Fri, 03 Sep 2010 17:51:07 -0700
6953144: Tiered compilation
Summary: Infrastructure for tiered compilation support (interpreter + c1 + c2) for 32 and 64 bit. Simple tiered policy implementation.
Reviewed-by: kvn, never, phh, twisti
1 /*
2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * 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->content_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->content_begin();
426 address blob_end = (address)locs_buf;
427 assert(blob->content_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.insts_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 _has_method_handle_invokes(false),
469 _node_bundling_limit(0),
470 _node_bundling_base(NULL),
471 _java_calls(0),
472 _inner_loops(0),
473 #ifndef PRODUCT
474 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
475 _printer(IdealGraphPrinter::printer()),
476 #endif
477 _congraph(NULL) {
478 C = this;
480 CompileWrapper cw(this);
481 #ifndef PRODUCT
482 if (TimeCompiler2) {
483 tty->print(" ");
484 target->holder()->name()->print();
485 tty->print(".");
486 target->print_short_name();
487 tty->print(" ");
488 }
489 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
490 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
491 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
492 if (!print_opto_assembly) {
493 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
494 if (print_assembly && !Disassembler::can_decode()) {
495 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
496 print_opto_assembly = true;
497 }
498 }
499 set_print_assembly(print_opto_assembly);
500 set_parsed_irreducible_loop(false);
501 #endif
503 if (ProfileTraps) {
504 // Make sure the method being compiled gets its own MDO,
505 // so we can at least track the decompile_count().
506 method()->build_method_data();
507 }
509 Init(::AliasLevel);
512 print_compile_messages();
514 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
515 _ilt = InlineTree::build_inline_tree_root();
516 else
517 _ilt = NULL;
519 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
520 assert(num_alias_types() >= AliasIdxRaw, "");
522 #define MINIMUM_NODE_HASH 1023
523 // Node list that Iterative GVN will start with
524 Unique_Node_List for_igvn(comp_arena());
525 set_for_igvn(&for_igvn);
527 // GVN that will be run immediately on new nodes
528 uint estimated_size = method()->code_size()*4+64;
529 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
530 PhaseGVN gvn(node_arena(), estimated_size);
531 set_initial_gvn(&gvn);
533 { // Scope for timing the parser
534 TracePhase t3("parse", &_t_parser, true);
536 // Put top into the hash table ASAP.
537 initial_gvn()->transform_no_reclaim(top());
539 // Set up tf(), start(), and find a CallGenerator.
540 CallGenerator* cg;
541 if (is_osr_compilation()) {
542 const TypeTuple *domain = StartOSRNode::osr_domain();
543 const TypeTuple *range = TypeTuple::make_range(method()->signature());
544 init_tf(TypeFunc::make(domain, range));
545 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
546 initial_gvn()->set_type_bottom(s);
547 init_start(s);
548 cg = CallGenerator::for_osr(method(), entry_bci());
549 } else {
550 // Normal case.
551 init_tf(TypeFunc::make(method()));
552 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
553 initial_gvn()->set_type_bottom(s);
554 init_start(s);
555 float past_uses = method()->interpreter_invocation_count();
556 float expected_uses = past_uses;
557 cg = CallGenerator::for_inline(method(), expected_uses);
558 }
559 if (failing()) return;
560 if (cg == NULL) {
561 record_method_not_compilable_all_tiers("cannot parse method");
562 return;
563 }
564 JVMState* jvms = build_start_state(start(), tf());
565 if ((jvms = cg->generate(jvms)) == NULL) {
566 record_method_not_compilable("method parse failed");
567 return;
568 }
569 GraphKit kit(jvms);
571 if (!kit.stopped()) {
572 // Accept return values, and transfer control we know not where.
573 // This is done by a special, unique ReturnNode bound to root.
574 return_values(kit.jvms());
575 }
577 if (kit.has_exceptions()) {
578 // Any exceptions that escape from this call must be rethrown
579 // to whatever caller is dynamically above us on the stack.
580 // This is done by a special, unique RethrowNode bound to root.
581 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
582 }
584 if (!failing() && has_stringbuilder()) {
585 {
586 // remove useless nodes to make the usage analysis simpler
587 ResourceMark rm;
588 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
589 }
591 {
592 ResourceMark rm;
593 print_method("Before StringOpts", 3);
594 PhaseStringOpts pso(initial_gvn(), &for_igvn);
595 print_method("After StringOpts", 3);
596 }
598 // now inline anything that we skipped the first time around
599 while (_late_inlines.length() > 0) {
600 CallGenerator* cg = _late_inlines.pop();
601 cg->do_late_inline();
602 }
603 }
604 assert(_late_inlines.length() == 0, "should have been processed");
606 print_method("Before RemoveUseless", 3);
608 // Remove clutter produced by parsing.
609 if (!failing()) {
610 ResourceMark rm;
611 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
612 }
613 }
615 // Note: Large methods are capped off in do_one_bytecode().
616 if (failing()) return;
618 // After parsing, node notes are no longer automagic.
619 // They must be propagated by register_new_node_with_optimizer(),
620 // clone(), or the like.
621 set_default_node_notes(NULL);
623 for (;;) {
624 int successes = Inline_Warm();
625 if (failing()) return;
626 if (successes == 0) break;
627 }
629 // Drain the list.
630 Finish_Warm();
631 #ifndef PRODUCT
632 if (_printer) {
633 _printer->print_inlining(this);
634 }
635 #endif
637 if (failing()) return;
638 NOT_PRODUCT( verify_graph_edges(); )
640 // Now optimize
641 Optimize();
642 if (failing()) return;
643 NOT_PRODUCT( verify_graph_edges(); )
645 #ifndef PRODUCT
646 if (PrintIdeal) {
647 ttyLocker ttyl; // keep the following output all in one block
648 // This output goes directly to the tty, not the compiler log.
649 // To enable tools to match it up with the compilation activity,
650 // be sure to tag this tty output with the compile ID.
651 if (xtty != NULL) {
652 xtty->head("ideal compile_id='%d'%s", compile_id(),
653 is_osr_compilation() ? " compile_kind='osr'" :
654 "");
655 }
656 root()->dump(9999);
657 if (xtty != NULL) {
658 xtty->tail("ideal");
659 }
660 }
661 #endif
663 // Now that we know the size of all the monitors we can add a fixed slot
664 // for the original deopt pc.
666 _orig_pc_slot = fixed_slots();
667 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
668 set_fixed_slots(next_slot);
670 // Now generate code
671 Code_Gen();
672 if (failing()) return;
674 // Check if we want to skip execution of all compiled code.
675 {
676 #ifndef PRODUCT
677 if (OptoNoExecute) {
678 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
679 return;
680 }
681 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
682 #endif
684 if (is_osr_compilation()) {
685 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
686 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
687 } else {
688 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
689 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
690 }
692 env()->register_method(_method, _entry_bci,
693 &_code_offsets,
694 _orig_pc_slot_offset_in_bytes,
695 code_buffer(),
696 frame_size_in_words(), _oop_map_set,
697 &_handler_table, &_inc_table,
698 compiler,
699 env()->comp_level(),
700 true, /*has_debug_info*/
701 has_unsafe_access()
702 );
703 }
704 }
706 //------------------------------Compile----------------------------------------
707 // Compile a runtime stub
708 Compile::Compile( ciEnv* ci_env,
709 TypeFunc_generator generator,
710 address stub_function,
711 const char *stub_name,
712 int is_fancy_jump,
713 bool pass_tls,
714 bool save_arg_registers,
715 bool return_pc )
716 : Phase(Compiler),
717 _env(ci_env),
718 _log(ci_env->log()),
719 _compile_id(-1),
720 _save_argument_registers(save_arg_registers),
721 _method(NULL),
722 _stub_name(stub_name),
723 _stub_function(stub_function),
724 _stub_entry_point(NULL),
725 _entry_bci(InvocationEntryBci),
726 _initial_gvn(NULL),
727 _for_igvn(NULL),
728 _warm_calls(NULL),
729 _orig_pc_slot(0),
730 _orig_pc_slot_offset_in_bytes(0),
731 _subsume_loads(true),
732 _do_escape_analysis(false),
733 _failure_reason(NULL),
734 _code_buffer("Compile::Fill_buffer"),
735 _has_method_handle_invokes(false),
736 _node_bundling_limit(0),
737 _node_bundling_base(NULL),
738 _java_calls(0),
739 _inner_loops(0),
740 #ifndef PRODUCT
741 _trace_opto_output(TraceOptoOutput),
742 _printer(NULL),
743 #endif
744 _congraph(NULL) {
745 C = this;
747 #ifndef PRODUCT
748 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
749 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
750 set_print_assembly(PrintFrameConverterAssembly);
751 set_parsed_irreducible_loop(false);
752 #endif
753 CompileWrapper cw(this);
754 Init(/*AliasLevel=*/ 0);
755 init_tf((*generator)());
757 {
758 // The following is a dummy for the sake of GraphKit::gen_stub
759 Unique_Node_List for_igvn(comp_arena());
760 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
761 PhaseGVN gvn(Thread::current()->resource_area(),255);
762 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
763 gvn.transform_no_reclaim(top());
765 GraphKit kit;
766 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
767 }
769 NOT_PRODUCT( verify_graph_edges(); )
770 Code_Gen();
771 if (failing()) return;
774 // Entry point will be accessed using compile->stub_entry_point();
775 if (code_buffer() == NULL) {
776 Matcher::soft_match_failure();
777 } else {
778 if (PrintAssembly && (WizardMode || Verbose))
779 tty->print_cr("### Stub::%s", stub_name);
781 if (!failing()) {
782 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
784 // Make the NMethod
785 // For now we mark the frame as never safe for profile stackwalking
786 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
787 code_buffer(),
788 CodeOffsets::frame_never_safe,
789 // _code_offsets.value(CodeOffsets::Frame_Complete),
790 frame_size_in_words(),
791 _oop_map_set,
792 save_arg_registers);
793 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
795 _stub_entry_point = rs->entry_point();
796 }
797 }
798 }
800 #ifndef PRODUCT
801 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
802 if(PrintOpto && Verbose) {
803 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
804 }
805 }
806 #endif
808 void Compile::print_codes() {
809 }
811 //------------------------------Init-------------------------------------------
812 // Prepare for a single compilation
813 void Compile::Init(int aliaslevel) {
814 _unique = 0;
815 _regalloc = NULL;
817 _tf = NULL; // filled in later
818 _top = NULL; // cached later
819 _matcher = NULL; // filled in later
820 _cfg = NULL; // filled in later
822 set_24_bit_selection_and_mode(Use24BitFP, false);
824 _node_note_array = NULL;
825 _default_node_notes = NULL;
827 _immutable_memory = NULL; // filled in at first inquiry
829 // Globally visible Nodes
830 // First set TOP to NULL to give safe behavior during creation of RootNode
831 set_cached_top_node(NULL);
832 set_root(new (this, 3) RootNode());
833 // Now that you have a Root to point to, create the real TOP
834 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
835 set_recent_alloc(NULL, NULL);
837 // Create Debug Information Recorder to record scopes, oopmaps, etc.
838 env()->set_oop_recorder(new OopRecorder(comp_arena()));
839 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
840 env()->set_dependencies(new Dependencies(env()));
842 _fixed_slots = 0;
843 set_has_split_ifs(false);
844 set_has_loops(has_method() && method()->has_loops()); // first approximation
845 set_has_stringbuilder(false);
846 _trap_can_recompile = false; // no traps emitted yet
847 _major_progress = true; // start out assuming good things will happen
848 set_has_unsafe_access(false);
849 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
850 set_decompile_count(0);
852 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
853 set_num_loop_opts(LoopOptsCount);
854 set_do_inlining(Inline);
855 set_max_inline_size(MaxInlineSize);
856 set_freq_inline_size(FreqInlineSize);
857 set_do_scheduling(OptoScheduling);
858 set_do_count_invocations(false);
859 set_do_method_data_update(false);
861 if (debug_info()->recording_non_safepoints()) {
862 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
863 (comp_arena(), 8, 0, NULL));
864 set_default_node_notes(Node_Notes::make(this));
865 }
867 // // -- Initialize types before each compile --
868 // // Update cached type information
869 // if( _method && _method->constants() )
870 // Type::update_loaded_types(_method, _method->constants());
872 // Init alias_type map.
873 if (!_do_escape_analysis && aliaslevel == 3)
874 aliaslevel = 2; // No unique types without escape analysis
875 _AliasLevel = aliaslevel;
876 const int grow_ats = 16;
877 _max_alias_types = grow_ats;
878 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
879 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
880 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
881 {
882 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
883 }
884 // Initialize the first few types.
885 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
886 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
887 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
888 _num_alias_types = AliasIdxRaw+1;
889 // Zero out the alias type cache.
890 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
891 // A NULL adr_type hits in the cache right away. Preload the right answer.
892 probe_alias_cache(NULL)->_index = AliasIdxTop;
894 _intrinsics = NULL;
895 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
896 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
897 register_library_intrinsics();
898 }
900 //---------------------------init_start----------------------------------------
901 // Install the StartNode on this compile object.
902 void Compile::init_start(StartNode* s) {
903 if (failing())
904 return; // already failing
905 assert(s == start(), "");
906 }
908 StartNode* Compile::start() const {
909 assert(!failing(), "");
910 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
911 Node* start = root()->fast_out(i);
912 if( start->is_Start() )
913 return start->as_Start();
914 }
915 ShouldNotReachHere();
916 return NULL;
917 }
919 //-------------------------------immutable_memory-------------------------------------
920 // Access immutable memory
921 Node* Compile::immutable_memory() {
922 if (_immutable_memory != NULL) {
923 return _immutable_memory;
924 }
925 StartNode* s = start();
926 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
927 Node *p = s->fast_out(i);
928 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
929 _immutable_memory = p;
930 return _immutable_memory;
931 }
932 }
933 ShouldNotReachHere();
934 return NULL;
935 }
937 //----------------------set_cached_top_node------------------------------------
938 // Install the cached top node, and make sure Node::is_top works correctly.
939 void Compile::set_cached_top_node(Node* tn) {
940 if (tn != NULL) verify_top(tn);
941 Node* old_top = _top;
942 _top = tn;
943 // Calling Node::setup_is_top allows the nodes the chance to adjust
944 // their _out arrays.
945 if (_top != NULL) _top->setup_is_top();
946 if (old_top != NULL) old_top->setup_is_top();
947 assert(_top == NULL || top()->is_top(), "");
948 }
950 #ifndef PRODUCT
951 void Compile::verify_top(Node* tn) const {
952 if (tn != NULL) {
953 assert(tn->is_Con(), "top node must be a constant");
954 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
955 assert(tn->in(0) != NULL, "must have live top node");
956 }
957 }
958 #endif
961 ///-------------------Managing Per-Node Debug & Profile Info-------------------
963 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
964 guarantee(arr != NULL, "");
965 int num_blocks = arr->length();
966 if (grow_by < num_blocks) grow_by = num_blocks;
967 int num_notes = grow_by * _node_notes_block_size;
968 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
969 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
970 while (num_notes > 0) {
971 arr->append(notes);
972 notes += _node_notes_block_size;
973 num_notes -= _node_notes_block_size;
974 }
975 assert(num_notes == 0, "exact multiple, please");
976 }
978 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
979 if (source == NULL || dest == NULL) return false;
981 if (dest->is_Con())
982 return false; // Do not push debug info onto constants.
984 #ifdef ASSERT
985 // Leave a bread crumb trail pointing to the original node:
986 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
987 dest->set_debug_orig(source);
988 }
989 #endif
991 if (node_note_array() == NULL)
992 return false; // Not collecting any notes now.
994 // This is a copy onto a pre-existing node, which may already have notes.
995 // If both nodes have notes, do not overwrite any pre-existing notes.
996 Node_Notes* source_notes = node_notes_at(source->_idx);
997 if (source_notes == NULL || source_notes->is_clear()) return false;
998 Node_Notes* dest_notes = node_notes_at(dest->_idx);
999 if (dest_notes == NULL || dest_notes->is_clear()) {
1000 return set_node_notes_at(dest->_idx, source_notes);
1001 }
1003 Node_Notes merged_notes = (*source_notes);
1004 // The order of operations here ensures that dest notes will win...
1005 merged_notes.update_from(dest_notes);
1006 return set_node_notes_at(dest->_idx, &merged_notes);
1007 }
1010 //--------------------------allow_range_check_smearing-------------------------
1011 // Gating condition for coalescing similar range checks.
1012 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1013 // single covering check that is at least as strong as any of them.
1014 // If the optimization succeeds, the simplified (strengthened) range check
1015 // will always succeed. If it fails, we will deopt, and then give up
1016 // on the optimization.
1017 bool Compile::allow_range_check_smearing() const {
1018 // If this method has already thrown a range-check,
1019 // assume it was because we already tried range smearing
1020 // and it failed.
1021 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1022 return !already_trapped;
1023 }
1026 //------------------------------flatten_alias_type-----------------------------
1027 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1028 int offset = tj->offset();
1029 TypePtr::PTR ptr = tj->ptr();
1031 // Known instance (scalarizable allocation) alias only with itself.
1032 bool is_known_inst = tj->isa_oopptr() != NULL &&
1033 tj->is_oopptr()->is_known_instance();
1035 // Process weird unsafe references.
1036 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1037 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1038 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1039 tj = TypeOopPtr::BOTTOM;
1040 ptr = tj->ptr();
1041 offset = tj->offset();
1042 }
1044 // Array pointers need some flattening
1045 const TypeAryPtr *ta = tj->isa_aryptr();
1046 if( ta && is_known_inst ) {
1047 if ( offset != Type::OffsetBot &&
1048 offset > arrayOopDesc::length_offset_in_bytes() ) {
1049 offset = Type::OffsetBot; // Flatten constant access into array body only
1050 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1051 }
1052 } else if( ta && _AliasLevel >= 2 ) {
1053 // For arrays indexed by constant indices, we flatten the alias
1054 // space to include all of the array body. Only the header, klass
1055 // and array length can be accessed un-aliased.
1056 if( offset != Type::OffsetBot ) {
1057 if( ta->const_oop() ) { // methodDataOop or methodOop
1058 offset = Type::OffsetBot; // Flatten constant access into array body
1059 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1060 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1061 // range is OK as-is.
1062 tj = ta = TypeAryPtr::RANGE;
1063 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1064 tj = TypeInstPtr::KLASS; // all klass loads look alike
1065 ta = TypeAryPtr::RANGE; // generic ignored junk
1066 ptr = TypePtr::BotPTR;
1067 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1068 tj = TypeInstPtr::MARK;
1069 ta = TypeAryPtr::RANGE; // generic ignored junk
1070 ptr = TypePtr::BotPTR;
1071 } else { // Random constant offset into array body
1072 offset = Type::OffsetBot; // Flatten constant access into array body
1073 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1074 }
1075 }
1076 // Arrays of fixed size alias with arrays of unknown size.
1077 if (ta->size() != TypeInt::POS) {
1078 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1079 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1080 }
1081 // Arrays of known objects become arrays of unknown objects.
1082 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1083 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1084 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1085 }
1086 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1087 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1088 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1089 }
1090 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1091 // cannot be distinguished by bytecode alone.
1092 if (ta->elem() == TypeInt::BOOL) {
1093 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1094 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1095 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1096 }
1097 // During the 2nd round of IterGVN, NotNull castings are removed.
1098 // Make sure the Bottom and NotNull variants alias the same.
1099 // Also, make sure exact and non-exact variants alias the same.
1100 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1101 if (ta->const_oop()) {
1102 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1103 } else {
1104 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1105 }
1106 }
1107 }
1109 // Oop pointers need some flattening
1110 const TypeInstPtr *to = tj->isa_instptr();
1111 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1112 if( ptr == TypePtr::Constant ) {
1113 // No constant oop pointers (such as Strings); they alias with
1114 // unknown strings.
1115 assert(!is_known_inst, "not scalarizable allocation");
1116 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1117 } else if( is_known_inst ) {
1118 tj = to; // Keep NotNull and klass_is_exact for instance type
1119 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1120 // During the 2nd round of IterGVN, NotNull castings are removed.
1121 // Make sure the Bottom and NotNull variants alias the same.
1122 // Also, make sure exact and non-exact variants alias the same.
1123 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1124 }
1125 // Canonicalize the holder of this field
1126 ciInstanceKlass *k = to->klass()->as_instance_klass();
1127 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1128 // First handle header references such as a LoadKlassNode, even if the
1129 // object's klass is unloaded at compile time (4965979).
1130 if (!is_known_inst) { // Do it only for non-instance types
1131 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1132 }
1133 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1134 to = NULL;
1135 tj = TypeOopPtr::BOTTOM;
1136 offset = tj->offset();
1137 } else {
1138 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1139 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1140 if( is_known_inst ) {
1141 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1142 } else {
1143 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1144 }
1145 }
1146 }
1147 }
1149 // Klass pointers to object array klasses need some flattening
1150 const TypeKlassPtr *tk = tj->isa_klassptr();
1151 if( tk ) {
1152 // If we are referencing a field within a Klass, we need
1153 // to assume the worst case of an Object. Both exact and
1154 // inexact types must flatten to the same alias class.
1155 // Since the flattened result for a klass is defined to be
1156 // precisely java.lang.Object, use a constant ptr.
1157 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1159 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1160 TypeKlassPtr::OBJECT->klass(),
1161 offset);
1162 }
1164 ciKlass* klass = tk->klass();
1165 if( klass->is_obj_array_klass() ) {
1166 ciKlass* k = TypeAryPtr::OOPS->klass();
1167 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1168 k = TypeInstPtr::BOTTOM->klass();
1169 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1170 }
1172 // Check for precise loads from the primary supertype array and force them
1173 // to the supertype cache alias index. Check for generic array loads from
1174 // the primary supertype array and also force them to the supertype cache
1175 // alias index. Since the same load can reach both, we need to merge
1176 // these 2 disparate memories into the same alias class. Since the
1177 // primary supertype array is read-only, there's no chance of confusion
1178 // where we bypass an array load and an array store.
1179 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1180 if( offset == Type::OffsetBot ||
1181 off2 < Klass::primary_super_limit()*wordSize ) {
1182 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1183 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1184 }
1185 }
1187 // Flatten all Raw pointers together.
1188 if (tj->base() == Type::RawPtr)
1189 tj = TypeRawPtr::BOTTOM;
1191 if (tj->base() == Type::AnyPtr)
1192 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1194 // Flatten all to bottom for now
1195 switch( _AliasLevel ) {
1196 case 0:
1197 tj = TypePtr::BOTTOM;
1198 break;
1199 case 1: // Flatten to: oop, static, field or array
1200 switch (tj->base()) {
1201 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1202 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1203 case Type::AryPtr: // do not distinguish arrays at all
1204 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1205 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1206 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1207 default: ShouldNotReachHere();
1208 }
1209 break;
1210 case 2: // No collapsing at level 2; keep all splits
1211 case 3: // No collapsing at level 3; keep all splits
1212 break;
1213 default:
1214 Unimplemented();
1215 }
1217 offset = tj->offset();
1218 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1220 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1221 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1222 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1223 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1224 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1225 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1226 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1227 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1228 assert( tj->ptr() != TypePtr::TopPTR &&
1229 tj->ptr() != TypePtr::AnyNull &&
1230 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1231 // assert( tj->ptr() != TypePtr::Constant ||
1232 // tj->base() == Type::RawPtr ||
1233 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1235 return tj;
1236 }
1238 void Compile::AliasType::Init(int i, const TypePtr* at) {
1239 _index = i;
1240 _adr_type = at;
1241 _field = NULL;
1242 _is_rewritable = true; // default
1243 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1244 if (atoop != NULL && atoop->is_known_instance()) {
1245 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1246 _general_index = Compile::current()->get_alias_index(gt);
1247 } else {
1248 _general_index = 0;
1249 }
1250 }
1252 //---------------------------------print_on------------------------------------
1253 #ifndef PRODUCT
1254 void Compile::AliasType::print_on(outputStream* st) {
1255 if (index() < 10)
1256 st->print("@ <%d> ", index());
1257 else st->print("@ <%d>", index());
1258 st->print(is_rewritable() ? " " : " RO");
1259 int offset = adr_type()->offset();
1260 if (offset == Type::OffsetBot)
1261 st->print(" +any");
1262 else st->print(" +%-3d", offset);
1263 st->print(" in ");
1264 adr_type()->dump_on(st);
1265 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1266 if (field() != NULL && tjp) {
1267 if (tjp->klass() != field()->holder() ||
1268 tjp->offset() != field()->offset_in_bytes()) {
1269 st->print(" != ");
1270 field()->print();
1271 st->print(" ***");
1272 }
1273 }
1274 }
1276 void print_alias_types() {
1277 Compile* C = Compile::current();
1278 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1279 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1280 C->alias_type(idx)->print_on(tty);
1281 tty->cr();
1282 }
1283 }
1284 #endif
1287 //----------------------------probe_alias_cache--------------------------------
1288 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1289 intptr_t key = (intptr_t) adr_type;
1290 key ^= key >> logAliasCacheSize;
1291 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1292 }
1295 //-----------------------------grow_alias_types--------------------------------
1296 void Compile::grow_alias_types() {
1297 const int old_ats = _max_alias_types; // how many before?
1298 const int new_ats = old_ats; // how many more?
1299 const int grow_ats = old_ats+new_ats; // how many now?
1300 _max_alias_types = grow_ats;
1301 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1302 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1303 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1304 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1305 }
1308 //--------------------------------find_alias_type------------------------------
1309 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1310 if (_AliasLevel == 0)
1311 return alias_type(AliasIdxBot);
1313 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1314 if (ace->_adr_type == adr_type) {
1315 return alias_type(ace->_index);
1316 }
1318 // Handle special cases.
1319 if (adr_type == NULL) return alias_type(AliasIdxTop);
1320 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1322 // Do it the slow way.
1323 const TypePtr* flat = flatten_alias_type(adr_type);
1325 #ifdef ASSERT
1326 assert(flat == flatten_alias_type(flat), "idempotent");
1327 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1328 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1329 const TypeOopPtr* foop = flat->is_oopptr();
1330 // Scalarizable allocations have exact klass always.
1331 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1332 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1333 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1334 }
1335 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1336 #endif
1338 int idx = AliasIdxTop;
1339 for (int i = 0; i < num_alias_types(); i++) {
1340 if (alias_type(i)->adr_type() == flat) {
1341 idx = i;
1342 break;
1343 }
1344 }
1346 if (idx == AliasIdxTop) {
1347 if (no_create) return NULL;
1348 // Grow the array if necessary.
1349 if (_num_alias_types == _max_alias_types) grow_alias_types();
1350 // Add a new alias type.
1351 idx = _num_alias_types++;
1352 _alias_types[idx]->Init(idx, flat);
1353 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1354 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1355 if (flat->isa_instptr()) {
1356 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1357 && flat->is_instptr()->klass() == env()->Class_klass())
1358 alias_type(idx)->set_rewritable(false);
1359 }
1360 if (flat->isa_klassptr()) {
1361 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1362 alias_type(idx)->set_rewritable(false);
1363 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1364 alias_type(idx)->set_rewritable(false);
1365 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1366 alias_type(idx)->set_rewritable(false);
1367 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1368 alias_type(idx)->set_rewritable(false);
1369 }
1370 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1371 // but the base pointer type is not distinctive enough to identify
1372 // references into JavaThread.)
1374 // Check for final instance fields.
1375 const TypeInstPtr* tinst = flat->isa_instptr();
1376 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1377 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1378 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1379 // Set field() and is_rewritable() attributes.
1380 if (field != NULL) alias_type(idx)->set_field(field);
1381 }
1382 const TypeKlassPtr* tklass = flat->isa_klassptr();
1383 // Check for final static fields.
1384 if (tklass && tklass->klass()->is_instance_klass()) {
1385 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1386 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1387 // Set field() and is_rewritable() attributes.
1388 if (field != NULL) alias_type(idx)->set_field(field);
1389 }
1390 }
1392 // Fill the cache for next time.
1393 ace->_adr_type = adr_type;
1394 ace->_index = idx;
1395 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1397 // Might as well try to fill the cache for the flattened version, too.
1398 AliasCacheEntry* face = probe_alias_cache(flat);
1399 if (face->_adr_type == NULL) {
1400 face->_adr_type = flat;
1401 face->_index = idx;
1402 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1403 }
1405 return alias_type(idx);
1406 }
1409 Compile::AliasType* Compile::alias_type(ciField* field) {
1410 const TypeOopPtr* t;
1411 if (field->is_static())
1412 t = TypeKlassPtr::make(field->holder());
1413 else
1414 t = TypeOopPtr::make_from_klass_raw(field->holder());
1415 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1416 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1417 return atp;
1418 }
1421 //------------------------------have_alias_type--------------------------------
1422 bool Compile::have_alias_type(const TypePtr* adr_type) {
1423 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1424 if (ace->_adr_type == adr_type) {
1425 return true;
1426 }
1428 // Handle special cases.
1429 if (adr_type == NULL) return true;
1430 if (adr_type == TypePtr::BOTTOM) return true;
1432 return find_alias_type(adr_type, true) != NULL;
1433 }
1435 //-----------------------------must_alias--------------------------------------
1436 // True if all values of the given address type are in the given alias category.
1437 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1438 if (alias_idx == AliasIdxBot) return true; // the universal category
1439 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1440 if (alias_idx == AliasIdxTop) return false; // the empty category
1441 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1443 // the only remaining possible overlap is identity
1444 int adr_idx = get_alias_index(adr_type);
1445 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1446 assert(adr_idx == alias_idx ||
1447 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1448 && adr_type != TypeOopPtr::BOTTOM),
1449 "should not be testing for overlap with an unsafe pointer");
1450 return adr_idx == alias_idx;
1451 }
1453 //------------------------------can_alias--------------------------------------
1454 // True if any values of the given address type are in the given alias category.
1455 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1456 if (alias_idx == AliasIdxTop) return false; // the empty category
1457 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1458 if (alias_idx == AliasIdxBot) return true; // the universal category
1459 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1461 // the only remaining possible overlap is identity
1462 int adr_idx = get_alias_index(adr_type);
1463 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1464 return adr_idx == alias_idx;
1465 }
1469 //---------------------------pop_warm_call-------------------------------------
1470 WarmCallInfo* Compile::pop_warm_call() {
1471 WarmCallInfo* wci = _warm_calls;
1472 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1473 return wci;
1474 }
1476 //----------------------------Inline_Warm--------------------------------------
1477 int Compile::Inline_Warm() {
1478 // If there is room, try to inline some more warm call sites.
1479 // %%% Do a graph index compaction pass when we think we're out of space?
1480 if (!InlineWarmCalls) return 0;
1482 int calls_made_hot = 0;
1483 int room_to_grow = NodeCountInliningCutoff - unique();
1484 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1485 int amount_grown = 0;
1486 WarmCallInfo* call;
1487 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1488 int est_size = (int)call->size();
1489 if (est_size > (room_to_grow - amount_grown)) {
1490 // This one won't fit anyway. Get rid of it.
1491 call->make_cold();
1492 continue;
1493 }
1494 call->make_hot();
1495 calls_made_hot++;
1496 amount_grown += est_size;
1497 amount_to_grow -= est_size;
1498 }
1500 if (calls_made_hot > 0) set_major_progress();
1501 return calls_made_hot;
1502 }
1505 //----------------------------Finish_Warm--------------------------------------
1506 void Compile::Finish_Warm() {
1507 if (!InlineWarmCalls) return;
1508 if (failing()) return;
1509 if (warm_calls() == NULL) return;
1511 // Clean up loose ends, if we are out of space for inlining.
1512 WarmCallInfo* call;
1513 while ((call = pop_warm_call()) != NULL) {
1514 call->make_cold();
1515 }
1516 }
1518 //---------------------cleanup_loop_predicates-----------------------
1519 // Remove the opaque nodes that protect the predicates so that all unused
1520 // checks and uncommon_traps will be eliminated from the ideal graph
1521 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1522 if (predicate_count()==0) return;
1523 for (int i = predicate_count(); i > 0; i--) {
1524 Node * n = predicate_opaque1_node(i-1);
1525 assert(n->Opcode() == Op_Opaque1, "must be");
1526 igvn.replace_node(n, n->in(1));
1527 }
1528 assert(predicate_count()==0, "should be clean!");
1529 igvn.optimize();
1530 }
1532 //------------------------------Optimize---------------------------------------
1533 // Given a graph, optimize it.
1534 void Compile::Optimize() {
1535 TracePhase t1("optimizer", &_t_optimizer, true);
1537 #ifndef PRODUCT
1538 if (env()->break_at_compile()) {
1539 BREAKPOINT;
1540 }
1542 #endif
1544 ResourceMark rm;
1545 int loop_opts_cnt;
1547 NOT_PRODUCT( verify_graph_edges(); )
1549 print_method("After Parsing");
1551 {
1552 // Iterative Global Value Numbering, including ideal transforms
1553 // Initialize IterGVN with types and values from parse-time GVN
1554 PhaseIterGVN igvn(initial_gvn());
1555 {
1556 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1557 igvn.optimize();
1558 }
1560 print_method("Iter GVN 1", 2);
1562 if (failing()) return;
1564 // Perform escape analysis
1565 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1566 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
1567 ConnectionGraph::do_analysis(this, &igvn);
1569 if (failing()) return;
1571 igvn.optimize();
1572 print_method("Iter GVN 3", 2);
1574 if (failing()) return;
1576 }
1578 // Loop transforms on the ideal graph. Range Check Elimination,
1579 // peeling, unrolling, etc.
1581 // Set loop opts counter
1582 loop_opts_cnt = num_loop_opts();
1583 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1584 {
1585 TracePhase t2("idealLoop", &_t_idealLoop, true);
1586 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate);
1587 loop_opts_cnt--;
1588 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1589 if (failing()) return;
1590 }
1591 // Loop opts pass if partial peeling occurred in previous pass
1592 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1593 TracePhase t3("idealLoop", &_t_idealLoop, true);
1594 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
1595 loop_opts_cnt--;
1596 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1597 if (failing()) return;
1598 }
1599 // Loop opts pass for loop-unrolling before CCP
1600 if(major_progress() && (loop_opts_cnt > 0)) {
1601 TracePhase t4("idealLoop", &_t_idealLoop, true);
1602 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
1603 loop_opts_cnt--;
1604 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1605 }
1606 if (!failing()) {
1607 // Verify that last round of loop opts produced a valid graph
1608 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1609 PhaseIdealLoop::verify(igvn);
1610 }
1611 }
1612 if (failing()) return;
1614 // Conditional Constant Propagation;
1615 PhaseCCP ccp( &igvn );
1616 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1617 {
1618 TracePhase t2("ccp", &_t_ccp, true);
1619 ccp.do_transform();
1620 }
1621 print_method("PhaseCPP 1", 2);
1623 assert( true, "Break here to ccp.dump_old2new_map()");
1625 // Iterative Global Value Numbering, including ideal transforms
1626 {
1627 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1628 igvn = ccp;
1629 igvn.optimize();
1630 }
1632 print_method("Iter GVN 2", 2);
1634 if (failing()) return;
1636 // Loop transforms on the ideal graph. Range Check Elimination,
1637 // peeling, unrolling, etc.
1638 if(loop_opts_cnt > 0) {
1639 debug_only( int cnt = 0; );
1640 bool loop_predication = UseLoopPredicate;
1641 while(major_progress() && (loop_opts_cnt > 0)) {
1642 TracePhase t2("idealLoop", &_t_idealLoop, true);
1643 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1644 PhaseIdealLoop ideal_loop( igvn, true, loop_predication);
1645 loop_opts_cnt--;
1646 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1647 if (failing()) return;
1648 // Perform loop predication optimization during first iteration after CCP.
1649 // After that switch it off and cleanup unused loop predicates.
1650 if (loop_predication) {
1651 loop_predication = false;
1652 cleanup_loop_predicates(igvn);
1653 if (failing()) return;
1654 }
1655 }
1656 }
1658 {
1659 // Verify that all previous optimizations produced a valid graph
1660 // at least to this point, even if no loop optimizations were done.
1661 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1662 PhaseIdealLoop::verify(igvn);
1663 }
1665 {
1666 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1667 PhaseMacroExpand mex(igvn);
1668 if (mex.expand_macro_nodes()) {
1669 assert(failing(), "must bail out w/ explicit message");
1670 return;
1671 }
1672 }
1674 } // (End scope of igvn; run destructor if necessary for asserts.)
1676 // A method with only infinite loops has no edges entering loops from root
1677 {
1678 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1679 if (final_graph_reshaping()) {
1680 assert(failing(), "must bail out w/ explicit message");
1681 return;
1682 }
1683 }
1685 print_method("Optimize finished", 2);
1686 }
1689 //------------------------------Code_Gen---------------------------------------
1690 // Given a graph, generate code for it
1691 void Compile::Code_Gen() {
1692 if (failing()) return;
1694 // Perform instruction selection. You might think we could reclaim Matcher
1695 // memory PDQ, but actually the Matcher is used in generating spill code.
1696 // Internals of the Matcher (including some VectorSets) must remain live
1697 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1698 // set a bit in reclaimed memory.
1700 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1701 // nodes. Mapping is only valid at the root of each matched subtree.
1702 NOT_PRODUCT( verify_graph_edges(); )
1704 Node_List proj_list;
1705 Matcher m(proj_list);
1706 _matcher = &m;
1707 {
1708 TracePhase t2("matcher", &_t_matcher, true);
1709 m.match();
1710 }
1711 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1712 // nodes. Mapping is only valid at the root of each matched subtree.
1713 NOT_PRODUCT( verify_graph_edges(); )
1715 // If you have too many nodes, or if matching has failed, bail out
1716 check_node_count(0, "out of nodes matching instructions");
1717 if (failing()) return;
1719 // Build a proper-looking CFG
1720 PhaseCFG cfg(node_arena(), root(), m);
1721 _cfg = &cfg;
1722 {
1723 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1724 cfg.Dominators();
1725 if (failing()) return;
1727 NOT_PRODUCT( verify_graph_edges(); )
1729 cfg.Estimate_Block_Frequency();
1730 cfg.GlobalCodeMotion(m,unique(),proj_list);
1732 print_method("Global code motion", 2);
1734 if (failing()) return;
1735 NOT_PRODUCT( verify_graph_edges(); )
1737 debug_only( cfg.verify(); )
1738 }
1739 NOT_PRODUCT( verify_graph_edges(); )
1741 PhaseChaitin regalloc(unique(),cfg,m);
1742 _regalloc = ®alloc;
1743 {
1744 TracePhase t2("regalloc", &_t_registerAllocation, true);
1745 // Perform any platform dependent preallocation actions. This is used,
1746 // for example, to avoid taking an implicit null pointer exception
1747 // using the frame pointer on win95.
1748 _regalloc->pd_preallocate_hook();
1750 // Perform register allocation. After Chaitin, use-def chains are
1751 // no longer accurate (at spill code) and so must be ignored.
1752 // Node->LRG->reg mappings are still accurate.
1753 _regalloc->Register_Allocate();
1755 // Bail out if the allocator builds too many nodes
1756 if (failing()) return;
1757 }
1759 // Prior to register allocation we kept empty basic blocks in case the
1760 // the allocator needed a place to spill. After register allocation we
1761 // are not adding any new instructions. If any basic block is empty, we
1762 // can now safely remove it.
1763 {
1764 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1765 cfg.remove_empty();
1766 if (do_freq_based_layout()) {
1767 PhaseBlockLayout layout(cfg);
1768 } else {
1769 cfg.set_loop_alignment();
1770 }
1771 cfg.fixup_flow();
1772 }
1774 // Perform any platform dependent postallocation verifications.
1775 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1777 // Apply peephole optimizations
1778 if( OptoPeephole ) {
1779 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1780 PhasePeephole peep( _regalloc, cfg);
1781 peep.do_transform();
1782 }
1784 // Convert Nodes to instruction bits in a buffer
1785 {
1786 // %%%% workspace merge brought two timers together for one job
1787 TracePhase t2a("output", &_t_output, true);
1788 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1789 Output();
1790 }
1792 print_method("Final Code");
1794 // He's dead, Jim.
1795 _cfg = (PhaseCFG*)0xdeadbeef;
1796 _regalloc = (PhaseChaitin*)0xdeadbeef;
1797 }
1800 //------------------------------dump_asm---------------------------------------
1801 // Dump formatted assembly
1802 #ifndef PRODUCT
1803 void Compile::dump_asm(int *pcs, uint pc_limit) {
1804 bool cut_short = false;
1805 tty->print_cr("#");
1806 tty->print("# "); _tf->dump(); tty->cr();
1807 tty->print_cr("#");
1809 // For all blocks
1810 int pc = 0x0; // Program counter
1811 char starts_bundle = ' ';
1812 _regalloc->dump_frame();
1814 Node *n = NULL;
1815 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1816 if (VMThread::should_terminate()) { cut_short = true; break; }
1817 Block *b = _cfg->_blocks[i];
1818 if (b->is_connector() && !Verbose) continue;
1819 n = b->_nodes[0];
1820 if (pcs && n->_idx < pc_limit)
1821 tty->print("%3.3x ", pcs[n->_idx]);
1822 else
1823 tty->print(" ");
1824 b->dump_head( &_cfg->_bbs );
1825 if (b->is_connector()) {
1826 tty->print_cr(" # Empty connector block");
1827 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1828 tty->print_cr(" # Block is sole successor of call");
1829 }
1831 // For all instructions
1832 Node *delay = NULL;
1833 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1834 if (VMThread::should_terminate()) { cut_short = true; break; }
1835 n = b->_nodes[j];
1836 if (valid_bundle_info(n)) {
1837 Bundle *bundle = node_bundling(n);
1838 if (bundle->used_in_unconditional_delay()) {
1839 delay = n;
1840 continue;
1841 }
1842 if (bundle->starts_bundle())
1843 starts_bundle = '+';
1844 }
1846 if (WizardMode) n->dump();
1848 if( !n->is_Region() && // Dont print in the Assembly
1849 !n->is_Phi() && // a few noisely useless nodes
1850 !n->is_Proj() &&
1851 !n->is_MachTemp() &&
1852 !n->is_SafePointScalarObject() &&
1853 !n->is_Catch() && // Would be nice to print exception table targets
1854 !n->is_MergeMem() && // Not very interesting
1855 !n->is_top() && // Debug info table constants
1856 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1857 ) {
1858 if (pcs && n->_idx < pc_limit)
1859 tty->print("%3.3x", pcs[n->_idx]);
1860 else
1861 tty->print(" ");
1862 tty->print(" %c ", starts_bundle);
1863 starts_bundle = ' ';
1864 tty->print("\t");
1865 n->format(_regalloc, tty);
1866 tty->cr();
1867 }
1869 // If we have an instruction with a delay slot, and have seen a delay,
1870 // then back up and print it
1871 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1872 assert(delay != NULL, "no unconditional delay instruction");
1873 if (WizardMode) delay->dump();
1875 if (node_bundling(delay)->starts_bundle())
1876 starts_bundle = '+';
1877 if (pcs && n->_idx < pc_limit)
1878 tty->print("%3.3x", pcs[n->_idx]);
1879 else
1880 tty->print(" ");
1881 tty->print(" %c ", starts_bundle);
1882 starts_bundle = ' ';
1883 tty->print("\t");
1884 delay->format(_regalloc, tty);
1885 tty->print_cr("");
1886 delay = NULL;
1887 }
1889 // Dump the exception table as well
1890 if( n->is_Catch() && (Verbose || WizardMode) ) {
1891 // Print the exception table for this offset
1892 _handler_table.print_subtable_for(pc);
1893 }
1894 }
1896 if (pcs && n->_idx < pc_limit)
1897 tty->print_cr("%3.3x", pcs[n->_idx]);
1898 else
1899 tty->print_cr("");
1901 assert(cut_short || delay == NULL, "no unconditional delay branch");
1903 } // End of per-block dump
1904 tty->print_cr("");
1906 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1907 }
1908 #endif
1910 //------------------------------Final_Reshape_Counts---------------------------
1911 // This class defines counters to help identify when a method
1912 // may/must be executed using hardware with only 24-bit precision.
1913 struct Final_Reshape_Counts : public StackObj {
1914 int _call_count; // count non-inlined 'common' calls
1915 int _float_count; // count float ops requiring 24-bit precision
1916 int _double_count; // count double ops requiring more precision
1917 int _java_call_count; // count non-inlined 'java' calls
1918 int _inner_loop_count; // count loops which need alignment
1919 VectorSet _visited; // Visitation flags
1920 Node_List _tests; // Set of IfNodes & PCTableNodes
1922 Final_Reshape_Counts() :
1923 _call_count(0), _float_count(0), _double_count(0),
1924 _java_call_count(0), _inner_loop_count(0),
1925 _visited( Thread::current()->resource_area() ) { }
1927 void inc_call_count () { _call_count ++; }
1928 void inc_float_count () { _float_count ++; }
1929 void inc_double_count() { _double_count++; }
1930 void inc_java_call_count() { _java_call_count++; }
1931 void inc_inner_loop_count() { _inner_loop_count++; }
1933 int get_call_count () const { return _call_count ; }
1934 int get_float_count () const { return _float_count ; }
1935 int get_double_count() const { return _double_count; }
1936 int get_java_call_count() const { return _java_call_count; }
1937 int get_inner_loop_count() const { return _inner_loop_count; }
1938 };
1940 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1941 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1942 // Make sure the offset goes inside the instance layout.
1943 return k->contains_field_offset(tp->offset());
1944 // Note that OffsetBot and OffsetTop are very negative.
1945 }
1947 //------------------------------final_graph_reshaping_impl----------------------
1948 // Implement items 1-5 from final_graph_reshaping below.
1949 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
1951 if ( n->outcnt() == 0 ) return; // dead node
1952 uint nop = n->Opcode();
1954 // Check for 2-input instruction with "last use" on right input.
1955 // Swap to left input. Implements item (2).
1956 if( n->req() == 3 && // two-input instruction
1957 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1958 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1959 n->in(2)->outcnt() == 1 &&// right use IS a last use
1960 !n->in(2)->is_Con() ) { // right use is not a constant
1961 // Check for commutative opcode
1962 switch( nop ) {
1963 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1964 case Op_MaxI: case Op_MinI:
1965 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1966 case Op_AndL: case Op_XorL: case Op_OrL:
1967 case Op_AndI: case Op_XorI: case Op_OrI: {
1968 // Move "last use" input to left by swapping inputs
1969 n->swap_edges(1, 2);
1970 break;
1971 }
1972 default:
1973 break;
1974 }
1975 }
1977 #ifdef ASSERT
1978 if( n->is_Mem() ) {
1979 Compile* C = Compile::current();
1980 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
1981 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
1982 // oop will be recorded in oop map if load crosses safepoint
1983 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
1984 LoadNode::is_immutable_value(n->in(MemNode::Address))),
1985 "raw memory operations should have control edge");
1986 }
1987 #endif
1988 // Count FPU ops and common calls, implements item (3)
1989 switch( nop ) {
1990 // Count all float operations that may use FPU
1991 case Op_AddF:
1992 case Op_SubF:
1993 case Op_MulF:
1994 case Op_DivF:
1995 case Op_NegF:
1996 case Op_ModF:
1997 case Op_ConvI2F:
1998 case Op_ConF:
1999 case Op_CmpF:
2000 case Op_CmpF3:
2001 // case Op_ConvL2F: // longs are split into 32-bit halves
2002 frc.inc_float_count();
2003 break;
2005 case Op_ConvF2D:
2006 case Op_ConvD2F:
2007 frc.inc_float_count();
2008 frc.inc_double_count();
2009 break;
2011 // Count all double operations that may use FPU
2012 case Op_AddD:
2013 case Op_SubD:
2014 case Op_MulD:
2015 case Op_DivD:
2016 case Op_NegD:
2017 case Op_ModD:
2018 case Op_ConvI2D:
2019 case Op_ConvD2I:
2020 // case Op_ConvL2D: // handled by leaf call
2021 // case Op_ConvD2L: // handled by leaf call
2022 case Op_ConD:
2023 case Op_CmpD:
2024 case Op_CmpD3:
2025 frc.inc_double_count();
2026 break;
2027 case Op_Opaque1: // Remove Opaque Nodes before matching
2028 case Op_Opaque2: // Remove Opaque Nodes before matching
2029 n->subsume_by(n->in(1));
2030 break;
2031 case Op_CallStaticJava:
2032 case Op_CallJava:
2033 case Op_CallDynamicJava:
2034 frc.inc_java_call_count(); // Count java call site;
2035 case Op_CallRuntime:
2036 case Op_CallLeaf:
2037 case Op_CallLeafNoFP: {
2038 assert( n->is_Call(), "" );
2039 CallNode *call = n->as_Call();
2040 // Count call sites where the FP mode bit would have to be flipped.
2041 // Do not count uncommon runtime calls:
2042 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2043 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2044 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2045 frc.inc_call_count(); // Count the call site
2046 } else { // See if uncommon argument is shared
2047 Node *n = call->in(TypeFunc::Parms);
2048 int nop = n->Opcode();
2049 // Clone shared simple arguments to uncommon calls, item (1).
2050 if( n->outcnt() > 1 &&
2051 !n->is_Proj() &&
2052 nop != Op_CreateEx &&
2053 nop != Op_CheckCastPP &&
2054 nop != Op_DecodeN &&
2055 !n->is_Mem() ) {
2056 Node *x = n->clone();
2057 call->set_req( TypeFunc::Parms, x );
2058 }
2059 }
2060 break;
2061 }
2063 case Op_StoreD:
2064 case Op_LoadD:
2065 case Op_LoadD_unaligned:
2066 frc.inc_double_count();
2067 goto handle_mem;
2068 case Op_StoreF:
2069 case Op_LoadF:
2070 frc.inc_float_count();
2071 goto handle_mem;
2073 case Op_StoreB:
2074 case Op_StoreC:
2075 case Op_StoreCM:
2076 case Op_StorePConditional:
2077 case Op_StoreI:
2078 case Op_StoreL:
2079 case Op_StoreIConditional:
2080 case Op_StoreLConditional:
2081 case Op_CompareAndSwapI:
2082 case Op_CompareAndSwapL:
2083 case Op_CompareAndSwapP:
2084 case Op_CompareAndSwapN:
2085 case Op_StoreP:
2086 case Op_StoreN:
2087 case Op_LoadB:
2088 case Op_LoadUB:
2089 case Op_LoadUS:
2090 case Op_LoadI:
2091 case Op_LoadUI2L:
2092 case Op_LoadKlass:
2093 case Op_LoadNKlass:
2094 case Op_LoadL:
2095 case Op_LoadL_unaligned:
2096 case Op_LoadPLocked:
2097 case Op_LoadLLocked:
2098 case Op_LoadP:
2099 case Op_LoadN:
2100 case Op_LoadRange:
2101 case Op_LoadS: {
2102 handle_mem:
2103 #ifdef ASSERT
2104 if( VerifyOptoOopOffsets ) {
2105 assert( n->is_Mem(), "" );
2106 MemNode *mem = (MemNode*)n;
2107 // Check to see if address types have grounded out somehow.
2108 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2109 assert( !tp || oop_offset_is_sane(tp), "" );
2110 }
2111 #endif
2112 break;
2113 }
2115 case Op_AddP: { // Assert sane base pointers
2116 Node *addp = n->in(AddPNode::Address);
2117 assert( !addp->is_AddP() ||
2118 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2119 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2120 "Base pointers must match" );
2121 #ifdef _LP64
2122 if (UseCompressedOops &&
2123 addp->Opcode() == Op_ConP &&
2124 addp == n->in(AddPNode::Base) &&
2125 n->in(AddPNode::Offset)->is_Con()) {
2126 // Use addressing with narrow klass to load with offset on x86.
2127 // On sparc loading 32-bits constant and decoding it have less
2128 // instructions (4) then load 64-bits constant (7).
2129 // Do this transformation here since IGVN will convert ConN back to ConP.
2130 const Type* t = addp->bottom_type();
2131 if (t->isa_oopptr()) {
2132 Node* nn = NULL;
2134 // Look for existing ConN node of the same exact type.
2135 Compile* C = Compile::current();
2136 Node* r = C->root();
2137 uint cnt = r->outcnt();
2138 for (uint i = 0; i < cnt; i++) {
2139 Node* m = r->raw_out(i);
2140 if (m!= NULL && m->Opcode() == Op_ConN &&
2141 m->bottom_type()->make_ptr() == t) {
2142 nn = m;
2143 break;
2144 }
2145 }
2146 if (nn != NULL) {
2147 // Decode a narrow oop to match address
2148 // [R12 + narrow_oop_reg<<3 + offset]
2149 nn = new (C, 2) DecodeNNode(nn, t);
2150 n->set_req(AddPNode::Base, nn);
2151 n->set_req(AddPNode::Address, nn);
2152 if (addp->outcnt() == 0) {
2153 addp->disconnect_inputs(NULL);
2154 }
2155 }
2156 }
2157 }
2158 #endif
2159 break;
2160 }
2162 #ifdef _LP64
2163 case Op_CastPP:
2164 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2165 Compile* C = Compile::current();
2166 Node* in1 = n->in(1);
2167 const Type* t = n->bottom_type();
2168 Node* new_in1 = in1->clone();
2169 new_in1->as_DecodeN()->set_type(t);
2171 if (!Matcher::narrow_oop_use_complex_address()) {
2172 //
2173 // x86, ARM and friends can handle 2 adds in addressing mode
2174 // and Matcher can fold a DecodeN node into address by using
2175 // a narrow oop directly and do implicit NULL check in address:
2176 //
2177 // [R12 + narrow_oop_reg<<3 + offset]
2178 // NullCheck narrow_oop_reg
2179 //
2180 // On other platforms (Sparc) we have to keep new DecodeN node and
2181 // use it to do implicit NULL check in address:
2182 //
2183 // decode_not_null narrow_oop_reg, base_reg
2184 // [base_reg + offset]
2185 // NullCheck base_reg
2186 //
2187 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2188 // to keep the information to which NULL check the new DecodeN node
2189 // corresponds to use it as value in implicit_null_check().
2190 //
2191 new_in1->set_req(0, n->in(0));
2192 }
2194 n->subsume_by(new_in1);
2195 if (in1->outcnt() == 0) {
2196 in1->disconnect_inputs(NULL);
2197 }
2198 }
2199 break;
2201 case Op_CmpP:
2202 // Do this transformation here to preserve CmpPNode::sub() and
2203 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2204 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2205 Node* in1 = n->in(1);
2206 Node* in2 = n->in(2);
2207 if (!in1->is_DecodeN()) {
2208 in2 = in1;
2209 in1 = n->in(2);
2210 }
2211 assert(in1->is_DecodeN(), "sanity");
2213 Compile* C = Compile::current();
2214 Node* new_in2 = NULL;
2215 if (in2->is_DecodeN()) {
2216 new_in2 = in2->in(1);
2217 } else if (in2->Opcode() == Op_ConP) {
2218 const Type* t = in2->bottom_type();
2219 if (t == TypePtr::NULL_PTR) {
2220 // Don't convert CmpP null check into CmpN if compressed
2221 // oops implicit null check is not generated.
2222 // This will allow to generate normal oop implicit null check.
2223 if (Matcher::gen_narrow_oop_implicit_null_checks())
2224 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2225 //
2226 // This transformation together with CastPP transformation above
2227 // will generated code for implicit NULL checks for compressed oops.
2228 //
2229 // The original code after Optimize()
2230 //
2231 // LoadN memory, narrow_oop_reg
2232 // decode narrow_oop_reg, base_reg
2233 // CmpP base_reg, NULL
2234 // CastPP base_reg // NotNull
2235 // Load [base_reg + offset], val_reg
2236 //
2237 // after these transformations will be
2238 //
2239 // LoadN memory, narrow_oop_reg
2240 // CmpN narrow_oop_reg, NULL
2241 // decode_not_null narrow_oop_reg, base_reg
2242 // Load [base_reg + offset], val_reg
2243 //
2244 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2245 // since narrow oops can be used in debug info now (see the code in
2246 // final_graph_reshaping_walk()).
2247 //
2248 // At the end the code will be matched to
2249 // on x86:
2250 //
2251 // Load_narrow_oop memory, narrow_oop_reg
2252 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2253 // NullCheck narrow_oop_reg
2254 //
2255 // and on sparc:
2256 //
2257 // Load_narrow_oop memory, narrow_oop_reg
2258 // decode_not_null narrow_oop_reg, base_reg
2259 // Load [base_reg + offset], val_reg
2260 // NullCheck base_reg
2261 //
2262 } else if (t->isa_oopptr()) {
2263 new_in2 = ConNode::make(C, t->make_narrowoop());
2264 }
2265 }
2266 if (new_in2 != NULL) {
2267 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2268 n->subsume_by( cmpN );
2269 if (in1->outcnt() == 0) {
2270 in1->disconnect_inputs(NULL);
2271 }
2272 if (in2->outcnt() == 0) {
2273 in2->disconnect_inputs(NULL);
2274 }
2275 }
2276 }
2277 break;
2279 case Op_DecodeN:
2280 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2281 // DecodeN could be pinned when it can't be fold into
2282 // an address expression, see the code for Op_CastPP above.
2283 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2284 break;
2286 case Op_EncodeP: {
2287 Node* in1 = n->in(1);
2288 if (in1->is_DecodeN()) {
2289 n->subsume_by(in1->in(1));
2290 } else if (in1->Opcode() == Op_ConP) {
2291 Compile* C = Compile::current();
2292 const Type* t = in1->bottom_type();
2293 if (t == TypePtr::NULL_PTR) {
2294 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2295 } else if (t->isa_oopptr()) {
2296 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2297 }
2298 }
2299 if (in1->outcnt() == 0) {
2300 in1->disconnect_inputs(NULL);
2301 }
2302 break;
2303 }
2305 case Op_Proj: {
2306 if (OptimizeStringConcat) {
2307 ProjNode* p = n->as_Proj();
2308 if (p->_is_io_use) {
2309 // Separate projections were used for the exception path which
2310 // are normally removed by a late inline. If it wasn't inlined
2311 // then they will hang around and should just be replaced with
2312 // the original one.
2313 Node* proj = NULL;
2314 // Replace with just one
2315 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2316 Node *use = i.get();
2317 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2318 proj = use;
2319 break;
2320 }
2321 }
2322 assert(p != NULL, "must be found");
2323 p->subsume_by(proj);
2324 }
2325 }
2326 break;
2327 }
2329 case Op_Phi:
2330 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2331 // The EncodeP optimization may create Phi with the same edges
2332 // for all paths. It is not handled well by Register Allocator.
2333 Node* unique_in = n->in(1);
2334 assert(unique_in != NULL, "");
2335 uint cnt = n->req();
2336 for (uint i = 2; i < cnt; i++) {
2337 Node* m = n->in(i);
2338 assert(m != NULL, "");
2339 if (unique_in != m)
2340 unique_in = NULL;
2341 }
2342 if (unique_in != NULL) {
2343 n->subsume_by(unique_in);
2344 }
2345 }
2346 break;
2348 #endif
2350 case Op_ModI:
2351 if (UseDivMod) {
2352 // Check if a%b and a/b both exist
2353 Node* d = n->find_similar(Op_DivI);
2354 if (d) {
2355 // Replace them with a fused divmod if supported
2356 Compile* C = Compile::current();
2357 if (Matcher::has_match_rule(Op_DivModI)) {
2358 DivModINode* divmod = DivModINode::make(C, n);
2359 d->subsume_by(divmod->div_proj());
2360 n->subsume_by(divmod->mod_proj());
2361 } else {
2362 // replace a%b with a-((a/b)*b)
2363 Node* mult = new (C, 3) MulINode(d, d->in(2));
2364 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2365 n->subsume_by( sub );
2366 }
2367 }
2368 }
2369 break;
2371 case Op_ModL:
2372 if (UseDivMod) {
2373 // Check if a%b and a/b both exist
2374 Node* d = n->find_similar(Op_DivL);
2375 if (d) {
2376 // Replace them with a fused divmod if supported
2377 Compile* C = Compile::current();
2378 if (Matcher::has_match_rule(Op_DivModL)) {
2379 DivModLNode* divmod = DivModLNode::make(C, n);
2380 d->subsume_by(divmod->div_proj());
2381 n->subsume_by(divmod->mod_proj());
2382 } else {
2383 // replace a%b with a-((a/b)*b)
2384 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2385 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2386 n->subsume_by( sub );
2387 }
2388 }
2389 }
2390 break;
2392 case Op_Load16B:
2393 case Op_Load8B:
2394 case Op_Load4B:
2395 case Op_Load8S:
2396 case Op_Load4S:
2397 case Op_Load2S:
2398 case Op_Load8C:
2399 case Op_Load4C:
2400 case Op_Load2C:
2401 case Op_Load4I:
2402 case Op_Load2I:
2403 case Op_Load2L:
2404 case Op_Load4F:
2405 case Op_Load2F:
2406 case Op_Load2D:
2407 case Op_Store16B:
2408 case Op_Store8B:
2409 case Op_Store4B:
2410 case Op_Store8C:
2411 case Op_Store4C:
2412 case Op_Store2C:
2413 case Op_Store4I:
2414 case Op_Store2I:
2415 case Op_Store2L:
2416 case Op_Store4F:
2417 case Op_Store2F:
2418 case Op_Store2D:
2419 break;
2421 case Op_PackB:
2422 case Op_PackS:
2423 case Op_PackC:
2424 case Op_PackI:
2425 case Op_PackF:
2426 case Op_PackL:
2427 case Op_PackD:
2428 if (n->req()-1 > 2) {
2429 // Replace many operand PackNodes with a binary tree for matching
2430 PackNode* p = (PackNode*) n;
2431 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2432 n->subsume_by(btp);
2433 }
2434 break;
2435 case Op_Loop:
2436 case Op_CountedLoop:
2437 if (n->as_Loop()->is_inner_loop()) {
2438 frc.inc_inner_loop_count();
2439 }
2440 break;
2441 default:
2442 assert( !n->is_Call(), "" );
2443 assert( !n->is_Mem(), "" );
2444 break;
2445 }
2447 // Collect CFG split points
2448 if (n->is_MultiBranch())
2449 frc._tests.push(n);
2450 }
2452 //------------------------------final_graph_reshaping_walk---------------------
2453 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2454 // requires that the walk visits a node's inputs before visiting the node.
2455 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2456 ResourceArea *area = Thread::current()->resource_area();
2457 Unique_Node_List sfpt(area);
2459 frc._visited.set(root->_idx); // first, mark node as visited
2460 uint cnt = root->req();
2461 Node *n = root;
2462 uint i = 0;
2463 while (true) {
2464 if (i < cnt) {
2465 // Place all non-visited non-null inputs onto stack
2466 Node* m = n->in(i);
2467 ++i;
2468 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2469 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2470 sfpt.push(m);
2471 cnt = m->req();
2472 nstack.push(n, i); // put on stack parent and next input's index
2473 n = m;
2474 i = 0;
2475 }
2476 } else {
2477 // Now do post-visit work
2478 final_graph_reshaping_impl( n, frc );
2479 if (nstack.is_empty())
2480 break; // finished
2481 n = nstack.node(); // Get node from stack
2482 cnt = n->req();
2483 i = nstack.index();
2484 nstack.pop(); // Shift to the next node on stack
2485 }
2486 }
2488 // Skip next transformation if compressed oops are not used.
2489 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2490 return;
2492 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2493 // It could be done for an uncommon traps or any safepoints/calls
2494 // if the DecodeN node is referenced only in a debug info.
2495 while (sfpt.size() > 0) {
2496 n = sfpt.pop();
2497 JVMState *jvms = n->as_SafePoint()->jvms();
2498 assert(jvms != NULL, "sanity");
2499 int start = jvms->debug_start();
2500 int end = n->req();
2501 bool is_uncommon = (n->is_CallStaticJava() &&
2502 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2503 for (int j = start; j < end; j++) {
2504 Node* in = n->in(j);
2505 if (in->is_DecodeN()) {
2506 bool safe_to_skip = true;
2507 if (!is_uncommon ) {
2508 // Is it safe to skip?
2509 for (uint i = 0; i < in->outcnt(); i++) {
2510 Node* u = in->raw_out(i);
2511 if (!u->is_SafePoint() ||
2512 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2513 safe_to_skip = false;
2514 }
2515 }
2516 }
2517 if (safe_to_skip) {
2518 n->set_req(j, in->in(1));
2519 }
2520 if (in->outcnt() == 0) {
2521 in->disconnect_inputs(NULL);
2522 }
2523 }
2524 }
2525 }
2526 }
2528 //------------------------------final_graph_reshaping--------------------------
2529 // Final Graph Reshaping.
2530 //
2531 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2532 // and not commoned up and forced early. Must come after regular
2533 // optimizations to avoid GVN undoing the cloning. Clone constant
2534 // inputs to Loop Phis; these will be split by the allocator anyways.
2535 // Remove Opaque nodes.
2536 // (2) Move last-uses by commutative operations to the left input to encourage
2537 // Intel update-in-place two-address operations and better register usage
2538 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2539 // calls canonicalizing them back.
2540 // (3) Count the number of double-precision FP ops, single-precision FP ops
2541 // and call sites. On Intel, we can get correct rounding either by
2542 // forcing singles to memory (requires extra stores and loads after each
2543 // FP bytecode) or we can set a rounding mode bit (requires setting and
2544 // clearing the mode bit around call sites). The mode bit is only used
2545 // if the relative frequency of single FP ops to calls is low enough.
2546 // This is a key transform for SPEC mpeg_audio.
2547 // (4) Detect infinite loops; blobs of code reachable from above but not
2548 // below. Several of the Code_Gen algorithms fail on such code shapes,
2549 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2550 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2551 // Detection is by looking for IfNodes where only 1 projection is
2552 // reachable from below or CatchNodes missing some targets.
2553 // (5) Assert for insane oop offsets in debug mode.
2555 bool Compile::final_graph_reshaping() {
2556 // an infinite loop may have been eliminated by the optimizer,
2557 // in which case the graph will be empty.
2558 if (root()->req() == 1) {
2559 record_method_not_compilable("trivial infinite loop");
2560 return true;
2561 }
2563 Final_Reshape_Counts frc;
2565 // Visit everybody reachable!
2566 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2567 Node_Stack nstack(unique() >> 1);
2568 final_graph_reshaping_walk(nstack, root(), frc);
2570 // Check for unreachable (from below) code (i.e., infinite loops).
2571 for( uint i = 0; i < frc._tests.size(); i++ ) {
2572 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2573 // Get number of CFG targets.
2574 // Note that PCTables include exception targets after calls.
2575 uint required_outcnt = n->required_outcnt();
2576 if (n->outcnt() != required_outcnt) {
2577 // Check for a few special cases. Rethrow Nodes never take the
2578 // 'fall-thru' path, so expected kids is 1 less.
2579 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2580 if (n->in(0)->in(0)->is_Call()) {
2581 CallNode *call = n->in(0)->in(0)->as_Call();
2582 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2583 required_outcnt--; // Rethrow always has 1 less kid
2584 } else if (call->req() > TypeFunc::Parms &&
2585 call->is_CallDynamicJava()) {
2586 // Check for null receiver. In such case, the optimizer has
2587 // detected that the virtual call will always result in a null
2588 // pointer exception. The fall-through projection of this CatchNode
2589 // will not be populated.
2590 Node *arg0 = call->in(TypeFunc::Parms);
2591 if (arg0->is_Type() &&
2592 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2593 required_outcnt--;
2594 }
2595 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2596 call->req() > TypeFunc::Parms+1 &&
2597 call->is_CallStaticJava()) {
2598 // Check for negative array length. In such case, the optimizer has
2599 // detected that the allocation attempt will always result in an
2600 // exception. There is no fall-through projection of this CatchNode .
2601 Node *arg1 = call->in(TypeFunc::Parms+1);
2602 if (arg1->is_Type() &&
2603 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2604 required_outcnt--;
2605 }
2606 }
2607 }
2608 }
2609 // Recheck with a better notion of 'required_outcnt'
2610 if (n->outcnt() != required_outcnt) {
2611 record_method_not_compilable("malformed control flow");
2612 return true; // Not all targets reachable!
2613 }
2614 }
2615 // Check that I actually visited all kids. Unreached kids
2616 // must be infinite loops.
2617 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2618 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2619 record_method_not_compilable("infinite loop");
2620 return true; // Found unvisited kid; must be unreach
2621 }
2622 }
2624 // If original bytecodes contained a mixture of floats and doubles
2625 // check if the optimizer has made it homogenous, item (3).
2626 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2627 frc.get_float_count() > 32 &&
2628 frc.get_double_count() == 0 &&
2629 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2630 set_24_bit_selection_and_mode( false, true );
2631 }
2633 set_java_calls(frc.get_java_call_count());
2634 set_inner_loops(frc.get_inner_loop_count());
2636 // No infinite loops, no reason to bail out.
2637 return false;
2638 }
2640 //-----------------------------too_many_traps----------------------------------
2641 // Report if there are too many traps at the current method and bci.
2642 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2643 bool Compile::too_many_traps(ciMethod* method,
2644 int bci,
2645 Deoptimization::DeoptReason reason) {
2646 ciMethodData* md = method->method_data();
2647 if (md->is_empty()) {
2648 // Assume the trap has not occurred, or that it occurred only
2649 // because of a transient condition during start-up in the interpreter.
2650 return false;
2651 }
2652 if (md->has_trap_at(bci, reason) != 0) {
2653 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2654 // Also, if there are multiple reasons, or if there is no per-BCI record,
2655 // assume the worst.
2656 if (log())
2657 log()->elem("observe trap='%s' count='%d'",
2658 Deoptimization::trap_reason_name(reason),
2659 md->trap_count(reason));
2660 return true;
2661 } else {
2662 // Ignore method/bci and see if there have been too many globally.
2663 return too_many_traps(reason, md);
2664 }
2665 }
2667 // Less-accurate variant which does not require a method and bci.
2668 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2669 ciMethodData* logmd) {
2670 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2671 // Too many traps globally.
2672 // Note that we use cumulative trap_count, not just md->trap_count.
2673 if (log()) {
2674 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2675 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2676 Deoptimization::trap_reason_name(reason),
2677 mcount, trap_count(reason));
2678 }
2679 return true;
2680 } else {
2681 // The coast is clear.
2682 return false;
2683 }
2684 }
2686 //--------------------------too_many_recompiles--------------------------------
2687 // Report if there are too many recompiles at the current method and bci.
2688 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2689 // Is not eager to return true, since this will cause the compiler to use
2690 // Action_none for a trap point, to avoid too many recompilations.
2691 bool Compile::too_many_recompiles(ciMethod* method,
2692 int bci,
2693 Deoptimization::DeoptReason reason) {
2694 ciMethodData* md = method->method_data();
2695 if (md->is_empty()) {
2696 // Assume the trap has not occurred, or that it occurred only
2697 // because of a transient condition during start-up in the interpreter.
2698 return false;
2699 }
2700 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2701 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2702 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2703 Deoptimization::DeoptReason per_bc_reason
2704 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2705 if ((per_bc_reason == Deoptimization::Reason_none
2706 || md->has_trap_at(bci, reason) != 0)
2707 // The trap frequency measure we care about is the recompile count:
2708 && md->trap_recompiled_at(bci)
2709 && md->overflow_recompile_count() >= bc_cutoff) {
2710 // Do not emit a trap here if it has already caused recompilations.
2711 // Also, if there are multiple reasons, or if there is no per-BCI record,
2712 // assume the worst.
2713 if (log())
2714 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2715 Deoptimization::trap_reason_name(reason),
2716 md->trap_count(reason),
2717 md->overflow_recompile_count());
2718 return true;
2719 } else if (trap_count(reason) != 0
2720 && decompile_count() >= m_cutoff) {
2721 // Too many recompiles globally, and we have seen this sort of trap.
2722 // Use cumulative decompile_count, not just md->decompile_count.
2723 if (log())
2724 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2725 Deoptimization::trap_reason_name(reason),
2726 md->trap_count(reason), trap_count(reason),
2727 md->decompile_count(), decompile_count());
2728 return true;
2729 } else {
2730 // The coast is clear.
2731 return false;
2732 }
2733 }
2736 #ifndef PRODUCT
2737 //------------------------------verify_graph_edges---------------------------
2738 // Walk the Graph and verify that there is a one-to-one correspondence
2739 // between Use-Def edges and Def-Use edges in the graph.
2740 void Compile::verify_graph_edges(bool no_dead_code) {
2741 if (VerifyGraphEdges) {
2742 ResourceArea *area = Thread::current()->resource_area();
2743 Unique_Node_List visited(area);
2744 // Call recursive graph walk to check edges
2745 _root->verify_edges(visited);
2746 if (no_dead_code) {
2747 // Now make sure that no visited node is used by an unvisited node.
2748 bool dead_nodes = 0;
2749 Unique_Node_List checked(area);
2750 while (visited.size() > 0) {
2751 Node* n = visited.pop();
2752 checked.push(n);
2753 for (uint i = 0; i < n->outcnt(); i++) {
2754 Node* use = n->raw_out(i);
2755 if (checked.member(use)) continue; // already checked
2756 if (visited.member(use)) continue; // already in the graph
2757 if (use->is_Con()) continue; // a dead ConNode is OK
2758 // At this point, we have found a dead node which is DU-reachable.
2759 if (dead_nodes++ == 0)
2760 tty->print_cr("*** Dead nodes reachable via DU edges:");
2761 use->dump(2);
2762 tty->print_cr("---");
2763 checked.push(use); // No repeats; pretend it is now checked.
2764 }
2765 }
2766 assert(dead_nodes == 0, "using nodes must be reachable from root");
2767 }
2768 }
2769 }
2770 #endif
2772 // The Compile object keeps track of failure reasons separately from the ciEnv.
2773 // This is required because there is not quite a 1-1 relation between the
2774 // ciEnv and its compilation task and the Compile object. Note that one
2775 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2776 // to backtrack and retry without subsuming loads. Other than this backtracking
2777 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2778 // by the logic in C2Compiler.
2779 void Compile::record_failure(const char* reason) {
2780 if (log() != NULL) {
2781 log()->elem("failure reason='%s' phase='compile'", reason);
2782 }
2783 if (_failure_reason == NULL) {
2784 // Record the first failure reason.
2785 _failure_reason = reason;
2786 }
2787 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2788 C->print_method(_failure_reason);
2789 }
2790 _root = NULL; // flush the graph, too
2791 }
2793 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2794 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2795 {
2796 if (dolog) {
2797 C = Compile::current();
2798 _log = C->log();
2799 } else {
2800 C = NULL;
2801 _log = NULL;
2802 }
2803 if (_log != NULL) {
2804 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2805 _log->stamp();
2806 _log->end_head();
2807 }
2808 }
2810 Compile::TracePhase::~TracePhase() {
2811 if (_log != NULL) {
2812 _log->done("phase nodes='%d'", C->unique());
2813 }
2814 }