Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
1 /*
2 * Copyright 1997-2010 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_compile.cpp.incl"
28 /// Support for intrinsics.
30 // Return the index at which m must be inserted (or already exists).
31 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
33 #ifdef ASSERT
34 for (int i = 1; i < _intrinsics->length(); i++) {
35 CallGenerator* cg1 = _intrinsics->at(i-1);
36 CallGenerator* cg2 = _intrinsics->at(i);
37 assert(cg1->method() != cg2->method()
38 ? cg1->method() < cg2->method()
39 : cg1->is_virtual() < cg2->is_virtual(),
40 "compiler intrinsics list must stay sorted");
41 }
42 #endif
43 // Binary search sorted list, in decreasing intervals [lo, hi].
44 int lo = 0, hi = _intrinsics->length()-1;
45 while (lo <= hi) {
46 int mid = (uint)(hi + lo) / 2;
47 ciMethod* mid_m = _intrinsics->at(mid)->method();
48 if (m < mid_m) {
49 hi = mid-1;
50 } else if (m > mid_m) {
51 lo = mid+1;
52 } else {
53 // look at minor sort key
54 bool mid_virt = _intrinsics->at(mid)->is_virtual();
55 if (is_virtual < mid_virt) {
56 hi = mid-1;
57 } else if (is_virtual > mid_virt) {
58 lo = mid+1;
59 } else {
60 return mid; // exact match
61 }
62 }
63 }
64 return lo; // inexact match
65 }
67 void Compile::register_intrinsic(CallGenerator* cg) {
68 if (_intrinsics == NULL) {
69 _intrinsics = new GrowableArray<CallGenerator*>(60);
70 }
71 // This code is stolen from ciObjectFactory::insert.
72 // Really, GrowableArray should have methods for
73 // insert_at, remove_at, and binary_search.
74 int len = _intrinsics->length();
75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
76 if (index == len) {
77 _intrinsics->append(cg);
78 } else {
79 #ifdef ASSERT
80 CallGenerator* oldcg = _intrinsics->at(index);
81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
82 #endif
83 _intrinsics->append(_intrinsics->at(len-1));
84 int pos;
85 for (pos = len-2; pos >= index; pos--) {
86 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
87 }
88 _intrinsics->at_put(index, cg);
89 }
90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
91 }
93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
94 assert(m->is_loaded(), "don't try this on unloaded methods");
95 if (_intrinsics != NULL) {
96 int index = intrinsic_insertion_index(m, is_virtual);
97 if (index < _intrinsics->length()
98 && _intrinsics->at(index)->method() == m
99 && _intrinsics->at(index)->is_virtual() == is_virtual) {
100 return _intrinsics->at(index);
101 }
102 }
103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
104 if (m->intrinsic_id() != vmIntrinsics::_none &&
105 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
106 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
107 if (cg != NULL) {
108 // Save it for next time:
109 register_intrinsic(cg);
110 return cg;
111 } else {
112 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
113 }
114 }
115 return NULL;
116 }
118 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
119 // in library_call.cpp.
122 #ifndef PRODUCT
123 // statistics gathering...
125 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
126 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
128 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
129 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
130 int oflags = _intrinsic_hist_flags[id];
131 assert(flags != 0, "what happened?");
132 if (is_virtual) {
133 flags |= _intrinsic_virtual;
134 }
135 bool changed = (flags != oflags);
136 if ((flags & _intrinsic_worked) != 0) {
137 juint count = (_intrinsic_hist_count[id] += 1);
138 if (count == 1) {
139 changed = true; // first time
140 }
141 // increment the overall count also:
142 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
143 }
144 if (changed) {
145 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
146 // Something changed about the intrinsic's virtuality.
147 if ((flags & _intrinsic_virtual) != 0) {
148 // This is the first use of this intrinsic as a virtual call.
149 if (oflags != 0) {
150 // We already saw it as a non-virtual, so note both cases.
151 flags |= _intrinsic_both;
152 }
153 } else if ((oflags & _intrinsic_both) == 0) {
154 // This is the first use of this intrinsic as a non-virtual
155 flags |= _intrinsic_both;
156 }
157 }
158 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
159 }
160 // update the overall flags also:
161 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
162 return changed;
163 }
165 static char* format_flags(int flags, char* buf) {
166 buf[0] = 0;
167 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
168 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
169 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
170 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
171 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
172 if (buf[0] == 0) strcat(buf, ",");
173 assert(buf[0] == ',', "must be");
174 return &buf[1];
175 }
177 void Compile::print_intrinsic_statistics() {
178 char flagsbuf[100];
179 ttyLocker ttyl;
180 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
181 tty->print_cr("Compiler intrinsic usage:");
182 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
183 if (total == 0) total = 1; // avoid div0 in case of no successes
184 #define PRINT_STAT_LINE(name, c, f) \
185 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
186 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
187 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
188 int flags = _intrinsic_hist_flags[id];
189 juint count = _intrinsic_hist_count[id];
190 if ((flags | count) != 0) {
191 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
192 }
193 }
194 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
195 if (xtty != NULL) xtty->tail("statistics");
196 }
198 void Compile::print_statistics() {
199 { ttyLocker ttyl;
200 if (xtty != NULL) xtty->head("statistics type='opto'");
201 Parse::print_statistics();
202 PhaseCCP::print_statistics();
203 PhaseRegAlloc::print_statistics();
204 Scheduling::print_statistics();
205 PhasePeephole::print_statistics();
206 PhaseIdealLoop::print_statistics();
207 if (xtty != NULL) xtty->tail("statistics");
208 }
209 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
210 // put this under its own <statistics> element.
211 print_intrinsic_statistics();
212 }
213 }
214 #endif //PRODUCT
216 // Support for bundling info
217 Bundle* Compile::node_bundling(const Node *n) {
218 assert(valid_bundle_info(n), "oob");
219 return &_node_bundling_base[n->_idx];
220 }
222 bool Compile::valid_bundle_info(const Node *n) {
223 return (_node_bundling_limit > n->_idx);
224 }
227 void Compile::gvn_replace_by(Node* n, Node* nn) {
228 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
229 Node* use = n->last_out(i);
230 bool is_in_table = initial_gvn()->hash_delete(use);
231 uint uses_found = 0;
232 for (uint j = 0; j < use->len(); j++) {
233 if (use->in(j) == n) {
234 if (j < use->req())
235 use->set_req(j, nn);
236 else
237 use->set_prec(j, nn);
238 uses_found++;
239 }
240 }
241 if (is_in_table) {
242 // reinsert into table
243 initial_gvn()->hash_find_insert(use);
244 }
245 record_for_igvn(use);
246 i -= uses_found; // we deleted 1 or more copies of this edge
247 }
248 }
253 // Identify all nodes that are reachable from below, useful.
254 // Use breadth-first pass that records state in a Unique_Node_List,
255 // recursive traversal is slower.
256 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
257 int estimated_worklist_size = unique();
258 useful.map( estimated_worklist_size, NULL ); // preallocate space
260 // Initialize worklist
261 if (root() != NULL) { useful.push(root()); }
262 // If 'top' is cached, declare it useful to preserve cached node
263 if( cached_top_node() ) { useful.push(cached_top_node()); }
265 // Push all useful nodes onto the list, breadthfirst
266 for( uint next = 0; next < useful.size(); ++next ) {
267 assert( next < unique(), "Unique useful nodes < total nodes");
268 Node *n = useful.at(next);
269 uint max = n->len();
270 for( uint i = 0; i < max; ++i ) {
271 Node *m = n->in(i);
272 if( m == NULL ) continue;
273 useful.push(m);
274 }
275 }
276 }
278 // Disconnect all useless nodes by disconnecting those at the boundary.
279 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
280 uint next = 0;
281 while( next < useful.size() ) {
282 Node *n = useful.at(next++);
283 // Use raw traversal of out edges since this code removes out edges
284 int max = n->outcnt();
285 for (int j = 0; j < max; ++j ) {
286 Node* child = n->raw_out(j);
287 if( ! useful.member(child) ) {
288 assert( !child->is_top() || child != top(),
289 "If top is cached in Compile object it is in useful list");
290 // Only need to remove this out-edge to the useless node
291 n->raw_del_out(j);
292 --j;
293 --max;
294 }
295 }
296 if (n->outcnt() == 1 && n->has_special_unique_user()) {
297 record_for_igvn( n->unique_out() );
298 }
299 }
300 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
301 }
303 //------------------------------frame_size_in_words-----------------------------
304 // frame_slots in units of words
305 int Compile::frame_size_in_words() const {
306 // shift is 0 in LP32 and 1 in LP64
307 const int shift = (LogBytesPerWord - LogBytesPerInt);
308 int words = _frame_slots >> shift;
309 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
310 return words;
311 }
313 // ============================================================================
314 //------------------------------CompileWrapper---------------------------------
315 class CompileWrapper : public StackObj {
316 Compile *const _compile;
317 public:
318 CompileWrapper(Compile* compile);
320 ~CompileWrapper();
321 };
323 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
324 // the Compile* pointer is stored in the current ciEnv:
325 ciEnv* env = compile->env();
326 assert(env == ciEnv::current(), "must already be a ciEnv active");
327 assert(env->compiler_data() == NULL, "compile already active?");
328 env->set_compiler_data(compile);
329 assert(compile == Compile::current(), "sanity");
331 compile->set_type_dict(NULL);
332 compile->set_type_hwm(NULL);
333 compile->set_type_last_size(0);
334 compile->set_last_tf(NULL, NULL);
335 compile->set_indexSet_arena(NULL);
336 compile->set_indexSet_free_block_list(NULL);
337 compile->init_type_arena();
338 Type::Initialize(compile);
339 _compile->set_scratch_buffer_blob(NULL);
340 _compile->begin_method();
341 }
342 CompileWrapper::~CompileWrapper() {
343 _compile->end_method();
344 if (_compile->scratch_buffer_blob() != NULL)
345 BufferBlob::free(_compile->scratch_buffer_blob());
346 _compile->env()->set_compiler_data(NULL);
347 }
350 //----------------------------print_compile_messages---------------------------
351 void Compile::print_compile_messages() {
352 #ifndef PRODUCT
353 // Check if recompiling
354 if (_subsume_loads == false && PrintOpto) {
355 // Recompiling without allowing machine instructions to subsume loads
356 tty->print_cr("*********************************************************");
357 tty->print_cr("** Bailout: Recompile without subsuming loads **");
358 tty->print_cr("*********************************************************");
359 }
360 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
361 // Recompiling without escape analysis
362 tty->print_cr("*********************************************************");
363 tty->print_cr("** Bailout: Recompile without escape analysis **");
364 tty->print_cr("*********************************************************");
365 }
366 if (env()->break_at_compile()) {
367 // Open the debugger when compiling this method.
368 tty->print("### Breaking when compiling: ");
369 method()->print_short_name();
370 tty->cr();
371 BREAKPOINT;
372 }
374 if( PrintOpto ) {
375 if (is_osr_compilation()) {
376 tty->print("[OSR]%3d", _compile_id);
377 } else {
378 tty->print("%3d", _compile_id);
379 }
380 }
381 #endif
382 }
385 void Compile::init_scratch_buffer_blob() {
386 if( scratch_buffer_blob() != NULL ) return;
388 // Construct a temporary CodeBuffer to have it construct a BufferBlob
389 // Cache this BufferBlob for this compile.
390 ResourceMark rm;
391 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size);
392 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size);
393 // Record the buffer blob for next time.
394 set_scratch_buffer_blob(blob);
395 // Have we run out of code space?
396 if (scratch_buffer_blob() == NULL) {
397 // Let CompilerBroker disable further compilations.
398 record_failure("Not enough space for scratch buffer in CodeCache");
399 return;
400 }
402 // Initialize the relocation buffers
403 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size;
404 set_scratch_locs_memory(locs_buf);
405 }
408 //-----------------------scratch_emit_size-------------------------------------
409 // Helper function that computes size by emitting code
410 uint Compile::scratch_emit_size(const Node* n) {
411 // Emit into a trash buffer and count bytes emitted.
412 // This is a pretty expensive way to compute a size,
413 // but it works well enough if seldom used.
414 // All common fixed-size instructions are given a size
415 // method by the AD file.
416 // Note that the scratch buffer blob and locs memory are
417 // allocated at the beginning of the compile task, and
418 // may be shared by several calls to scratch_emit_size.
419 // The allocation of the scratch buffer blob is particularly
420 // expensive, since it has to grab the code cache lock.
421 BufferBlob* blob = this->scratch_buffer_blob();
422 assert(blob != NULL, "Initialize BufferBlob at start");
423 assert(blob->size() > MAX_inst_size, "sanity");
424 relocInfo* locs_buf = scratch_locs_memory();
425 address blob_begin = blob->instructions_begin();
426 address blob_end = (address)locs_buf;
427 assert(blob->instructions_contains(blob_end), "sanity");
428 CodeBuffer buf(blob_begin, blob_end - blob_begin);
429 buf.initialize_consts_size(MAX_const_size);
430 buf.initialize_stubs_size(MAX_stubs_size);
431 assert(locs_buf != NULL, "sanity");
432 int lsize = MAX_locs_size / 2;
433 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize);
434 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize);
435 n->emit(buf, this->regalloc());
436 return buf.code_size();
437 }
440 // ============================================================================
441 //------------------------------Compile standard-------------------------------
442 debug_only( int Compile::_debug_idx = 100000; )
444 // Compile a method. entry_bci is -1 for normal compilations and indicates
445 // the continuation bci for on stack replacement.
448 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
449 : Phase(Compiler),
450 _env(ci_env),
451 _log(ci_env->log()),
452 _compile_id(ci_env->compile_id()),
453 _save_argument_registers(false),
454 _stub_name(NULL),
455 _stub_function(NULL),
456 _stub_entry_point(NULL),
457 _method(target),
458 _entry_bci(osr_bci),
459 _initial_gvn(NULL),
460 _for_igvn(NULL),
461 _warm_calls(NULL),
462 _subsume_loads(subsume_loads),
463 _do_escape_analysis(do_escape_analysis),
464 _failure_reason(NULL),
465 _code_buffer("Compile::Fill_buffer"),
466 _orig_pc_slot(0),
467 _orig_pc_slot_offset_in_bytes(0),
468 _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 // Perform escape analysis
641 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
642 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
643 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction.
644 PhaseGVN* igvn = initial_gvn();
645 Node* oop_null = igvn->zerocon(T_OBJECT);
646 Node* noop_null = igvn->zerocon(T_NARROWOOP);
648 _congraph = new(comp_arena()) ConnectionGraph(this);
649 bool has_non_escaping_obj = _congraph->compute_escape();
651 #ifndef PRODUCT
652 if (PrintEscapeAnalysis) {
653 _congraph->dump();
654 }
655 #endif
656 // Cleanup.
657 if (oop_null->outcnt() == 0)
658 igvn->hash_delete(oop_null);
659 if (noop_null->outcnt() == 0)
660 igvn->hash_delete(noop_null);
662 if (!has_non_escaping_obj) {
663 _congraph = NULL;
664 }
666 if (failing()) return;
667 }
668 // Now optimize
669 Optimize();
670 if (failing()) return;
671 NOT_PRODUCT( verify_graph_edges(); )
673 #ifndef PRODUCT
674 if (PrintIdeal) {
675 ttyLocker ttyl; // keep the following output all in one block
676 // This output goes directly to the tty, not the compiler log.
677 // To enable tools to match it up with the compilation activity,
678 // be sure to tag this tty output with the compile ID.
679 if (xtty != NULL) {
680 xtty->head("ideal compile_id='%d'%s", compile_id(),
681 is_osr_compilation() ? " compile_kind='osr'" :
682 "");
683 }
684 root()->dump(9999);
685 if (xtty != NULL) {
686 xtty->tail("ideal");
687 }
688 }
689 #endif
691 // Now that we know the size of all the monitors we can add a fixed slot
692 // for the original deopt pc.
694 _orig_pc_slot = fixed_slots();
695 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
696 set_fixed_slots(next_slot);
698 // Now generate code
699 Code_Gen();
700 if (failing()) return;
702 // Check if we want to skip execution of all compiled code.
703 {
704 #ifndef PRODUCT
705 if (OptoNoExecute) {
706 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
707 return;
708 }
709 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
710 #endif
712 if (is_osr_compilation()) {
713 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
714 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
715 } else {
716 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
717 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
718 }
720 env()->register_method(_method, _entry_bci,
721 &_code_offsets,
722 _orig_pc_slot_offset_in_bytes,
723 code_buffer(),
724 frame_size_in_words(), _oop_map_set,
725 &_handler_table, &_inc_table,
726 compiler,
727 env()->comp_level(),
728 true, /*has_debug_info*/
729 has_unsafe_access()
730 );
731 }
732 }
734 //------------------------------Compile----------------------------------------
735 // Compile a runtime stub
736 Compile::Compile( ciEnv* ci_env,
737 TypeFunc_generator generator,
738 address stub_function,
739 const char *stub_name,
740 int is_fancy_jump,
741 bool pass_tls,
742 bool save_arg_registers,
743 bool return_pc )
744 : Phase(Compiler),
745 _env(ci_env),
746 _log(ci_env->log()),
747 _compile_id(-1),
748 _save_argument_registers(save_arg_registers),
749 _method(NULL),
750 _stub_name(stub_name),
751 _stub_function(stub_function),
752 _stub_entry_point(NULL),
753 _entry_bci(InvocationEntryBci),
754 _initial_gvn(NULL),
755 _for_igvn(NULL),
756 _warm_calls(NULL),
757 _orig_pc_slot(0),
758 _orig_pc_slot_offset_in_bytes(0),
759 _subsume_loads(true),
760 _do_escape_analysis(false),
761 _failure_reason(NULL),
762 _code_buffer("Compile::Fill_buffer"),
763 _has_method_handle_invokes(false),
764 _node_bundling_limit(0),
765 _node_bundling_base(NULL),
766 _java_calls(0),
767 _inner_loops(0),
768 #ifndef PRODUCT
769 _trace_opto_output(TraceOptoOutput),
770 _printer(NULL),
771 #endif
772 _congraph(NULL) {
773 C = this;
775 #ifndef PRODUCT
776 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
777 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
778 set_print_assembly(PrintFrameConverterAssembly);
779 set_parsed_irreducible_loop(false);
780 #endif
781 CompileWrapper cw(this);
782 Init(/*AliasLevel=*/ 0);
783 init_tf((*generator)());
785 {
786 // The following is a dummy for the sake of GraphKit::gen_stub
787 Unique_Node_List for_igvn(comp_arena());
788 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
789 PhaseGVN gvn(Thread::current()->resource_area(),255);
790 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
791 gvn.transform_no_reclaim(top());
793 GraphKit kit;
794 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
795 }
797 NOT_PRODUCT( verify_graph_edges(); )
798 Code_Gen();
799 if (failing()) return;
802 // Entry point will be accessed using compile->stub_entry_point();
803 if (code_buffer() == NULL) {
804 Matcher::soft_match_failure();
805 } else {
806 if (PrintAssembly && (WizardMode || Verbose))
807 tty->print_cr("### Stub::%s", stub_name);
809 if (!failing()) {
810 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
812 // Make the NMethod
813 // For now we mark the frame as never safe for profile stackwalking
814 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
815 code_buffer(),
816 CodeOffsets::frame_never_safe,
817 // _code_offsets.value(CodeOffsets::Frame_Complete),
818 frame_size_in_words(),
819 _oop_map_set,
820 save_arg_registers);
821 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
823 _stub_entry_point = rs->entry_point();
824 }
825 }
826 }
828 #ifndef PRODUCT
829 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
830 if(PrintOpto && Verbose) {
831 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
832 }
833 }
834 #endif
836 void Compile::print_codes() {
837 }
839 //------------------------------Init-------------------------------------------
840 // Prepare for a single compilation
841 void Compile::Init(int aliaslevel) {
842 _unique = 0;
843 _regalloc = NULL;
845 _tf = NULL; // filled in later
846 _top = NULL; // cached later
847 _matcher = NULL; // filled in later
848 _cfg = NULL; // filled in later
850 set_24_bit_selection_and_mode(Use24BitFP, false);
852 _node_note_array = NULL;
853 _default_node_notes = NULL;
855 _immutable_memory = NULL; // filled in at first inquiry
857 // Globally visible Nodes
858 // First set TOP to NULL to give safe behavior during creation of RootNode
859 set_cached_top_node(NULL);
860 set_root(new (this, 3) RootNode());
861 // Now that you have a Root to point to, create the real TOP
862 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
863 set_recent_alloc(NULL, NULL);
865 // Create Debug Information Recorder to record scopes, oopmaps, etc.
866 env()->set_oop_recorder(new OopRecorder(comp_arena()));
867 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
868 env()->set_dependencies(new Dependencies(env()));
870 _fixed_slots = 0;
871 set_has_split_ifs(false);
872 set_has_loops(has_method() && method()->has_loops()); // first approximation
873 set_has_stringbuilder(false);
874 _deopt_happens = true; // start out assuming the worst
875 _trap_can_recompile = false; // no traps emitted yet
876 _major_progress = true; // start out assuming good things will happen
877 set_has_unsafe_access(false);
878 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
879 set_decompile_count(0);
881 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
882 // Compilation level related initialization
883 if (env()->comp_level() == CompLevel_fast_compile) {
884 set_num_loop_opts(Tier1LoopOptsCount);
885 set_do_inlining(Tier1Inline != 0);
886 set_max_inline_size(Tier1MaxInlineSize);
887 set_freq_inline_size(Tier1FreqInlineSize);
888 set_do_scheduling(false);
889 set_do_count_invocations(Tier1CountInvocations);
890 set_do_method_data_update(Tier1UpdateMethodData);
891 } else {
892 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level");
893 set_num_loop_opts(LoopOptsCount);
894 set_do_inlining(Inline);
895 set_max_inline_size(MaxInlineSize);
896 set_freq_inline_size(FreqInlineSize);
897 set_do_scheduling(OptoScheduling);
898 set_do_count_invocations(false);
899 set_do_method_data_update(false);
900 }
902 if (debug_info()->recording_non_safepoints()) {
903 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
904 (comp_arena(), 8, 0, NULL));
905 set_default_node_notes(Node_Notes::make(this));
906 }
908 // // -- Initialize types before each compile --
909 // // Update cached type information
910 // if( _method && _method->constants() )
911 // Type::update_loaded_types(_method, _method->constants());
913 // Init alias_type map.
914 if (!_do_escape_analysis && aliaslevel == 3)
915 aliaslevel = 2; // No unique types without escape analysis
916 _AliasLevel = aliaslevel;
917 const int grow_ats = 16;
918 _max_alias_types = grow_ats;
919 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
920 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
921 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
922 {
923 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
924 }
925 // Initialize the first few types.
926 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
927 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
928 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
929 _num_alias_types = AliasIdxRaw+1;
930 // Zero out the alias type cache.
931 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
932 // A NULL adr_type hits in the cache right away. Preload the right answer.
933 probe_alias_cache(NULL)->_index = AliasIdxTop;
935 _intrinsics = NULL;
936 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
937 _predicate_opaqs = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
938 register_library_intrinsics();
939 }
941 //---------------------------init_start----------------------------------------
942 // Install the StartNode on this compile object.
943 void Compile::init_start(StartNode* s) {
944 if (failing())
945 return; // already failing
946 assert(s == start(), "");
947 }
949 StartNode* Compile::start() const {
950 assert(!failing(), "");
951 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
952 Node* start = root()->fast_out(i);
953 if( start->is_Start() )
954 return start->as_Start();
955 }
956 ShouldNotReachHere();
957 return NULL;
958 }
960 //-------------------------------immutable_memory-------------------------------------
961 // Access immutable memory
962 Node* Compile::immutable_memory() {
963 if (_immutable_memory != NULL) {
964 return _immutable_memory;
965 }
966 StartNode* s = start();
967 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
968 Node *p = s->fast_out(i);
969 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
970 _immutable_memory = p;
971 return _immutable_memory;
972 }
973 }
974 ShouldNotReachHere();
975 return NULL;
976 }
978 //----------------------set_cached_top_node------------------------------------
979 // Install the cached top node, and make sure Node::is_top works correctly.
980 void Compile::set_cached_top_node(Node* tn) {
981 if (tn != NULL) verify_top(tn);
982 Node* old_top = _top;
983 _top = tn;
984 // Calling Node::setup_is_top allows the nodes the chance to adjust
985 // their _out arrays.
986 if (_top != NULL) _top->setup_is_top();
987 if (old_top != NULL) old_top->setup_is_top();
988 assert(_top == NULL || top()->is_top(), "");
989 }
991 #ifndef PRODUCT
992 void Compile::verify_top(Node* tn) const {
993 if (tn != NULL) {
994 assert(tn->is_Con(), "top node must be a constant");
995 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
996 assert(tn->in(0) != NULL, "must have live top node");
997 }
998 }
999 #endif
1002 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1004 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1005 guarantee(arr != NULL, "");
1006 int num_blocks = arr->length();
1007 if (grow_by < num_blocks) grow_by = num_blocks;
1008 int num_notes = grow_by * _node_notes_block_size;
1009 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1010 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1011 while (num_notes > 0) {
1012 arr->append(notes);
1013 notes += _node_notes_block_size;
1014 num_notes -= _node_notes_block_size;
1015 }
1016 assert(num_notes == 0, "exact multiple, please");
1017 }
1019 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1020 if (source == NULL || dest == NULL) return false;
1022 if (dest->is_Con())
1023 return false; // Do not push debug info onto constants.
1025 #ifdef ASSERT
1026 // Leave a bread crumb trail pointing to the original node:
1027 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1028 dest->set_debug_orig(source);
1029 }
1030 #endif
1032 if (node_note_array() == NULL)
1033 return false; // Not collecting any notes now.
1035 // This is a copy onto a pre-existing node, which may already have notes.
1036 // If both nodes have notes, do not overwrite any pre-existing notes.
1037 Node_Notes* source_notes = node_notes_at(source->_idx);
1038 if (source_notes == NULL || source_notes->is_clear()) return false;
1039 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1040 if (dest_notes == NULL || dest_notes->is_clear()) {
1041 return set_node_notes_at(dest->_idx, source_notes);
1042 }
1044 Node_Notes merged_notes = (*source_notes);
1045 // The order of operations here ensures that dest notes will win...
1046 merged_notes.update_from(dest_notes);
1047 return set_node_notes_at(dest->_idx, &merged_notes);
1048 }
1051 //--------------------------allow_range_check_smearing-------------------------
1052 // Gating condition for coalescing similar range checks.
1053 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1054 // single covering check that is at least as strong as any of them.
1055 // If the optimization succeeds, the simplified (strengthened) range check
1056 // will always succeed. If it fails, we will deopt, and then give up
1057 // on the optimization.
1058 bool Compile::allow_range_check_smearing() const {
1059 // If this method has already thrown a range-check,
1060 // assume it was because we already tried range smearing
1061 // and it failed.
1062 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1063 return !already_trapped;
1064 }
1067 //------------------------------flatten_alias_type-----------------------------
1068 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1069 int offset = tj->offset();
1070 TypePtr::PTR ptr = tj->ptr();
1072 // Known instance (scalarizable allocation) alias only with itself.
1073 bool is_known_inst = tj->isa_oopptr() != NULL &&
1074 tj->is_oopptr()->is_known_instance();
1076 // Process weird unsafe references.
1077 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1078 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1079 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1080 tj = TypeOopPtr::BOTTOM;
1081 ptr = tj->ptr();
1082 offset = tj->offset();
1083 }
1085 // Array pointers need some flattening
1086 const TypeAryPtr *ta = tj->isa_aryptr();
1087 if( ta && is_known_inst ) {
1088 if ( offset != Type::OffsetBot &&
1089 offset > arrayOopDesc::length_offset_in_bytes() ) {
1090 offset = Type::OffsetBot; // Flatten constant access into array body only
1091 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1092 }
1093 } else if( ta && _AliasLevel >= 2 ) {
1094 // For arrays indexed by constant indices, we flatten the alias
1095 // space to include all of the array body. Only the header, klass
1096 // and array length can be accessed un-aliased.
1097 if( offset != Type::OffsetBot ) {
1098 if( ta->const_oop() ) { // methodDataOop or methodOop
1099 offset = Type::OffsetBot; // Flatten constant access into array body
1100 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1101 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1102 // range is OK as-is.
1103 tj = ta = TypeAryPtr::RANGE;
1104 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1105 tj = TypeInstPtr::KLASS; // all klass loads look alike
1106 ta = TypeAryPtr::RANGE; // generic ignored junk
1107 ptr = TypePtr::BotPTR;
1108 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1109 tj = TypeInstPtr::MARK;
1110 ta = TypeAryPtr::RANGE; // generic ignored junk
1111 ptr = TypePtr::BotPTR;
1112 } else { // Random constant offset into array body
1113 offset = Type::OffsetBot; // Flatten constant access into array body
1114 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1115 }
1116 }
1117 // Arrays of fixed size alias with arrays of unknown size.
1118 if (ta->size() != TypeInt::POS) {
1119 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1120 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1121 }
1122 // Arrays of known objects become arrays of unknown objects.
1123 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1124 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1125 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1126 }
1127 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1128 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1129 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1130 }
1131 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1132 // cannot be distinguished by bytecode alone.
1133 if (ta->elem() == TypeInt::BOOL) {
1134 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1135 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1136 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1137 }
1138 // During the 2nd round of IterGVN, NotNull castings are removed.
1139 // Make sure the Bottom and NotNull variants alias the same.
1140 // Also, make sure exact and non-exact variants alias the same.
1141 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1142 if (ta->const_oop()) {
1143 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1144 } else {
1145 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1146 }
1147 }
1148 }
1150 // Oop pointers need some flattening
1151 const TypeInstPtr *to = tj->isa_instptr();
1152 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1153 if( ptr == TypePtr::Constant ) {
1154 // No constant oop pointers (such as Strings); they alias with
1155 // unknown strings.
1156 assert(!is_known_inst, "not scalarizable allocation");
1157 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1158 } else if( is_known_inst ) {
1159 tj = to; // Keep NotNull and klass_is_exact for instance type
1160 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1161 // During the 2nd round of IterGVN, NotNull castings are removed.
1162 // Make sure the Bottom and NotNull variants alias the same.
1163 // Also, make sure exact and non-exact variants alias the same.
1164 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1165 }
1166 // Canonicalize the holder of this field
1167 ciInstanceKlass *k = to->klass()->as_instance_klass();
1168 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1169 // First handle header references such as a LoadKlassNode, even if the
1170 // object's klass is unloaded at compile time (4965979).
1171 if (!is_known_inst) { // Do it only for non-instance types
1172 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1173 }
1174 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1175 to = NULL;
1176 tj = TypeOopPtr::BOTTOM;
1177 offset = tj->offset();
1178 } else {
1179 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1180 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1181 if( is_known_inst ) {
1182 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1183 } else {
1184 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1185 }
1186 }
1187 }
1188 }
1190 // Klass pointers to object array klasses need some flattening
1191 const TypeKlassPtr *tk = tj->isa_klassptr();
1192 if( tk ) {
1193 // If we are referencing a field within a Klass, we need
1194 // to assume the worst case of an Object. Both exact and
1195 // inexact types must flatten to the same alias class.
1196 // Since the flattened result for a klass is defined to be
1197 // precisely java.lang.Object, use a constant ptr.
1198 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1200 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1201 TypeKlassPtr::OBJECT->klass(),
1202 offset);
1203 }
1205 ciKlass* klass = tk->klass();
1206 if( klass->is_obj_array_klass() ) {
1207 ciKlass* k = TypeAryPtr::OOPS->klass();
1208 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1209 k = TypeInstPtr::BOTTOM->klass();
1210 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1211 }
1213 // Check for precise loads from the primary supertype array and force them
1214 // to the supertype cache alias index. Check for generic array loads from
1215 // the primary supertype array and also force them to the supertype cache
1216 // alias index. Since the same load can reach both, we need to merge
1217 // these 2 disparate memories into the same alias class. Since the
1218 // primary supertype array is read-only, there's no chance of confusion
1219 // where we bypass an array load and an array store.
1220 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1221 if( offset == Type::OffsetBot ||
1222 off2 < Klass::primary_super_limit()*wordSize ) {
1223 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1224 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1225 }
1226 }
1228 // Flatten all Raw pointers together.
1229 if (tj->base() == Type::RawPtr)
1230 tj = TypeRawPtr::BOTTOM;
1232 if (tj->base() == Type::AnyPtr)
1233 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1235 // Flatten all to bottom for now
1236 switch( _AliasLevel ) {
1237 case 0:
1238 tj = TypePtr::BOTTOM;
1239 break;
1240 case 1: // Flatten to: oop, static, field or array
1241 switch (tj->base()) {
1242 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1243 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1244 case Type::AryPtr: // do not distinguish arrays at all
1245 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1246 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1247 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1248 default: ShouldNotReachHere();
1249 }
1250 break;
1251 case 2: // No collapsing at level 2; keep all splits
1252 case 3: // No collapsing at level 3; keep all splits
1253 break;
1254 default:
1255 Unimplemented();
1256 }
1258 offset = tj->offset();
1259 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1261 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1262 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1263 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1264 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1265 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1266 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1267 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1268 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1269 assert( tj->ptr() != TypePtr::TopPTR &&
1270 tj->ptr() != TypePtr::AnyNull &&
1271 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1272 // assert( tj->ptr() != TypePtr::Constant ||
1273 // tj->base() == Type::RawPtr ||
1274 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1276 return tj;
1277 }
1279 void Compile::AliasType::Init(int i, const TypePtr* at) {
1280 _index = i;
1281 _adr_type = at;
1282 _field = NULL;
1283 _is_rewritable = true; // default
1284 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1285 if (atoop != NULL && atoop->is_known_instance()) {
1286 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1287 _general_index = Compile::current()->get_alias_index(gt);
1288 } else {
1289 _general_index = 0;
1290 }
1291 }
1293 //---------------------------------print_on------------------------------------
1294 #ifndef PRODUCT
1295 void Compile::AliasType::print_on(outputStream* st) {
1296 if (index() < 10)
1297 st->print("@ <%d> ", index());
1298 else st->print("@ <%d>", index());
1299 st->print(is_rewritable() ? " " : " RO");
1300 int offset = adr_type()->offset();
1301 if (offset == Type::OffsetBot)
1302 st->print(" +any");
1303 else st->print(" +%-3d", offset);
1304 st->print(" in ");
1305 adr_type()->dump_on(st);
1306 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1307 if (field() != NULL && tjp) {
1308 if (tjp->klass() != field()->holder() ||
1309 tjp->offset() != field()->offset_in_bytes()) {
1310 st->print(" != ");
1311 field()->print();
1312 st->print(" ***");
1313 }
1314 }
1315 }
1317 void print_alias_types() {
1318 Compile* C = Compile::current();
1319 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1320 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1321 C->alias_type(idx)->print_on(tty);
1322 tty->cr();
1323 }
1324 }
1325 #endif
1328 //----------------------------probe_alias_cache--------------------------------
1329 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1330 intptr_t key = (intptr_t) adr_type;
1331 key ^= key >> logAliasCacheSize;
1332 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1333 }
1336 //-----------------------------grow_alias_types--------------------------------
1337 void Compile::grow_alias_types() {
1338 const int old_ats = _max_alias_types; // how many before?
1339 const int new_ats = old_ats; // how many more?
1340 const int grow_ats = old_ats+new_ats; // how many now?
1341 _max_alias_types = grow_ats;
1342 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1343 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1344 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1345 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1346 }
1349 //--------------------------------find_alias_type------------------------------
1350 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) {
1351 if (_AliasLevel == 0)
1352 return alias_type(AliasIdxBot);
1354 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1355 if (ace->_adr_type == adr_type) {
1356 return alias_type(ace->_index);
1357 }
1359 // Handle special cases.
1360 if (adr_type == NULL) return alias_type(AliasIdxTop);
1361 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1363 // Do it the slow way.
1364 const TypePtr* flat = flatten_alias_type(adr_type);
1366 #ifdef ASSERT
1367 assert(flat == flatten_alias_type(flat), "idempotent");
1368 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1369 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1370 const TypeOopPtr* foop = flat->is_oopptr();
1371 // Scalarizable allocations have exact klass always.
1372 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1373 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1374 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1375 }
1376 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1377 #endif
1379 int idx = AliasIdxTop;
1380 for (int i = 0; i < num_alias_types(); i++) {
1381 if (alias_type(i)->adr_type() == flat) {
1382 idx = i;
1383 break;
1384 }
1385 }
1387 if (idx == AliasIdxTop) {
1388 if (no_create) return NULL;
1389 // Grow the array if necessary.
1390 if (_num_alias_types == _max_alias_types) grow_alias_types();
1391 // Add a new alias type.
1392 idx = _num_alias_types++;
1393 _alias_types[idx]->Init(idx, flat);
1394 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1395 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1396 if (flat->isa_instptr()) {
1397 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1398 && flat->is_instptr()->klass() == env()->Class_klass())
1399 alias_type(idx)->set_rewritable(false);
1400 }
1401 if (flat->isa_klassptr()) {
1402 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1403 alias_type(idx)->set_rewritable(false);
1404 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1405 alias_type(idx)->set_rewritable(false);
1406 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1407 alias_type(idx)->set_rewritable(false);
1408 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1409 alias_type(idx)->set_rewritable(false);
1410 }
1411 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1412 // but the base pointer type is not distinctive enough to identify
1413 // references into JavaThread.)
1415 // Check for final instance fields.
1416 const TypeInstPtr* tinst = flat->isa_instptr();
1417 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1418 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1419 ciField* field = k->get_field_by_offset(tinst->offset(), false);
1420 // Set field() and is_rewritable() attributes.
1421 if (field != NULL) alias_type(idx)->set_field(field);
1422 }
1423 const TypeKlassPtr* tklass = flat->isa_klassptr();
1424 // Check for final static fields.
1425 if (tklass && tklass->klass()->is_instance_klass()) {
1426 ciInstanceKlass *k = tklass->klass()->as_instance_klass();
1427 ciField* field = k->get_field_by_offset(tklass->offset(), true);
1428 // Set field() and is_rewritable() attributes.
1429 if (field != NULL) alias_type(idx)->set_field(field);
1430 }
1431 }
1433 // Fill the cache for next time.
1434 ace->_adr_type = adr_type;
1435 ace->_index = idx;
1436 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1438 // Might as well try to fill the cache for the flattened version, too.
1439 AliasCacheEntry* face = probe_alias_cache(flat);
1440 if (face->_adr_type == NULL) {
1441 face->_adr_type = flat;
1442 face->_index = idx;
1443 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1444 }
1446 return alias_type(idx);
1447 }
1450 Compile::AliasType* Compile::alias_type(ciField* field) {
1451 const TypeOopPtr* t;
1452 if (field->is_static())
1453 t = TypeKlassPtr::make(field->holder());
1454 else
1455 t = TypeOopPtr::make_from_klass_raw(field->holder());
1456 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()));
1457 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1458 return atp;
1459 }
1462 //------------------------------have_alias_type--------------------------------
1463 bool Compile::have_alias_type(const TypePtr* adr_type) {
1464 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1465 if (ace->_adr_type == adr_type) {
1466 return true;
1467 }
1469 // Handle special cases.
1470 if (adr_type == NULL) return true;
1471 if (adr_type == TypePtr::BOTTOM) return true;
1473 return find_alias_type(adr_type, true) != NULL;
1474 }
1476 //-----------------------------must_alias--------------------------------------
1477 // True if all values of the given address type are in the given alias category.
1478 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1479 if (alias_idx == AliasIdxBot) return true; // the universal category
1480 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1481 if (alias_idx == AliasIdxTop) return false; // the empty category
1482 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1484 // the only remaining possible overlap is identity
1485 int adr_idx = get_alias_index(adr_type);
1486 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1487 assert(adr_idx == alias_idx ||
1488 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1489 && adr_type != TypeOopPtr::BOTTOM),
1490 "should not be testing for overlap with an unsafe pointer");
1491 return adr_idx == alias_idx;
1492 }
1494 //------------------------------can_alias--------------------------------------
1495 // True if any values of the given address type are in the given alias category.
1496 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1497 if (alias_idx == AliasIdxTop) return false; // the empty category
1498 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1499 if (alias_idx == AliasIdxBot) return true; // the universal category
1500 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1502 // the only remaining possible overlap is identity
1503 int adr_idx = get_alias_index(adr_type);
1504 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1505 return adr_idx == alias_idx;
1506 }
1510 //---------------------------pop_warm_call-------------------------------------
1511 WarmCallInfo* Compile::pop_warm_call() {
1512 WarmCallInfo* wci = _warm_calls;
1513 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1514 return wci;
1515 }
1517 //----------------------------Inline_Warm--------------------------------------
1518 int Compile::Inline_Warm() {
1519 // If there is room, try to inline some more warm call sites.
1520 // %%% Do a graph index compaction pass when we think we're out of space?
1521 if (!InlineWarmCalls) return 0;
1523 int calls_made_hot = 0;
1524 int room_to_grow = NodeCountInliningCutoff - unique();
1525 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1526 int amount_grown = 0;
1527 WarmCallInfo* call;
1528 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1529 int est_size = (int)call->size();
1530 if (est_size > (room_to_grow - amount_grown)) {
1531 // This one won't fit anyway. Get rid of it.
1532 call->make_cold();
1533 continue;
1534 }
1535 call->make_hot();
1536 calls_made_hot++;
1537 amount_grown += est_size;
1538 amount_to_grow -= est_size;
1539 }
1541 if (calls_made_hot > 0) set_major_progress();
1542 return calls_made_hot;
1543 }
1546 //----------------------------Finish_Warm--------------------------------------
1547 void Compile::Finish_Warm() {
1548 if (!InlineWarmCalls) return;
1549 if (failing()) return;
1550 if (warm_calls() == NULL) return;
1552 // Clean up loose ends, if we are out of space for inlining.
1553 WarmCallInfo* call;
1554 while ((call = pop_warm_call()) != NULL) {
1555 call->make_cold();
1556 }
1557 }
1559 //---------------------cleanup_loop_predicates-----------------------
1560 // Remove the opaque nodes that protect the predicates so that all unused
1561 // checks and uncommon_traps will be eliminated from the ideal graph
1562 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1563 if (predicate_count()==0) return;
1564 for (int i = predicate_count(); i > 0; i--) {
1565 Node * n = predicate_opaque1_node(i-1);
1566 assert(n->Opcode() == Op_Opaque1, "must be");
1567 igvn.replace_node(n, n->in(1));
1568 }
1569 assert(predicate_count()==0, "should be clean!");
1570 igvn.optimize();
1571 }
1573 //------------------------------Optimize---------------------------------------
1574 // Given a graph, optimize it.
1575 void Compile::Optimize() {
1576 TracePhase t1("optimizer", &_t_optimizer, true);
1578 #ifndef PRODUCT
1579 if (env()->break_at_compile()) {
1580 BREAKPOINT;
1581 }
1583 #endif
1585 ResourceMark rm;
1586 int loop_opts_cnt;
1588 NOT_PRODUCT( verify_graph_edges(); )
1590 print_method("After Parsing");
1592 {
1593 // Iterative Global Value Numbering, including ideal transforms
1594 // Initialize IterGVN with types and values from parse-time GVN
1595 PhaseIterGVN igvn(initial_gvn());
1596 {
1597 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1598 igvn.optimize();
1599 }
1601 print_method("Iter GVN 1", 2);
1603 if (failing()) return;
1605 // Loop transforms on the ideal graph. Range Check Elimination,
1606 // peeling, unrolling, etc.
1608 // Set loop opts counter
1609 loop_opts_cnt = num_loop_opts();
1610 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1611 {
1612 TracePhase t2("idealLoop", &_t_idealLoop, true);
1613 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate);
1614 loop_opts_cnt--;
1615 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1616 if (failing()) return;
1617 }
1618 // Loop opts pass if partial peeling occurred in previous pass
1619 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1620 TracePhase t3("idealLoop", &_t_idealLoop, true);
1621 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
1622 loop_opts_cnt--;
1623 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1624 if (failing()) return;
1625 }
1626 // Loop opts pass for loop-unrolling before CCP
1627 if(major_progress() && (loop_opts_cnt > 0)) {
1628 TracePhase t4("idealLoop", &_t_idealLoop, true);
1629 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
1630 loop_opts_cnt--;
1631 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1632 }
1633 if (!failing()) {
1634 // Verify that last round of loop opts produced a valid graph
1635 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1636 PhaseIdealLoop::verify(igvn);
1637 }
1638 }
1639 if (failing()) return;
1641 // Conditional Constant Propagation;
1642 PhaseCCP ccp( &igvn );
1643 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1644 {
1645 TracePhase t2("ccp", &_t_ccp, true);
1646 ccp.do_transform();
1647 }
1648 print_method("PhaseCPP 1", 2);
1650 assert( true, "Break here to ccp.dump_old2new_map()");
1652 // Iterative Global Value Numbering, including ideal transforms
1653 {
1654 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1655 igvn = ccp;
1656 igvn.optimize();
1657 }
1659 print_method("Iter GVN 2", 2);
1661 if (failing()) return;
1663 // Loop transforms on the ideal graph. Range Check Elimination,
1664 // peeling, unrolling, etc.
1665 if(loop_opts_cnt > 0) {
1666 debug_only( int cnt = 0; );
1667 bool loop_predication = UseLoopPredicate;
1668 while(major_progress() && (loop_opts_cnt > 0)) {
1669 TracePhase t2("idealLoop", &_t_idealLoop, true);
1670 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1671 PhaseIdealLoop ideal_loop( igvn, true, loop_predication);
1672 loop_opts_cnt--;
1673 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1674 if (failing()) return;
1675 // Perform loop predication optimization during first iteration after CCP.
1676 // After that switch it off and cleanup unused loop predicates.
1677 if (loop_predication) {
1678 loop_predication = false;
1679 cleanup_loop_predicates(igvn);
1680 if (failing()) return;
1681 }
1682 }
1683 }
1685 {
1686 // Verify that all previous optimizations produced a valid graph
1687 // at least to this point, even if no loop optimizations were done.
1688 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1689 PhaseIdealLoop::verify(igvn);
1690 }
1692 {
1693 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1694 PhaseMacroExpand mex(igvn);
1695 if (mex.expand_macro_nodes()) {
1696 assert(failing(), "must bail out w/ explicit message");
1697 return;
1698 }
1699 }
1701 } // (End scope of igvn; run destructor if necessary for asserts.)
1703 // A method with only infinite loops has no edges entering loops from root
1704 {
1705 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1706 if (final_graph_reshaping()) {
1707 assert(failing(), "must bail out w/ explicit message");
1708 return;
1709 }
1710 }
1712 print_method("Optimize finished", 2);
1713 }
1716 //------------------------------Code_Gen---------------------------------------
1717 // Given a graph, generate code for it
1718 void Compile::Code_Gen() {
1719 if (failing()) return;
1721 // Perform instruction selection. You might think we could reclaim Matcher
1722 // memory PDQ, but actually the Matcher is used in generating spill code.
1723 // Internals of the Matcher (including some VectorSets) must remain live
1724 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1725 // set a bit in reclaimed memory.
1727 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1728 // nodes. Mapping is only valid at the root of each matched subtree.
1729 NOT_PRODUCT( verify_graph_edges(); )
1731 Node_List proj_list;
1732 Matcher m(proj_list);
1733 _matcher = &m;
1734 {
1735 TracePhase t2("matcher", &_t_matcher, true);
1736 m.match();
1737 }
1738 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1739 // nodes. Mapping is only valid at the root of each matched subtree.
1740 NOT_PRODUCT( verify_graph_edges(); )
1742 // If you have too many nodes, or if matching has failed, bail out
1743 check_node_count(0, "out of nodes matching instructions");
1744 if (failing()) return;
1746 // Build a proper-looking CFG
1747 PhaseCFG cfg(node_arena(), root(), m);
1748 _cfg = &cfg;
1749 {
1750 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1751 cfg.Dominators();
1752 if (failing()) return;
1754 NOT_PRODUCT( verify_graph_edges(); )
1756 cfg.Estimate_Block_Frequency();
1757 cfg.GlobalCodeMotion(m,unique(),proj_list);
1759 print_method("Global code motion", 2);
1761 if (failing()) return;
1762 NOT_PRODUCT( verify_graph_edges(); )
1764 debug_only( cfg.verify(); )
1765 }
1766 NOT_PRODUCT( verify_graph_edges(); )
1768 PhaseChaitin regalloc(unique(),cfg,m);
1769 _regalloc = ®alloc;
1770 {
1771 TracePhase t2("regalloc", &_t_registerAllocation, true);
1772 // Perform any platform dependent preallocation actions. This is used,
1773 // for example, to avoid taking an implicit null pointer exception
1774 // using the frame pointer on win95.
1775 _regalloc->pd_preallocate_hook();
1777 // Perform register allocation. After Chaitin, use-def chains are
1778 // no longer accurate (at spill code) and so must be ignored.
1779 // Node->LRG->reg mappings are still accurate.
1780 _regalloc->Register_Allocate();
1782 // Bail out if the allocator builds too many nodes
1783 if (failing()) return;
1784 }
1786 // Prior to register allocation we kept empty basic blocks in case the
1787 // the allocator needed a place to spill. After register allocation we
1788 // are not adding any new instructions. If any basic block is empty, we
1789 // can now safely remove it.
1790 {
1791 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1792 cfg.remove_empty();
1793 if (do_freq_based_layout()) {
1794 PhaseBlockLayout layout(cfg);
1795 } else {
1796 cfg.set_loop_alignment();
1797 }
1798 cfg.fixup_flow();
1799 }
1801 // Perform any platform dependent postallocation verifications.
1802 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1804 // Apply peephole optimizations
1805 if( OptoPeephole ) {
1806 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1807 PhasePeephole peep( _regalloc, cfg);
1808 peep.do_transform();
1809 }
1811 // Convert Nodes to instruction bits in a buffer
1812 {
1813 // %%%% workspace merge brought two timers together for one job
1814 TracePhase t2a("output", &_t_output, true);
1815 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1816 Output();
1817 }
1819 print_method("Final Code");
1821 // He's dead, Jim.
1822 _cfg = (PhaseCFG*)0xdeadbeef;
1823 _regalloc = (PhaseChaitin*)0xdeadbeef;
1824 }
1827 //------------------------------dump_asm---------------------------------------
1828 // Dump formatted assembly
1829 #ifndef PRODUCT
1830 void Compile::dump_asm(int *pcs, uint pc_limit) {
1831 bool cut_short = false;
1832 tty->print_cr("#");
1833 tty->print("# "); _tf->dump(); tty->cr();
1834 tty->print_cr("#");
1836 // For all blocks
1837 int pc = 0x0; // Program counter
1838 char starts_bundle = ' ';
1839 _regalloc->dump_frame();
1841 Node *n = NULL;
1842 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1843 if (VMThread::should_terminate()) { cut_short = true; break; }
1844 Block *b = _cfg->_blocks[i];
1845 if (b->is_connector() && !Verbose) continue;
1846 n = b->_nodes[0];
1847 if (pcs && n->_idx < pc_limit)
1848 tty->print("%3.3x ", pcs[n->_idx]);
1849 else
1850 tty->print(" ");
1851 b->dump_head( &_cfg->_bbs );
1852 if (b->is_connector()) {
1853 tty->print_cr(" # Empty connector block");
1854 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1855 tty->print_cr(" # Block is sole successor of call");
1856 }
1858 // For all instructions
1859 Node *delay = NULL;
1860 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1861 if (VMThread::should_terminate()) { cut_short = true; break; }
1862 n = b->_nodes[j];
1863 if (valid_bundle_info(n)) {
1864 Bundle *bundle = node_bundling(n);
1865 if (bundle->used_in_unconditional_delay()) {
1866 delay = n;
1867 continue;
1868 }
1869 if (bundle->starts_bundle())
1870 starts_bundle = '+';
1871 }
1873 if (WizardMode) n->dump();
1875 if( !n->is_Region() && // Dont print in the Assembly
1876 !n->is_Phi() && // a few noisely useless nodes
1877 !n->is_Proj() &&
1878 !n->is_MachTemp() &&
1879 !n->is_SafePointScalarObject() &&
1880 !n->is_Catch() && // Would be nice to print exception table targets
1881 !n->is_MergeMem() && // Not very interesting
1882 !n->is_top() && // Debug info table constants
1883 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1884 ) {
1885 if (pcs && n->_idx < pc_limit)
1886 tty->print("%3.3x", pcs[n->_idx]);
1887 else
1888 tty->print(" ");
1889 tty->print(" %c ", starts_bundle);
1890 starts_bundle = ' ';
1891 tty->print("\t");
1892 n->format(_regalloc, tty);
1893 tty->cr();
1894 }
1896 // If we have an instruction with a delay slot, and have seen a delay,
1897 // then back up and print it
1898 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1899 assert(delay != NULL, "no unconditional delay instruction");
1900 if (WizardMode) delay->dump();
1902 if (node_bundling(delay)->starts_bundle())
1903 starts_bundle = '+';
1904 if (pcs && n->_idx < pc_limit)
1905 tty->print("%3.3x", pcs[n->_idx]);
1906 else
1907 tty->print(" ");
1908 tty->print(" %c ", starts_bundle);
1909 starts_bundle = ' ';
1910 tty->print("\t");
1911 delay->format(_regalloc, tty);
1912 tty->print_cr("");
1913 delay = NULL;
1914 }
1916 // Dump the exception table as well
1917 if( n->is_Catch() && (Verbose || WizardMode) ) {
1918 // Print the exception table for this offset
1919 _handler_table.print_subtable_for(pc);
1920 }
1921 }
1923 if (pcs && n->_idx < pc_limit)
1924 tty->print_cr("%3.3x", pcs[n->_idx]);
1925 else
1926 tty->print_cr("");
1928 assert(cut_short || delay == NULL, "no unconditional delay branch");
1930 } // End of per-block dump
1931 tty->print_cr("");
1933 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
1934 }
1935 #endif
1937 //------------------------------Final_Reshape_Counts---------------------------
1938 // This class defines counters to help identify when a method
1939 // may/must be executed using hardware with only 24-bit precision.
1940 struct Final_Reshape_Counts : public StackObj {
1941 int _call_count; // count non-inlined 'common' calls
1942 int _float_count; // count float ops requiring 24-bit precision
1943 int _double_count; // count double ops requiring more precision
1944 int _java_call_count; // count non-inlined 'java' calls
1945 int _inner_loop_count; // count loops which need alignment
1946 VectorSet _visited; // Visitation flags
1947 Node_List _tests; // Set of IfNodes & PCTableNodes
1949 Final_Reshape_Counts() :
1950 _call_count(0), _float_count(0), _double_count(0),
1951 _java_call_count(0), _inner_loop_count(0),
1952 _visited( Thread::current()->resource_area() ) { }
1954 void inc_call_count () { _call_count ++; }
1955 void inc_float_count () { _float_count ++; }
1956 void inc_double_count() { _double_count++; }
1957 void inc_java_call_count() { _java_call_count++; }
1958 void inc_inner_loop_count() { _inner_loop_count++; }
1960 int get_call_count () const { return _call_count ; }
1961 int get_float_count () const { return _float_count ; }
1962 int get_double_count() const { return _double_count; }
1963 int get_java_call_count() const { return _java_call_count; }
1964 int get_inner_loop_count() const { return _inner_loop_count; }
1965 };
1967 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
1968 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1969 // Make sure the offset goes inside the instance layout.
1970 return k->contains_field_offset(tp->offset());
1971 // Note that OffsetBot and OffsetTop are very negative.
1972 }
1974 //------------------------------final_graph_reshaping_impl----------------------
1975 // Implement items 1-5 from final_graph_reshaping below.
1976 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
1978 if ( n->outcnt() == 0 ) return; // dead node
1979 uint nop = n->Opcode();
1981 // Check for 2-input instruction with "last use" on right input.
1982 // Swap to left input. Implements item (2).
1983 if( n->req() == 3 && // two-input instruction
1984 n->in(1)->outcnt() > 1 && // left use is NOT a last use
1985 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
1986 n->in(2)->outcnt() == 1 &&// right use IS a last use
1987 !n->in(2)->is_Con() ) { // right use is not a constant
1988 // Check for commutative opcode
1989 switch( nop ) {
1990 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
1991 case Op_MaxI: case Op_MinI:
1992 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
1993 case Op_AndL: case Op_XorL: case Op_OrL:
1994 case Op_AndI: case Op_XorI: case Op_OrI: {
1995 // Move "last use" input to left by swapping inputs
1996 n->swap_edges(1, 2);
1997 break;
1998 }
1999 default:
2000 break;
2001 }
2002 }
2004 // Count FPU ops and common calls, implements item (3)
2005 switch( nop ) {
2006 // Count all float operations that may use FPU
2007 case Op_AddF:
2008 case Op_SubF:
2009 case Op_MulF:
2010 case Op_DivF:
2011 case Op_NegF:
2012 case Op_ModF:
2013 case Op_ConvI2F:
2014 case Op_ConF:
2015 case Op_CmpF:
2016 case Op_CmpF3:
2017 // case Op_ConvL2F: // longs are split into 32-bit halves
2018 frc.inc_float_count();
2019 break;
2021 case Op_ConvF2D:
2022 case Op_ConvD2F:
2023 frc.inc_float_count();
2024 frc.inc_double_count();
2025 break;
2027 // Count all double operations that may use FPU
2028 case Op_AddD:
2029 case Op_SubD:
2030 case Op_MulD:
2031 case Op_DivD:
2032 case Op_NegD:
2033 case Op_ModD:
2034 case Op_ConvI2D:
2035 case Op_ConvD2I:
2036 // case Op_ConvL2D: // handled by leaf call
2037 // case Op_ConvD2L: // handled by leaf call
2038 case Op_ConD:
2039 case Op_CmpD:
2040 case Op_CmpD3:
2041 frc.inc_double_count();
2042 break;
2043 case Op_Opaque1: // Remove Opaque Nodes before matching
2044 case Op_Opaque2: // Remove Opaque Nodes before matching
2045 n->subsume_by(n->in(1));
2046 break;
2047 case Op_CallStaticJava:
2048 case Op_CallJava:
2049 case Op_CallDynamicJava:
2050 frc.inc_java_call_count(); // Count java call site;
2051 case Op_CallRuntime:
2052 case Op_CallLeaf:
2053 case Op_CallLeafNoFP: {
2054 assert( n->is_Call(), "" );
2055 CallNode *call = n->as_Call();
2056 // Count call sites where the FP mode bit would have to be flipped.
2057 // Do not count uncommon runtime calls:
2058 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2059 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2060 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2061 frc.inc_call_count(); // Count the call site
2062 } else { // See if uncommon argument is shared
2063 Node *n = call->in(TypeFunc::Parms);
2064 int nop = n->Opcode();
2065 // Clone shared simple arguments to uncommon calls, item (1).
2066 if( n->outcnt() > 1 &&
2067 !n->is_Proj() &&
2068 nop != Op_CreateEx &&
2069 nop != Op_CheckCastPP &&
2070 nop != Op_DecodeN &&
2071 !n->is_Mem() ) {
2072 Node *x = n->clone();
2073 call->set_req( TypeFunc::Parms, x );
2074 }
2075 }
2076 break;
2077 }
2079 case Op_StoreD:
2080 case Op_LoadD:
2081 case Op_LoadD_unaligned:
2082 frc.inc_double_count();
2083 goto handle_mem;
2084 case Op_StoreF:
2085 case Op_LoadF:
2086 frc.inc_float_count();
2087 goto handle_mem;
2089 case Op_StoreB:
2090 case Op_StoreC:
2091 case Op_StoreCM:
2092 case Op_StorePConditional:
2093 case Op_StoreI:
2094 case Op_StoreL:
2095 case Op_StoreIConditional:
2096 case Op_StoreLConditional:
2097 case Op_CompareAndSwapI:
2098 case Op_CompareAndSwapL:
2099 case Op_CompareAndSwapP:
2100 case Op_CompareAndSwapN:
2101 case Op_StoreP:
2102 case Op_StoreN:
2103 case Op_LoadB:
2104 case Op_LoadUB:
2105 case Op_LoadUS:
2106 case Op_LoadI:
2107 case Op_LoadUI2L:
2108 case Op_LoadKlass:
2109 case Op_LoadNKlass:
2110 case Op_LoadL:
2111 case Op_LoadL_unaligned:
2112 case Op_LoadPLocked:
2113 case Op_LoadLLocked:
2114 case Op_LoadP:
2115 case Op_LoadN:
2116 case Op_LoadRange:
2117 case Op_LoadS: {
2118 handle_mem:
2119 #ifdef ASSERT
2120 if( VerifyOptoOopOffsets ) {
2121 assert( n->is_Mem(), "" );
2122 MemNode *mem = (MemNode*)n;
2123 // Check to see if address types have grounded out somehow.
2124 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2125 assert( !tp || oop_offset_is_sane(tp), "" );
2126 }
2127 #endif
2128 break;
2129 }
2131 case Op_AddP: { // Assert sane base pointers
2132 Node *addp = n->in(AddPNode::Address);
2133 assert( !addp->is_AddP() ||
2134 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2135 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2136 "Base pointers must match" );
2137 #ifdef _LP64
2138 if (UseCompressedOops &&
2139 addp->Opcode() == Op_ConP &&
2140 addp == n->in(AddPNode::Base) &&
2141 n->in(AddPNode::Offset)->is_Con()) {
2142 // Use addressing with narrow klass to load with offset on x86.
2143 // On sparc loading 32-bits constant and decoding it have less
2144 // instructions (4) then load 64-bits constant (7).
2145 // Do this transformation here since IGVN will convert ConN back to ConP.
2146 const Type* t = addp->bottom_type();
2147 if (t->isa_oopptr()) {
2148 Node* nn = NULL;
2150 // Look for existing ConN node of the same exact type.
2151 Compile* C = Compile::current();
2152 Node* r = C->root();
2153 uint cnt = r->outcnt();
2154 for (uint i = 0; i < cnt; i++) {
2155 Node* m = r->raw_out(i);
2156 if (m!= NULL && m->Opcode() == Op_ConN &&
2157 m->bottom_type()->make_ptr() == t) {
2158 nn = m;
2159 break;
2160 }
2161 }
2162 if (nn != NULL) {
2163 // Decode a narrow oop to match address
2164 // [R12 + narrow_oop_reg<<3 + offset]
2165 nn = new (C, 2) DecodeNNode(nn, t);
2166 n->set_req(AddPNode::Base, nn);
2167 n->set_req(AddPNode::Address, nn);
2168 if (addp->outcnt() == 0) {
2169 addp->disconnect_inputs(NULL);
2170 }
2171 }
2172 }
2173 }
2174 #endif
2175 break;
2176 }
2178 #ifdef _LP64
2179 case Op_CastPP:
2180 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) {
2181 Compile* C = Compile::current();
2182 Node* in1 = n->in(1);
2183 const Type* t = n->bottom_type();
2184 Node* new_in1 = in1->clone();
2185 new_in1->as_DecodeN()->set_type(t);
2187 if (!Matcher::clone_shift_expressions) {
2188 //
2189 // x86, ARM and friends can handle 2 adds in addressing mode
2190 // and Matcher can fold a DecodeN node into address by using
2191 // a narrow oop directly and do implicit NULL check in address:
2192 //
2193 // [R12 + narrow_oop_reg<<3 + offset]
2194 // NullCheck narrow_oop_reg
2195 //
2196 // On other platforms (Sparc) we have to keep new DecodeN node and
2197 // use it to do implicit NULL check in address:
2198 //
2199 // decode_not_null narrow_oop_reg, base_reg
2200 // [base_reg + offset]
2201 // NullCheck base_reg
2202 //
2203 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2204 // to keep the information to which NULL check the new DecodeN node
2205 // corresponds to use it as value in implicit_null_check().
2206 //
2207 new_in1->set_req(0, n->in(0));
2208 }
2210 n->subsume_by(new_in1);
2211 if (in1->outcnt() == 0) {
2212 in1->disconnect_inputs(NULL);
2213 }
2214 }
2215 break;
2217 case Op_CmpP:
2218 // Do this transformation here to preserve CmpPNode::sub() and
2219 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2220 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2221 Node* in1 = n->in(1);
2222 Node* in2 = n->in(2);
2223 if (!in1->is_DecodeN()) {
2224 in2 = in1;
2225 in1 = n->in(2);
2226 }
2227 assert(in1->is_DecodeN(), "sanity");
2229 Compile* C = Compile::current();
2230 Node* new_in2 = NULL;
2231 if (in2->is_DecodeN()) {
2232 new_in2 = in2->in(1);
2233 } else if (in2->Opcode() == Op_ConP) {
2234 const Type* t = in2->bottom_type();
2235 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) {
2236 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2237 //
2238 // This transformation together with CastPP transformation above
2239 // will generated code for implicit NULL checks for compressed oops.
2240 //
2241 // The original code after Optimize()
2242 //
2243 // LoadN memory, narrow_oop_reg
2244 // decode narrow_oop_reg, base_reg
2245 // CmpP base_reg, NULL
2246 // CastPP base_reg // NotNull
2247 // Load [base_reg + offset], val_reg
2248 //
2249 // after these transformations will be
2250 //
2251 // LoadN memory, narrow_oop_reg
2252 // CmpN narrow_oop_reg, NULL
2253 // decode_not_null narrow_oop_reg, base_reg
2254 // Load [base_reg + offset], val_reg
2255 //
2256 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2257 // since narrow oops can be used in debug info now (see the code in
2258 // final_graph_reshaping_walk()).
2259 //
2260 // At the end the code will be matched to
2261 // on x86:
2262 //
2263 // Load_narrow_oop memory, narrow_oop_reg
2264 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2265 // NullCheck narrow_oop_reg
2266 //
2267 // and on sparc:
2268 //
2269 // Load_narrow_oop memory, narrow_oop_reg
2270 // decode_not_null narrow_oop_reg, base_reg
2271 // Load [base_reg + offset], val_reg
2272 // NullCheck base_reg
2273 //
2274 } else if (t->isa_oopptr()) {
2275 new_in2 = ConNode::make(C, t->make_narrowoop());
2276 }
2277 }
2278 if (new_in2 != NULL) {
2279 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2280 n->subsume_by( cmpN );
2281 if (in1->outcnt() == 0) {
2282 in1->disconnect_inputs(NULL);
2283 }
2284 if (in2->outcnt() == 0) {
2285 in2->disconnect_inputs(NULL);
2286 }
2287 }
2288 }
2289 break;
2291 case Op_DecodeN:
2292 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2293 // DecodeN could be pinned on Sparc where it can't be fold into
2294 // an address expression, see the code for Op_CastPP above.
2295 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc");
2296 break;
2298 case Op_EncodeP: {
2299 Node* in1 = n->in(1);
2300 if (in1->is_DecodeN()) {
2301 n->subsume_by(in1->in(1));
2302 } else if (in1->Opcode() == Op_ConP) {
2303 Compile* C = Compile::current();
2304 const Type* t = in1->bottom_type();
2305 if (t == TypePtr::NULL_PTR) {
2306 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2307 } else if (t->isa_oopptr()) {
2308 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2309 }
2310 }
2311 if (in1->outcnt() == 0) {
2312 in1->disconnect_inputs(NULL);
2313 }
2314 break;
2315 }
2317 case Op_Proj: {
2318 if (OptimizeStringConcat) {
2319 ProjNode* p = n->as_Proj();
2320 if (p->_is_io_use) {
2321 // Separate projections were used for the exception path which
2322 // are normally removed by a late inline. If it wasn't inlined
2323 // then they will hang around and should just be replaced with
2324 // the original one.
2325 Node* proj = NULL;
2326 // Replace with just one
2327 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2328 Node *use = i.get();
2329 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2330 proj = use;
2331 break;
2332 }
2333 }
2334 assert(p != NULL, "must be found");
2335 p->subsume_by(proj);
2336 }
2337 }
2338 break;
2339 }
2341 case Op_Phi:
2342 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2343 // The EncodeP optimization may create Phi with the same edges
2344 // for all paths. It is not handled well by Register Allocator.
2345 Node* unique_in = n->in(1);
2346 assert(unique_in != NULL, "");
2347 uint cnt = n->req();
2348 for (uint i = 2; i < cnt; i++) {
2349 Node* m = n->in(i);
2350 assert(m != NULL, "");
2351 if (unique_in != m)
2352 unique_in = NULL;
2353 }
2354 if (unique_in != NULL) {
2355 n->subsume_by(unique_in);
2356 }
2357 }
2358 break;
2360 #endif
2362 case Op_ModI:
2363 if (UseDivMod) {
2364 // Check if a%b and a/b both exist
2365 Node* d = n->find_similar(Op_DivI);
2366 if (d) {
2367 // Replace them with a fused divmod if supported
2368 Compile* C = Compile::current();
2369 if (Matcher::has_match_rule(Op_DivModI)) {
2370 DivModINode* divmod = DivModINode::make(C, n);
2371 d->subsume_by(divmod->div_proj());
2372 n->subsume_by(divmod->mod_proj());
2373 } else {
2374 // replace a%b with a-((a/b)*b)
2375 Node* mult = new (C, 3) MulINode(d, d->in(2));
2376 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2377 n->subsume_by( sub );
2378 }
2379 }
2380 }
2381 break;
2383 case Op_ModL:
2384 if (UseDivMod) {
2385 // Check if a%b and a/b both exist
2386 Node* d = n->find_similar(Op_DivL);
2387 if (d) {
2388 // Replace them with a fused divmod if supported
2389 Compile* C = Compile::current();
2390 if (Matcher::has_match_rule(Op_DivModL)) {
2391 DivModLNode* divmod = DivModLNode::make(C, n);
2392 d->subsume_by(divmod->div_proj());
2393 n->subsume_by(divmod->mod_proj());
2394 } else {
2395 // replace a%b with a-((a/b)*b)
2396 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2397 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2398 n->subsume_by( sub );
2399 }
2400 }
2401 }
2402 break;
2404 case Op_Load16B:
2405 case Op_Load8B:
2406 case Op_Load4B:
2407 case Op_Load8S:
2408 case Op_Load4S:
2409 case Op_Load2S:
2410 case Op_Load8C:
2411 case Op_Load4C:
2412 case Op_Load2C:
2413 case Op_Load4I:
2414 case Op_Load2I:
2415 case Op_Load2L:
2416 case Op_Load4F:
2417 case Op_Load2F:
2418 case Op_Load2D:
2419 case Op_Store16B:
2420 case Op_Store8B:
2421 case Op_Store4B:
2422 case Op_Store8C:
2423 case Op_Store4C:
2424 case Op_Store2C:
2425 case Op_Store4I:
2426 case Op_Store2I:
2427 case Op_Store2L:
2428 case Op_Store4F:
2429 case Op_Store2F:
2430 case Op_Store2D:
2431 break;
2433 case Op_PackB:
2434 case Op_PackS:
2435 case Op_PackC:
2436 case Op_PackI:
2437 case Op_PackF:
2438 case Op_PackL:
2439 case Op_PackD:
2440 if (n->req()-1 > 2) {
2441 // Replace many operand PackNodes with a binary tree for matching
2442 PackNode* p = (PackNode*) n;
2443 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2444 n->subsume_by(btp);
2445 }
2446 break;
2447 case Op_Loop:
2448 case Op_CountedLoop:
2449 if (n->as_Loop()->is_inner_loop()) {
2450 frc.inc_inner_loop_count();
2451 }
2452 break;
2453 default:
2454 assert( !n->is_Call(), "" );
2455 assert( !n->is_Mem(), "" );
2456 break;
2457 }
2459 // Collect CFG split points
2460 if (n->is_MultiBranch())
2461 frc._tests.push(n);
2462 }
2464 //------------------------------final_graph_reshaping_walk---------------------
2465 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2466 // requires that the walk visits a node's inputs before visiting the node.
2467 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2468 ResourceArea *area = Thread::current()->resource_area();
2469 Unique_Node_List sfpt(area);
2471 frc._visited.set(root->_idx); // first, mark node as visited
2472 uint cnt = root->req();
2473 Node *n = root;
2474 uint i = 0;
2475 while (true) {
2476 if (i < cnt) {
2477 // Place all non-visited non-null inputs onto stack
2478 Node* m = n->in(i);
2479 ++i;
2480 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2481 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2482 sfpt.push(m);
2483 cnt = m->req();
2484 nstack.push(n, i); // put on stack parent and next input's index
2485 n = m;
2486 i = 0;
2487 }
2488 } else {
2489 // Now do post-visit work
2490 final_graph_reshaping_impl( n, frc );
2491 if (nstack.is_empty())
2492 break; // finished
2493 n = nstack.node(); // Get node from stack
2494 cnt = n->req();
2495 i = nstack.index();
2496 nstack.pop(); // Shift to the next node on stack
2497 }
2498 }
2500 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2501 // It could be done for an uncommon traps or any safepoints/calls
2502 // if the DecodeN node is referenced only in a debug info.
2503 while (sfpt.size() > 0) {
2504 n = sfpt.pop();
2505 JVMState *jvms = n->as_SafePoint()->jvms();
2506 assert(jvms != NULL, "sanity");
2507 int start = jvms->debug_start();
2508 int end = n->req();
2509 bool is_uncommon = (n->is_CallStaticJava() &&
2510 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2511 for (int j = start; j < end; j++) {
2512 Node* in = n->in(j);
2513 if (in->is_DecodeN()) {
2514 bool safe_to_skip = true;
2515 if (!is_uncommon ) {
2516 // Is it safe to skip?
2517 for (uint i = 0; i < in->outcnt(); i++) {
2518 Node* u = in->raw_out(i);
2519 if (!u->is_SafePoint() ||
2520 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2521 safe_to_skip = false;
2522 }
2523 }
2524 }
2525 if (safe_to_skip) {
2526 n->set_req(j, in->in(1));
2527 }
2528 if (in->outcnt() == 0) {
2529 in->disconnect_inputs(NULL);
2530 }
2531 }
2532 }
2533 }
2534 }
2536 //------------------------------final_graph_reshaping--------------------------
2537 // Final Graph Reshaping.
2538 //
2539 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2540 // and not commoned up and forced early. Must come after regular
2541 // optimizations to avoid GVN undoing the cloning. Clone constant
2542 // inputs to Loop Phis; these will be split by the allocator anyways.
2543 // Remove Opaque nodes.
2544 // (2) Move last-uses by commutative operations to the left input to encourage
2545 // Intel update-in-place two-address operations and better register usage
2546 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2547 // calls canonicalizing them back.
2548 // (3) Count the number of double-precision FP ops, single-precision FP ops
2549 // and call sites. On Intel, we can get correct rounding either by
2550 // forcing singles to memory (requires extra stores and loads after each
2551 // FP bytecode) or we can set a rounding mode bit (requires setting and
2552 // clearing the mode bit around call sites). The mode bit is only used
2553 // if the relative frequency of single FP ops to calls is low enough.
2554 // This is a key transform for SPEC mpeg_audio.
2555 // (4) Detect infinite loops; blobs of code reachable from above but not
2556 // below. Several of the Code_Gen algorithms fail on such code shapes,
2557 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2558 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2559 // Detection is by looking for IfNodes where only 1 projection is
2560 // reachable from below or CatchNodes missing some targets.
2561 // (5) Assert for insane oop offsets in debug mode.
2563 bool Compile::final_graph_reshaping() {
2564 // an infinite loop may have been eliminated by the optimizer,
2565 // in which case the graph will be empty.
2566 if (root()->req() == 1) {
2567 record_method_not_compilable("trivial infinite loop");
2568 return true;
2569 }
2571 Final_Reshape_Counts frc;
2573 // Visit everybody reachable!
2574 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2575 Node_Stack nstack(unique() >> 1);
2576 final_graph_reshaping_walk(nstack, root(), frc);
2578 // Check for unreachable (from below) code (i.e., infinite loops).
2579 for( uint i = 0; i < frc._tests.size(); i++ ) {
2580 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2581 // Get number of CFG targets.
2582 // Note that PCTables include exception targets after calls.
2583 uint required_outcnt = n->required_outcnt();
2584 if (n->outcnt() != required_outcnt) {
2585 // Check for a few special cases. Rethrow Nodes never take the
2586 // 'fall-thru' path, so expected kids is 1 less.
2587 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2588 if (n->in(0)->in(0)->is_Call()) {
2589 CallNode *call = n->in(0)->in(0)->as_Call();
2590 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2591 required_outcnt--; // Rethrow always has 1 less kid
2592 } else if (call->req() > TypeFunc::Parms &&
2593 call->is_CallDynamicJava()) {
2594 // Check for null receiver. In such case, the optimizer has
2595 // detected that the virtual call will always result in a null
2596 // pointer exception. The fall-through projection of this CatchNode
2597 // will not be populated.
2598 Node *arg0 = call->in(TypeFunc::Parms);
2599 if (arg0->is_Type() &&
2600 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2601 required_outcnt--;
2602 }
2603 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2604 call->req() > TypeFunc::Parms+1 &&
2605 call->is_CallStaticJava()) {
2606 // Check for negative array length. In such case, the optimizer has
2607 // detected that the allocation attempt will always result in an
2608 // exception. There is no fall-through projection of this CatchNode .
2609 Node *arg1 = call->in(TypeFunc::Parms+1);
2610 if (arg1->is_Type() &&
2611 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2612 required_outcnt--;
2613 }
2614 }
2615 }
2616 }
2617 // Recheck with a better notion of 'required_outcnt'
2618 if (n->outcnt() != required_outcnt) {
2619 record_method_not_compilable("malformed control flow");
2620 return true; // Not all targets reachable!
2621 }
2622 }
2623 // Check that I actually visited all kids. Unreached kids
2624 // must be infinite loops.
2625 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2626 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2627 record_method_not_compilable("infinite loop");
2628 return true; // Found unvisited kid; must be unreach
2629 }
2630 }
2632 // If original bytecodes contained a mixture of floats and doubles
2633 // check if the optimizer has made it homogenous, item (3).
2634 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2635 frc.get_float_count() > 32 &&
2636 frc.get_double_count() == 0 &&
2637 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2638 set_24_bit_selection_and_mode( false, true );
2639 }
2641 set_java_calls(frc.get_java_call_count());
2642 set_inner_loops(frc.get_inner_loop_count());
2644 // No infinite loops, no reason to bail out.
2645 return false;
2646 }
2648 //-----------------------------too_many_traps----------------------------------
2649 // Report if there are too many traps at the current method and bci.
2650 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2651 bool Compile::too_many_traps(ciMethod* method,
2652 int bci,
2653 Deoptimization::DeoptReason reason) {
2654 ciMethodData* md = method->method_data();
2655 if (md->is_empty()) {
2656 // Assume the trap has not occurred, or that it occurred only
2657 // because of a transient condition during start-up in the interpreter.
2658 return false;
2659 }
2660 if (md->has_trap_at(bci, reason) != 0) {
2661 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2662 // Also, if there are multiple reasons, or if there is no per-BCI record,
2663 // assume the worst.
2664 if (log())
2665 log()->elem("observe trap='%s' count='%d'",
2666 Deoptimization::trap_reason_name(reason),
2667 md->trap_count(reason));
2668 return true;
2669 } else {
2670 // Ignore method/bci and see if there have been too many globally.
2671 return too_many_traps(reason, md);
2672 }
2673 }
2675 // Less-accurate variant which does not require a method and bci.
2676 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2677 ciMethodData* logmd) {
2678 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2679 // Too many traps globally.
2680 // Note that we use cumulative trap_count, not just md->trap_count.
2681 if (log()) {
2682 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2683 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2684 Deoptimization::trap_reason_name(reason),
2685 mcount, trap_count(reason));
2686 }
2687 return true;
2688 } else {
2689 // The coast is clear.
2690 return false;
2691 }
2692 }
2694 //--------------------------too_many_recompiles--------------------------------
2695 // Report if there are too many recompiles at the current method and bci.
2696 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2697 // Is not eager to return true, since this will cause the compiler to use
2698 // Action_none for a trap point, to avoid too many recompilations.
2699 bool Compile::too_many_recompiles(ciMethod* method,
2700 int bci,
2701 Deoptimization::DeoptReason reason) {
2702 ciMethodData* md = method->method_data();
2703 if (md->is_empty()) {
2704 // Assume the trap has not occurred, or that it occurred only
2705 // because of a transient condition during start-up in the interpreter.
2706 return false;
2707 }
2708 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2709 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2710 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2711 Deoptimization::DeoptReason per_bc_reason
2712 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2713 if ((per_bc_reason == Deoptimization::Reason_none
2714 || md->has_trap_at(bci, reason) != 0)
2715 // The trap frequency measure we care about is the recompile count:
2716 && md->trap_recompiled_at(bci)
2717 && md->overflow_recompile_count() >= bc_cutoff) {
2718 // Do not emit a trap here if it has already caused recompilations.
2719 // Also, if there are multiple reasons, or if there is no per-BCI record,
2720 // assume the worst.
2721 if (log())
2722 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2723 Deoptimization::trap_reason_name(reason),
2724 md->trap_count(reason),
2725 md->overflow_recompile_count());
2726 return true;
2727 } else if (trap_count(reason) != 0
2728 && decompile_count() >= m_cutoff) {
2729 // Too many recompiles globally, and we have seen this sort of trap.
2730 // Use cumulative decompile_count, not just md->decompile_count.
2731 if (log())
2732 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2733 Deoptimization::trap_reason_name(reason),
2734 md->trap_count(reason), trap_count(reason),
2735 md->decompile_count(), decompile_count());
2736 return true;
2737 } else {
2738 // The coast is clear.
2739 return false;
2740 }
2741 }
2744 #ifndef PRODUCT
2745 //------------------------------verify_graph_edges---------------------------
2746 // Walk the Graph and verify that there is a one-to-one correspondence
2747 // between Use-Def edges and Def-Use edges in the graph.
2748 void Compile::verify_graph_edges(bool no_dead_code) {
2749 if (VerifyGraphEdges) {
2750 ResourceArea *area = Thread::current()->resource_area();
2751 Unique_Node_List visited(area);
2752 // Call recursive graph walk to check edges
2753 _root->verify_edges(visited);
2754 if (no_dead_code) {
2755 // Now make sure that no visited node is used by an unvisited node.
2756 bool dead_nodes = 0;
2757 Unique_Node_List checked(area);
2758 while (visited.size() > 0) {
2759 Node* n = visited.pop();
2760 checked.push(n);
2761 for (uint i = 0; i < n->outcnt(); i++) {
2762 Node* use = n->raw_out(i);
2763 if (checked.member(use)) continue; // already checked
2764 if (visited.member(use)) continue; // already in the graph
2765 if (use->is_Con()) continue; // a dead ConNode is OK
2766 // At this point, we have found a dead node which is DU-reachable.
2767 if (dead_nodes++ == 0)
2768 tty->print_cr("*** Dead nodes reachable via DU edges:");
2769 use->dump(2);
2770 tty->print_cr("---");
2771 checked.push(use); // No repeats; pretend it is now checked.
2772 }
2773 }
2774 assert(dead_nodes == 0, "using nodes must be reachable from root");
2775 }
2776 }
2777 }
2778 #endif
2780 // The Compile object keeps track of failure reasons separately from the ciEnv.
2781 // This is required because there is not quite a 1-1 relation between the
2782 // ciEnv and its compilation task and the Compile object. Note that one
2783 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2784 // to backtrack and retry without subsuming loads. Other than this backtracking
2785 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2786 // by the logic in C2Compiler.
2787 void Compile::record_failure(const char* reason) {
2788 if (log() != NULL) {
2789 log()->elem("failure reason='%s' phase='compile'", reason);
2790 }
2791 if (_failure_reason == NULL) {
2792 // Record the first failure reason.
2793 _failure_reason = reason;
2794 }
2795 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2796 C->print_method(_failure_reason);
2797 }
2798 _root = NULL; // flush the graph, too
2799 }
2801 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2802 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2803 {
2804 if (dolog) {
2805 C = Compile::current();
2806 _log = C->log();
2807 } else {
2808 C = NULL;
2809 _log = NULL;
2810 }
2811 if (_log != NULL) {
2812 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2813 _log->stamp();
2814 _log->end_head();
2815 }
2816 }
2818 Compile::TracePhase::~TracePhase() {
2819 if (_log != NULL) {
2820 _log->done("phase nodes='%d'", C->unique());
2821 }
2822 }