 #ifdef _LP64
2113 if (UseCompressedOops &&
2114 addp->Opcode() == Op_ConP &&
2115 addp == n->in(AddPNode::Base) &&
2116 n->in(AddPNode::Offset)->is_Con()) {
2117 // Use addressing with narrow klass to load with offset on x86.
2118 // On sparc loading 32-bits constant and decoding it have less
2119 // instructions (4) then load 64-bits constant (7).
2120 // Do this transformation here since IGVN will convert ConN back to ConP.
2121 const Type* t = addp->bottom_type();
2122 if (t->isa_oopptr()) {
2123 Node* nn = NULL;
2125 // Look for existing ConN node of the same exact type.
2126 Compile* C = Compile::current();
2127 Node* r = C->root();
2128 uint cnt = r->outcnt();
2129 for (uint i = 0; i < cnt; i++) {
2130 Node* m = r->raw_out(i);
2131 if (m!= NULL && m->Opcode() == Op_ConN &&
2132 m->bottom_type()->make_ptr() == t) {
2133 nn = m;
2134 break;
2135 }
2136 }
2137 if (nn != NULL) {
2138 // Decode a narrow oop to match address
2139 // [R12 + narrow_oop_reg<<3 + offset]
2140 nn = new (C, 2) DecodeNNode(nn, t);
2141 n->set_req(AddPNode::Base, nn);
2142 n->set_req(AddPNode::Address, nn);
2143 if (addp->outcnt() == 0) {
2144 addp->disconnect_inputs(NULL);
2145 }
2146 }
2147 }
2148 }
2149 #endif
2150 break;
2151 }
2153 #ifdef _LP64
2154 case Op_CastPP:
2155 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) {
2156 Compile* C = Compile::current();
2157 Node* in1 = n->in(1);
2158 const Type* t = n->bottom_type();
2159 Node* new_in1 = in1->clone();
2160 new_in1->as_DecodeN()->set_type(t);
2162 if (!Matcher::clone_shift_expressions) {
2163 //
2164 // x86, ARM and friends can handle 2 adds in addressing mode
2165 // and Matcher can fold a DecodeN node into address by using
2166 // a narrow oop directly and do implicit NULL check in address:
2167 //
2168 // [R12 + narrow_oop_reg<<3 + offset]
2169 // NullCheck narrow_oop_reg
2170 //
2171 // On other platforms (Sparc) we have to keep new DecodeN node and
2172 // use it to do implicit NULL check in address:
2173 //
2174 // decode_not_null narrow_oop_reg, base_reg
2175 // [base_reg + offset]
2176 // NullCheck base_reg
2177 //
2178 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2179 // to keep the information to which NULL check the new DecodeN node
2180 // corresponds to use it as value in implicit_null_check().
2181 //
2182 new_in1->set_req(0, n->in(0));
2183 }
2185 n->subsume_by(new_in1);
2186 if (in1->outcnt() == 0) {
2187 in1->disconnect_inputs(NULL);
2188 }
2189 }
2190 break;
2192 case Op_CmpP:
2193 // Do this transformation here to preserve CmpPNode::sub() and
2194 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2195 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2196 Node* in1 = n->in(1);
2197 Node* in2 = n->in(2);
2198 if (!in1->is_DecodeN()) {
2199 in2 = in1;
2200 in1 = n->in(2);
2201 }
2202 assert(in1->is_DecodeN(), "sanity");
2204 Compile* C = Compile::current();
2205 Node* new_in2 = NULL;
2206 if (in2->is_DecodeN()) {
2207 new_in2 = in2->in(1);
2208 } else if (in2->Opcode() == Op_ConP) {
2209 const Type* t = in2->bottom_type();
2210 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) {
2211 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2212 //
2213 // This transformation together with CastPP transformation above
2214 // will generated code for implicit NULL checks for compressed oops.
2215 //
2216 // The original code after Optimize()
2217 //
2218 // LoadN memory, narrow_oop_reg
2219 // decode narrow_oop_reg, base_reg
2220 // CmpP base_reg, NULL
2221 // CastPP base_reg // NotNull
2222 // Load [base_reg + offset], val_reg
2223 //
2224 // after these transformations will be
2225 //
2226 // LoadN memory, narrow_oop_reg
2227 // CmpN narrow_oop_reg, NULL
2228 // decode_not_null narrow_oop_reg, base_reg
2229 // Load [base_reg + offset], val_reg
2230 //
2231 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2232 // since narrow oops can be used in debug info now (see the code in
2233 // final_graph_reshaping_walk()).
2234 //
2235 // At the end the code will be matched to
2236 // on x86:
2237 //
2238 // Load_narrow_oop memory, narrow_oop_reg
2239 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2240 // NullCheck narrow_oop_reg
2241 //
2242 // and on sparc:
2243 //
2244 // Load_narrow_oop memory, narrow_oop_reg
2245 // decode_not_null narrow_oop_reg, base_reg
2246 // Load [base_reg + offset], val_reg
2247 // NullCheck base_reg
2248 //
2249 } else if (t->isa_oopptr()) {
2250 new_in2 = ConNode::make(C, t->make_narrowoop());
2251 }
2252 }
2253 if (new_in2 != NULL) {
2254 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2255 n->subsume_by( cmpN );
2256 if (in1->outcnt() == 0) {
2257 in1->disconnect_inputs(NULL);
2258 }
2259 if (in2->outcnt() == 0) {
2260 in2->disconnect_inputs(NULL);
2261 }
2262 }
2263 }
2264 break;
2266 case Op_DecodeN:
2267 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2268 // DecodeN could be pinned on Sparc where it can't be fold into
2269 // an address expression, see the code for Op_CastPP above.
2270 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc");
2271 break;
2273 case Op_EncodeP: {
2274 Node* in1 = n->in(1);
2275 if (in1->is_DecodeN()) {
2276 n->subsume_by(in1->in(1));
2277 } else if (in1->Opcode() == Op_ConP) {
2278 Compile* C = Compile::current();
2279 const Type* t = in1->bottom_type();
2280 if (t == TypePtr::NULL_PTR) {
2281 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2282 } else if (t->isa_oopptr()) {
2283 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2284 }
2285 }
2286 if (in1->outcnt() == 0) {
2287 in1->disconnect_inputs(NULL);
2288 }
2289 break;
2290 }
2292 case Op_Proj: {
2293 if (OptimizeStringConcat) {
2294 ProjNode* p = n->as_Proj();
2295 if (p->_is_io_use) {
2296 // Separate projections were used for the exception path which
2297 // are normally removed by a late inline. If it wasn't inlined
2298 // then they will hang around and should just be replaced with
2299 // the original one.
2300 Node* proj = NULL;
2301 // Replace with just one
2302 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2303 Node *use = i.get();
2304 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2305 proj = use;
2306 break;
2307 }
2308 }
2309 assert(p != NULL, "must be found");
2310 p->subsume_by(proj);
2311 }
2312 }
2313 break;
2314 }
2316 case Op_Phi:
2317 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2318 // The EncodeP optimization may create Phi with the same edges
2319 // for all paths. It is not handled well by Register Allocator.
2320 Node* unique_in = n->in(1);
2321 assert(unique_in != NULL, "");
2322 uint cnt = n->req();
2323 for (uint i = 2; i < cnt; i++) {
2324 Node* m = n->in(i);
2325 assert(m != NULL, "");
2326 if (unique_in != m)
2327 unique_in = NULL;
2328 }
2329 if (unique_in != NULL) {
2330 n->subsume_by(unique_in);
2331 }
2332 }
2333 break;
2335 #endif
2337 case Op_ModI:
2338 if (UseDivMod) {
2339 // Check if a%b and a/b both exist
2340 Node* d = n->find_similar(Op_DivI);
2341 if (d) {
2342 // Replace them with a fused divmod if supported
2343 Compile* C = Compile::current();
2344 if (Matcher::has_match_rule(Op_DivModI)) {
2345 DivModINode* divmod = DivModINode::make(C, n);
2346 d->subsume_by(divmod->div_proj());
2347 n->subsume_by(divmod->mod_proj());
2348 } else {
2349 // replace a%b with a-((a/b)*b)
2350 Node* mult = new (C, 3) MulINode(d, d->in(2));
2351 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2352 n->subsume_by( sub );
2353 }
2354 }
2355 }
2356 break;
2358 case Op_ModL:
2359 if (UseDivMod) {
2360 // Check if a%b and a/b both exist
2361 Node* d = n->find_similar(Op_DivL);
2362 if (d) {
2363 // Replace them with a fused divmod if supported
2364 Compile* C = Compile::current();
2365 if (Matcher::has_match_rule(Op_DivModL)) {
2366 DivModLNode* divmod = DivModLNode::make(C, n);
2367 d->subsume_by(divmod->div_proj());
2368 n->subsume_by(divmod->mod_proj());
2369 } else {
2370 // replace a%b with a-((a/b)*b)
2371 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2372 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2373 n->subsume_by( sub );
2374 }
2375 }
2376 }
2377 break;
2379 case Op_Load16B:
2380 case Op_Load8B:
2381 case Op_Load4B:
2382 case Op_Load8S:
2383 case Op_Load4S:
2384 case Op_Load2S:
2385 case Op_Load8C:
2386 case Op_Load4C:
2387 case Op_Load2C:
2388 case Op_Load4I:
2389 case Op_Load2I:
2390 case Op_Load2L:
2391 case Op_Load4F:
2392 case Op_Load2F:
2393 case Op_Load2D:
2394 case Op_Store16B:
2395 case Op_Store8B:
2396 case Op_Store4B:
2397 case Op_Store8C:
2398 case Op_Store4C:
2399 case Op_Store2C:
2400 case Op_Store4I:
2401 case Op_Store2I:
2402 case Op_Store2L:
2403 case Op_Store4F:
2404 case Op_Store2F:
2405 case Op_Store2D:
2406 break;
2408 case Op_PackB:
2409 case Op_PackS:
2410 case Op_PackC:
2411 case Op_PackI:
2412 case Op_PackF:
2413 case Op_PackL:
2414 case Op_PackD:
2415 if (n->req()-1 > 2) {
2416 // Replace many operand PackNodes with a binary tree for matching
2417 PackNode* p = (PackNode*) n;
2418 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2419 n->subsume_by(btp);
2420 }
2421 break;
2422 case Op_Loop:
2423 case Op_CountedLoop:
2424 if (n->as_Loop()->is_inner_loop()) {
2425 frc.inc_inner_loop_count();
2426 }
2427 break;
2428 default:
2429 assert( !n->is_Call(), "" );
2430 assert( !n->is_Mem(), "" );
2431 break;
2432 }
2434 // Collect CFG split points
2435 if (n->is_MultiBranch())
2436 frc._tests.push(n);
2437 }
2439 //------------------------------final_graph_reshaping_walk---------------------
2440 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2441 // requires that the walk visits a node's inputs before visiting the node.
2442 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2443 ResourceArea *area = Thread::current()->resource_area();
2444 Unique_Node_List sfpt(area);
2446 frc._visited.set(root->_idx); // first, mark node as visited
2447 uint cnt = root->req();
2448 Node *n = root;
2449 uint i = 0;
2450 while (true) {
2451 if (i < cnt) {
2452 // Place all non-visited non-null inputs onto stack
2453 Node* m = n->in(i);
2454 ++i;
2455 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2456 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2457 sfpt.push(m);
2458 cnt = m->req();
2459 nstack.push(n, i); // put on stack parent and next input's index
2460 n = m;
2461 i = 0;
2462 }
2463 } else {
2464 // Now do post-visit work
2465 final_graph_reshaping_impl( n, frc );
2466 if (nstack.is_empty())
2467 break; // finished
2468 n = nstack.node(); // Get node from stack
2469 cnt = n->req();
2470 i = nstack.index();
2471 nstack.pop(); // Shift to the next node on stack
2472 }
2473 }
2475 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2476 // It could be done for an uncommon traps or any safepoints/calls
2477 // if the DecodeN node is referenced only in a debug info.
2478 while (sfpt.size() > 0) {
2479 n = sfpt.pop();
2480 JVMState *jvms = n->as_SafePoint()->jvms();
2481 assert(jvms != NULL, "sanity");
2482 int start = jvms->debug_start();
2483 int end = n->req();
2484 bool is_uncommon = (n->is_CallStaticJava() &&
2485 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2486 for (int j = start; j < end; j++) {
2487 Node* in = n->in(j);
2488 if (in->is_DecodeN()) {
2489 bool safe_to_skip = true;
2490 if (!is_uncommon ) {
2491 // Is it safe to skip?
2492 for (uint i = 0; i < in->outcnt(); i++) {
2493 Node* u = in->raw_out(i);
2494 if (!u->is_SafePoint() ||
2495 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2496 safe_to_skip = false;
2497 }
2498 }
2499 }
2500 if (safe_to_skip) {
2501 n->set_req(j, in->in(1));
2502 }
2503 if (in->outcnt() == 0) {
2504 in->disconnect_inputs(NULL);
2505 }
2506 }
2507 }
2508 }
2509 }
2511 //------------------------------final_graph_reshaping--------------------------
2512 // Final Graph Reshaping.
2513 //
2514 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2515 // and not commoned up and forced early. Must come after regular
2516 // optimizations to avoid GVN undoing the cloning. Clone constant
2517 // inputs to Loop Phis; these will be split by the allocator anyways.
2518 // Remove Opaque nodes.
2519 // (2) Move last-uses by commutative operations to the left input to encourage
2520 // Intel update-in-place two-address operations and better register usage
2521 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2522 // calls canonicalizing them back.
2523 // (3) Count the number of double-precision FP ops, single-precision FP ops
2524 // and call sites. On Intel, we can get correct rounding either by
2525 // forcing singles to memory (requires extra stores and loads after each
2526 // FP bytecode) or we can set a rounding mode bit (requires setting and
2527 // clearing the mode bit around call sites). The mode bit is only used
2528 // if the relative frequency of single FP ops to calls is low enough.
2529 // This is a key transform for SPEC mpeg_audio.
2530 // (4) Detect infinite loops; blobs of code reachable from above but not
2531 // below. Several of the Code_Gen algorithms fail on such code shapes,
2532 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2533 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2534 // Detection is by looking for IfNodes where only 1 projection is
2535 // reachable from below or CatchNodes missing some targets.
2536 // (5) Assert for insane oop offsets in debug mode.
2538 bool Compile::final_graph_reshaping() {
2539 // an infinite loop may have been eliminated by the optimizer,
2540 // in which case the graph will be empty.
2541 if (root()->req() == 1) {
2542 record_method_not_compilable("trivial infinite loop");
2543 return true;
2544 }
2546 Final_Reshape_Counts frc;
2548 // Visit everybody reachable!
2549 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2550 Node_Stack nstack(unique() >> 1);
2551 final_graph_reshaping_walk(nstack, root(), frc);
2553 // Check for unreachable (from below) code (i.e., infinite loops).
2554 for( uint i = 0; i < frc._tests.size(); i++ ) {
2555 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2556 // Get number of CFG targets.
2557 // Note that PCTables include exception targets after calls.
2558 uint required_outcnt = n->required_outcnt();
2559 if (n->outcnt() != required_outcnt) {
2560 // Check for a few special cases. Rethrow Nodes never take the
2561 // 'fall-thru' path, so expected kids is 1 less.
2562 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2563 if (n->in(0)->in(0)->is_Call()) {
2564 CallNode *call = n->in(0)->in(0)->as_Call();
2565 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2566 required_outcnt--; // Rethrow always has 1 less kid
2567 } else if (call->req() > TypeFunc::Parms &&
2568 call->is_CallDynamicJava()) {
2569 // Check for null receiver. In such case, the optimizer has
2570 // detected that the virtual call will always result in a null
2571 // pointer exception. The fall-through projection of this CatchNode
2572 // will not be populated.
2573 Node *arg0 = call->in(TypeFunc::Parms);
2574 if (arg0->is_Type() &&
2575 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2576 required_outcnt--;
2577 }
2578 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2579 call->req() > TypeFunc::Parms+1 &&
2580 call->is_CallStaticJava()) {
2581 // Check for negative array length. In such case, the optimizer has
2582 // detected that the allocation attempt will always result in an
2583 // exception. There is no fall-through projection of this CatchNode .
2584 Node *arg1 = call->in(TypeFunc::Parms+1);
2585 if (arg1->is_Type() &&
2586 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2587 required_outcnt--;
2588 }
2589 }
2590 }
2591 }
2592 // Recheck with a better notion of 'required_outcnt'
2593 if (n->outcnt() != required_outcnt) {
2594 record_method_not_compilable("malformed control flow");
2595 return true; // Not all targets reachable!
2596 }
2597 }
2598 // Check that I actually visited all kids. Unreached kids
2599 // must be infinite loops.
2600 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2601 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2602 record_method_not_compilable("infinite loop");
2603 return true; // Found unvisited kid; must be unreach
2604 }
2605 }
2607 // If original bytecodes contained a mixture of floats and doubles
2608 // check if the optimizer has made it homogenous, item (3).
2609 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2610 frc.get_float_count() > 32 &&
2611 frc.get_double_count() == 0 &&
2612 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2613 set_24_bit_selection_and_mode( false, true );
2614 }
2616 set_java_calls(frc.get_java_call_count());
2617 set_inner_loops(frc.get_inner_loop_count());
2619 // No infinite loops, no reason to bail out.
2620 return false;
2621 }
2623 //-----------------------------too_many_traps----------------------------------
2624 // Report if there are too many traps at the current method and bci.
2625 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2626 bool Compile::too_many_traps(ciMethod* method,
2627 int bci,
2628 Deoptimization::DeoptReason reason) {
2629 ciMethodData* md = method->method_data();
2630 if (md->is_empty()) {
2631 // Assume the trap has not occurred, or that it occurred only
2632 // because of a transient condition during start-up in the interpreter.
2633 return false;
2634 }
2635 if (md->has_trap_at(bci, reason) != 0) {
2636 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2637 // Also, if there are multiple reasons, or if there is no per-BCI record,
2638 // assume the worst.
2639 if (log())
2640 log()->elem("observe trap='%s' count='%d'",
2641 Deoptimization::trap_reason_name(reason),
2642 md->trap_count(reason));
2643 return true;
2644 } else {
2645 // Ignore method/bci and see if there have been too many globally.
2646 return too_many_traps(reason, md);
2647 }
2648 }
2650 // Less-accurate variant which does not require a method and bci.
2651 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2652 ciMethodData* logmd) {
2653 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2654 // Too many traps globally.
2655 // Note that we use cumulative trap_count, not just md->trap_count.
2656 if (log()) {
2657 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2658 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2659 Deoptimization::trap_reason_name(reason),
2660 mcount, trap_count(reason));
2661 }
2662 return true;
2663 } else {
2664 // The coast is clear.
2665 return false;
2666 }
2667 }
2669 //--------------------------too_many_recompiles--------------------------------
2670 // Report if there are too many recompiles at the current method and bci.
2671 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2672 // Is not eager to return true, since this will cause the compiler to use
2673 // Action_none for a trap point, to avoid too many recompilations.
2674 bool Compile::too_many_recompiles(ciMethod* method,
2675 int bci,
2676 Deoptimization::DeoptReason reason) {
2677 ciMethodData* md = method->method_data();
2678 if (md->is_empty()) {
2679 // Assume the trap has not occurred, or that it occurred only
2680 // because of a transient condition during start-up in the interpreter.
2681 return false;
2682 }
2683 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2684 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2685 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2686 Deoptimization::DeoptReason per_bc_reason
2687 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2688 if ((per_bc_reason == Deoptimization::Reason_none
2689 || md->has_trap_at(bci, reason) != 0)
2690 // The trap frequency measure we care about is the recompile count:
2691 && md->trap_recompiled_at(bci)
2692 && md->overflow_recompile_count() >= bc_cutoff) {
2693 // Do not emit a trap here if it has already caused recompilations.
2694 // Also, if there are multiple reasons, or if there is no per-BCI record,
2695 // assume the worst.
2696 if (log())
2697 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2698 Deoptimization::trap_reason_name(reason),
2699 md->trap_count(reason),
2700 md->overflow_recompile_count());
2701 return true;
2702 } else if (trap_count(reason) != 0
2703 && decompile_count() >= m_cutoff) {
2704 // Too many recompiles globally, and we have seen this sort of trap.
2705 // Use cumulative decompile_count, not just md->decompile_count.
2706 if (log())
2707 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2708 Deoptimization::trap_reason_name(reason),
2709 md->trap_count(reason), trap_count(reason),
2710 md->decompile_count(), decompile_count());
2711 return true;
2712 } else {
2713 // The coast is clear.
2714 return false;
2715 }
2716 }
2719 #ifndef PRODUCT
2720 //------------------------------verify_graph_edges---------------------------
2721 // Walk the Graph and verify that there is a one-to-one correspondence
2722 // between Use-Def edges and Def-Use edges in the graph.
2723 void Compile::verify_graph_edges(bool no_dead_code) {
2724 if (VerifyGraphEdges) {
2725 ResourceArea *area = Thread::current()->resource_area();
2726 Unique_Node_List visited(area);
2727 // Call recursive graph walk to check edges
2728 _root->verify_edges(visited);
2729 if (no_dead_code) {
2730 // Now make sure that no visited node is used by an unvisited node.
2731 bool dead_nodes = 0;
2732 Unique_Node_List checked(area);
2733 while (visited.size() > 0) {
2734 Node* n = visited.pop();
2735 checked.push(n);
2736 for (uint i = 0; i < n->outcnt(); i++) {
2737 Node* use = n->raw_out(i);
2738 if (checked.member(use)) continue; // already checked
2739 if (visited.member(use)) continue; // already in the graph
2740 if (use->is_Con()) continue; // a dead ConNode is OK
2741 // At this point, we have found a dead node which is DU-reachable.
2742 if (dead_nodes++ == 0)
2743 tty->print_cr("*** Dead nodes reachable via DU edges:");
2744 use->dump(2);
2745 tty->print_cr("---");
2746 checked.push(use); // No repeats; pretend it is now checked.
2747 }
2748 }
2749 assert(dead_nodes == 0, "using nodes must be reachable from root");
2750 }
2751 }
2752 }
2753 #endif
2755 // The Compile object keeps track of failure reasons separately from the ciEnv.
2756 // This is required because there is not quite a 1-1 relation between the
2757 // ciEnv and its compilation task and the Compile object. Note that one
2758 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2759 // to backtrack and retry without subsuming loads. Other than this backtracking
2760 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2761 // by the logic in C2Compiler.
2762 void Compile::record_failure(const char* reason) {
2763 if (log() != NULL) {
2764 log()->elem("failure reason='%s' phase='compile'", reason);
2765 }
2766 if (_failure_reason == NULL) {
2767 // Record the first failure reason.
2768 _failure_reason = reason;
2769 }
2770 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2771 C->print_method(_failure_reason);
2772 }
2773 _root = NULL; // flush the graph, too
2774 }
2776 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2777 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2778 {
2779 if (dolog) {
2780 C = Compile::current();
2781 _log = C->log();
2782 } else {
2783 C = NULL;
2784 _log = NULL;
2785 }
2786 if (_log != NULL) {
2787 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2788 _log->stamp();
2789 _log->end_head();
2790 }
2791 }
2793 Compile::TracePhase::~TracePhase() {
2794 if (_log != NULL) {
2795 _log->done("phase nodes='%d'", C->unique());
2796 }
2797 }