1 /*
2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vector.hpp"
71 #include "opto/vectornode.hpp"
72 #include "runtime/arguments.hpp"
73 #include "runtime/sharedRuntime.hpp"
74 #include "runtime/signature.hpp"
75 #include "runtime/stubRoutines.hpp"
76 #include "runtime/timer.hpp"
77 #include "utilities/align.hpp"
78 #include "utilities/copy.hpp"
79 #include "utilities/macros.hpp"
80
81
82 // -------------------- Compile::mach_constant_base_node -----------------------
83 // Constant table base node singleton.
84 MachConstantBaseNode* Compile::mach_constant_base_node() {
85 if (_mach_constant_base_node == NULL) {
86 _mach_constant_base_node = new MachConstantBaseNode();
87 _mach_constant_base_node->add_req(C->root());
88 }
89 return _mach_constant_base_node;
90 }
91
92
93 /// Support for intrinsics.
94
95 // Return the index at which m must be inserted (or already exists).
96 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
97 class IntrinsicDescPair {
98 private:
99 ciMethod* _m;
100 bool _is_virtual;
101 public:
102 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
103 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
104 ciMethod* m= elt->method();
105 ciMethod* key_m = key->_m;
106 if (key_m < m) return -1;
107 else if (key_m > m) return 1;
108 else {
109 bool is_virtual = elt->is_virtual();
110 bool key_virtual = key->_is_virtual;
111 if (key_virtual < is_virtual) return -1;
112 else if (key_virtual > is_virtual) return 1;
113 else return 0;
114 }
115 }
116 };
117 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
118 #ifdef ASSERT
119 for (int i = 1; i < _intrinsics->length(); i++) {
120 CallGenerator* cg1 = _intrinsics->at(i-1);
121 CallGenerator* cg2 = _intrinsics->at(i);
122 assert(cg1->method() != cg2->method()
123 ? cg1->method() < cg2->method()
124 : cg1->is_virtual() < cg2->is_virtual(),
125 "compiler intrinsics list must stay sorted");
126 }
127 #endif
128 IntrinsicDescPair pair(m, is_virtual);
129 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
130 }
131
132 void Compile::register_intrinsic(CallGenerator* cg) {
133 if (_intrinsics == NULL) {
134 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
135 }
136 int len = _intrinsics->length();
137 bool found = false;
138 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
139 assert(!found, "registering twice");
140 _intrinsics->insert_before(index, cg);
141 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
142 }
143
144 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
145 assert(m->is_loaded(), "don't try this on unloaded methods");
146 if (_intrinsics != NULL) {
147 bool found = false;
148 int index = intrinsic_insertion_index(m, is_virtual, found);
149 if (found) {
150 return _intrinsics->at(index);
151 }
152 }
153 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
154 if (m->intrinsic_id() != vmIntrinsics::_none &&
155 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
156 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
157 if (cg != NULL) {
158 // Save it for next time:
159 register_intrinsic(cg);
160 return cg;
161 } else {
162 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
163 }
164 }
165 return NULL;
166 }
167
168 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
169 // in library_call.cpp.
170
171
172 #ifndef PRODUCT
173 // statistics gathering...
174
175 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
176 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
177
178 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
179 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
180 int oflags = _intrinsic_hist_flags[id];
181 assert(flags != 0, "what happened?");
182 if (is_virtual) {
183 flags |= _intrinsic_virtual;
184 }
185 bool changed = (flags != oflags);
186 if ((flags & _intrinsic_worked) != 0) {
187 juint count = (_intrinsic_hist_count[id] += 1);
188 if (count == 1) {
189 changed = true; // first time
190 }
191 // increment the overall count also:
192 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
193 }
194 if (changed) {
195 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
196 // Something changed about the intrinsic's virtuality.
197 if ((flags & _intrinsic_virtual) != 0) {
198 // This is the first use of this intrinsic as a virtual call.
199 if (oflags != 0) {
200 // We already saw it as a non-virtual, so note both cases.
201 flags |= _intrinsic_both;
202 }
203 } else if ((oflags & _intrinsic_both) == 0) {
204 // This is the first use of this intrinsic as a non-virtual
205 flags |= _intrinsic_both;
206 }
207 }
208 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
209 }
210 // update the overall flags also:
211 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
212 return changed;
213 }
214
215 static char* format_flags(int flags, char* buf) {
216 buf[0] = 0;
217 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
218 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
219 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
220 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
221 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
222 if (buf[0] == 0) strcat(buf, ",");
223 assert(buf[0] == ',', "must be");
224 return &buf[1];
225 }
226
227 void Compile::print_intrinsic_statistics() {
228 char flagsbuf[100];
229 ttyLocker ttyl;
230 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
231 tty->print_cr("Compiler intrinsic usage:");
232 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
233 if (total == 0) total = 1; // avoid div0 in case of no successes
234 #define PRINT_STAT_LINE(name, c, f) \
235 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
236 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
237 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
238 int flags = _intrinsic_hist_flags[id];
239 juint count = _intrinsic_hist_count[id];
240 if ((flags | count) != 0) {
241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
242 }
243 }
244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
245 if (xtty != NULL) xtty->tail("statistics");
246 }
247
248 void Compile::print_statistics() {
249 { ttyLocker ttyl;
250 if (xtty != NULL) xtty->head("statistics type='opto'");
251 Parse::print_statistics();
252 PhaseCCP::print_statistics();
253 PhaseRegAlloc::print_statistics();
254 PhaseOutput::print_statistics();
255 PhasePeephole::print_statistics();
256 PhaseIdealLoop::print_statistics();
257 if (xtty != NULL) xtty->tail("statistics");
258 }
259 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
260 // put this under its own <statistics> element.
261 print_intrinsic_statistics();
262 }
263 }
264 #endif //PRODUCT
265
266 void Compile::gvn_replace_by(Node* n, Node* nn) {
267 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
268 Node* use = n->last_out(i);
269 bool is_in_table = initial_gvn()->hash_delete(use);
270 uint uses_found = 0;
271 for (uint j = 0; j < use->len(); j++) {
272 if (use->in(j) == n) {
273 if (j < use->req())
274 use->set_req(j, nn);
275 else
276 use->set_prec(j, nn);
277 uses_found++;
278 }
279 }
280 if (is_in_table) {
281 // reinsert into table
282 initial_gvn()->hash_find_insert(use);
283 }
284 record_for_igvn(use);
285 i -= uses_found; // we deleted 1 or more copies of this edge
286 }
287 }
288
289
290 static inline bool not_a_node(const Node* n) {
291 if (n == NULL) return true;
292 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
293 if (*(address*)n == badAddress) return true; // kill by Node::destruct
294 return false;
295 }
296
297 // Identify all nodes that are reachable from below, useful.
298 // Use breadth-first pass that records state in a Unique_Node_List,
299 // recursive traversal is slower.
300 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
301 int estimated_worklist_size = live_nodes();
302 useful.map( estimated_worklist_size, NULL ); // preallocate space
303
304 // Initialize worklist
305 if (root() != NULL) { useful.push(root()); }
306 // If 'top' is cached, declare it useful to preserve cached node
307 if( cached_top_node() ) { useful.push(cached_top_node()); }
308
309 // Push all useful nodes onto the list, breadthfirst
310 for( uint next = 0; next < useful.size(); ++next ) {
311 assert( next < unique(), "Unique useful nodes < total nodes");
312 Node *n = useful.at(next);
313 uint max = n->len();
314 for( uint i = 0; i < max; ++i ) {
315 Node *m = n->in(i);
316 if (not_a_node(m)) continue;
317 useful.push(m);
318 }
319 }
320 }
321
322 // Update dead_node_list with any missing dead nodes using useful
323 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
324 void Compile::update_dead_node_list(Unique_Node_List &useful) {
325 uint max_idx = unique();
326 VectorSet& useful_node_set = useful.member_set();
327
328 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
329 // If node with index node_idx is not in useful set,
330 // mark it as dead in dead node list.
331 if (!useful_node_set.test(node_idx)) {
332 record_dead_node(node_idx);
333 }
334 }
335 }
336
337 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
338 int shift = 0;
339 for (int i = 0; i < inlines->length(); i++) {
340 CallGenerator* cg = inlines->at(i);
341 CallNode* call = cg->call_node();
342 if (shift > 0) {
343 inlines->at_put(i-shift, cg);
344 }
345 if (!useful.member(call)) {
346 shift++;
347 }
348 }
349 inlines->trunc_to(inlines->length()-shift);
350 }
351
352 // Disconnect all useless nodes by disconnecting those at the boundary.
353 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
354 uint next = 0;
355 while (next < useful.size()) {
356 Node *n = useful.at(next++);
357 if (n->is_SafePoint()) {
358 // We're done with a parsing phase. Replaced nodes are not valid
359 // beyond that point.
360 n->as_SafePoint()->delete_replaced_nodes();
361 }
362 // Use raw traversal of out edges since this code removes out edges
363 int max = n->outcnt();
364 for (int j = 0; j < max; ++j) {
365 Node* child = n->raw_out(j);
366 if (! useful.member(child)) {
367 assert(!child->is_top() || child != top(),
368 "If top is cached in Compile object it is in useful list");
369 // Only need to remove this out-edge to the useless node
370 n->raw_del_out(j);
371 --j;
372 --max;
373 }
374 }
375 if (n->outcnt() == 1 && n->has_special_unique_user()) {
376 record_for_igvn(n->unique_out());
377 }
378 }
379 // Remove useless macro and predicate opaq nodes
380 for (int i = C->macro_count()-1; i >= 0; i--) {
381 Node* n = C->macro_node(i);
382 if (!useful.member(n)) {
383 remove_macro_node(n);
384 }
385 }
386 // Remove useless CastII nodes with range check dependency
387 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
388 Node* cast = range_check_cast_node(i);
389 if (!useful.member(cast)) {
390 remove_range_check_cast(cast);
391 }
392 }
393 // Remove useless expensive nodes
394 for (int i = C->expensive_count()-1; i >= 0; i--) {
395 Node* n = C->expensive_node(i);
396 if (!useful.member(n)) {
397 remove_expensive_node(n);
398 }
399 }
400 // Remove useless Opaque4 nodes
401 for (int i = opaque4_count() - 1; i >= 0; i--) {
402 Node* opaq = opaque4_node(i);
403 if (!useful.member(opaq)) {
404 remove_opaque4_node(opaq);
405 }
406 }
407 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
408 bs->eliminate_useless_gc_barriers(useful, this);
409 // clean up the late inline lists
410 remove_useless_late_inlines(&_string_late_inlines, useful);
411 remove_useless_late_inlines(&_boxing_late_inlines, useful);
412 remove_useless_late_inlines(&_late_inlines, useful);
413 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
414 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
415 }
416
417 // ============================================================================
418 //------------------------------CompileWrapper---------------------------------
419 class CompileWrapper : public StackObj {
420 Compile *const _compile;
421 public:
422 CompileWrapper(Compile* compile);
423
424 ~CompileWrapper();
425 };
426
427 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
428 // the Compile* pointer is stored in the current ciEnv:
429 ciEnv* env = compile->env();
430 assert(env == ciEnv::current(), "must already be a ciEnv active");
431 assert(env->compiler_data() == NULL, "compile already active?");
432 env->set_compiler_data(compile);
433 assert(compile == Compile::current(), "sanity");
434
435 compile->set_type_dict(NULL);
436 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
437 compile->clone_map().set_clone_idx(0);
438 compile->set_type_last_size(0);
439 compile->set_last_tf(NULL, NULL);
440 compile->set_indexSet_arena(NULL);
441 compile->set_indexSet_free_block_list(NULL);
442 compile->init_type_arena();
443 Type::Initialize(compile);
444 _compile->begin_method();
445 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
446 }
447 CompileWrapper::~CompileWrapper() {
448 _compile->end_method();
449 _compile->env()->set_compiler_data(NULL);
450 }
451
452
453 //----------------------------print_compile_messages---------------------------
454 void Compile::print_compile_messages() {
455 #ifndef PRODUCT
456 // Check if recompiling
457 if (_subsume_loads == false && PrintOpto) {
458 // Recompiling without allowing machine instructions to subsume loads
459 tty->print_cr("*********************************************************");
460 tty->print_cr("** Bailout: Recompile without subsuming loads **");
461 tty->print_cr("*********************************************************");
462 }
463 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
464 // Recompiling without escape analysis
465 tty->print_cr("*********************************************************");
466 tty->print_cr("** Bailout: Recompile without escape analysis **");
467 tty->print_cr("*********************************************************");
468 }
469 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
470 // Recompiling without boxing elimination
471 tty->print_cr("*********************************************************");
472 tty->print_cr("** Bailout: Recompile without boxing elimination **");
473 tty->print_cr("*********************************************************");
474 }
475 if (C->directive()->BreakAtCompileOption) {
476 // Open the debugger when compiling this method.
477 tty->print("### Breaking when compiling: ");
478 method()->print_short_name();
479 tty->cr();
480 BREAKPOINT;
481 }
482
483 if( PrintOpto ) {
484 if (is_osr_compilation()) {
485 tty->print("[OSR]%3d", _compile_id);
486 } else {
487 tty->print("%3d", _compile_id);
488 }
489 }
490 #endif
491 }
492
493 // ============================================================================
494 //------------------------------Compile standard-------------------------------
495 debug_only( int Compile::_debug_idx = 100000; )
496
497 // Compile a method. entry_bci is -1 for normal compilations and indicates
498 // the continuation bci for on stack replacement.
499
500
501 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
502 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
503 : Phase(Compiler),
504 _compile_id(ci_env->compile_id()),
505 _save_argument_registers(false),
506 _subsume_loads(subsume_loads),
507 _do_escape_analysis(do_escape_analysis),
508 _eliminate_boxing(eliminate_boxing),
509 _method(target),
510 _entry_bci(osr_bci),
511 _stub_function(NULL),
512 _stub_name(NULL),
513 _stub_entry_point(NULL),
514 _max_node_limit(MaxNodeLimit),
515 _inlining_progress(false),
516 _inlining_incrementally(false),
517 _do_cleanup(false),
518 _has_reserved_stack_access(target->has_reserved_stack_access()),
519 #ifndef PRODUCT
520 _trace_opto_output(directive->TraceOptoOutputOption),
521 _print_ideal(directive->PrintIdealOption),
522 #endif
523 _has_method_handle_invokes(false),
524 _clinit_barrier_on_entry(false),
525 _comp_arena(mtCompiler),
526 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
527 _env(ci_env),
528 _directive(directive),
529 _log(ci_env->log()),
530 _failure_reason(NULL),
531 _congraph(NULL),
532 #ifndef PRODUCT
533 _printer(IdealGraphPrinter::printer()),
534 #endif
535 _dead_node_list(comp_arena()),
536 _dead_node_count(0),
537 _node_arena(mtCompiler),
538 _old_arena(mtCompiler),
539 _mach_constant_base_node(NULL),
540 _Compile_types(mtCompiler),
541 _initial_gvn(NULL),
542 _for_igvn(NULL),
543 _warm_calls(NULL),
544 _late_inlines(comp_arena(), 2, 0, NULL),
545 _string_late_inlines(comp_arena(), 2, 0, NULL),
546 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
547 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
548 _late_inlines_pos(0),
549 _number_of_mh_late_inlines(0),
550 _print_inlining_stream(NULL),
551 _print_inlining_list(NULL),
552 _print_inlining_idx(0),
553 _print_inlining_output(NULL),
554 _replay_inline_data(NULL),
555 _java_calls(0),
556 _inner_loops(0),
557 _interpreter_frame_size(0)
558 #ifndef PRODUCT
559 , _in_dump_cnt(0)
560 #endif
561 {
562 C = this;
563 #ifndef PRODUCT
564 if (_printer != NULL) {
565 _printer->set_compile(this);
566 }
567 #endif
568 CompileWrapper cw(this);
569
570 if (CITimeVerbose) {
571 tty->print(" ");
572 target->holder()->name()->print();
573 tty->print(".");
574 target->print_short_name();
575 tty->print(" ");
576 }
577 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
578 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
579
580 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
581 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
582 // We can always print a disassembly, either abstract (hex dump) or
583 // with the help of a suitable hsdis library. Thus, we should not
584 // couple print_assembly and print_opto_assembly controls.
585 // But: always print opto and regular assembly on compile command 'print'.
586 bool print_assembly = directive->PrintAssemblyOption;
587 set_print_assembly(print_opto_assembly || print_assembly);
588 #else
589 set_print_assembly(false); // must initialize.
590 #endif
591
592 #ifndef PRODUCT
593 set_parsed_irreducible_loop(false);
594
595 if (directive->ReplayInlineOption) {
596 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
597 }
598 #endif
599 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
600 set_print_intrinsics(directive->PrintIntrinsicsOption);
601 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
602
603 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
604 // Make sure the method being compiled gets its own MDO,
605 // so we can at least track the decompile_count().
606 // Need MDO to record RTM code generation state.
607 method()->ensure_method_data();
608 }
609
610 Init(::AliasLevel);
611
612
613 print_compile_messages();
614
615 _ilt = InlineTree::build_inline_tree_root();
616
617 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
618 assert(num_alias_types() >= AliasIdxRaw, "");
619
620 #define MINIMUM_NODE_HASH 1023
621 // Node list that Iterative GVN will start with
622 Unique_Node_List for_igvn(comp_arena());
623 set_for_igvn(&for_igvn);
624
625 // GVN that will be run immediately on new nodes
626 uint estimated_size = method()->code_size()*4+64;
627 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
628 PhaseGVN gvn(node_arena(), estimated_size);
629 set_initial_gvn(&gvn);
630
631 print_inlining_init();
632 { // Scope for timing the parser
633 TracePhase tp("parse", &timers[_t_parser]);
634
635 // Put top into the hash table ASAP.
636 initial_gvn()->transform_no_reclaim(top());
637
638 // Set up tf(), start(), and find a CallGenerator.
639 CallGenerator* cg = NULL;
640 if (is_osr_compilation()) {
641 const TypeTuple *domain = StartOSRNode::osr_domain();
642 const TypeTuple *range = TypeTuple::make_range(method()->signature());
643 init_tf(TypeFunc::make(domain, range));
644 StartNode* s = new StartOSRNode(root(), domain);
645 initial_gvn()->set_type_bottom(s);
646 init_start(s);
647 cg = CallGenerator::for_osr(method(), entry_bci());
648 } else {
649 // Normal case.
650 init_tf(TypeFunc::make(method()));
651 StartNode* s = new StartNode(root(), tf()->domain());
652 initial_gvn()->set_type_bottom(s);
653 init_start(s);
654 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
655 // With java.lang.ref.reference.get() we must go through the
656 // intrinsic - even when get() is the root
657 // method of the compile - so that, if necessary, the value in
658 // the referent field of the reference object gets recorded by
659 // the pre-barrier code.
660 cg = find_intrinsic(method(), false);
661 }
662 if (cg == NULL) {
663 float past_uses = method()->interpreter_invocation_count();
664 float expected_uses = past_uses;
665 cg = CallGenerator::for_inline(method(), expected_uses);
666 }
667 }
668 if (failing()) return;
669 if (cg == NULL) {
670 record_method_not_compilable("cannot parse method");
671 return;
672 }
673 JVMState* jvms = build_start_state(start(), tf());
674 if ((jvms = cg->generate(jvms)) == NULL) {
675 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
676 record_method_not_compilable("method parse failed");
677 }
678 return;
679 }
680 GraphKit kit(jvms);
681
682 if (!kit.stopped()) {
683 // Accept return values, and transfer control we know not where.
684 // This is done by a special, unique ReturnNode bound to root.
685 return_values(kit.jvms());
686 }
687
688 if (kit.has_exceptions()) {
689 // Any exceptions that escape from this call must be rethrown
690 // to whatever caller is dynamically above us on the stack.
691 // This is done by a special, unique RethrowNode bound to root.
692 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
693 }
694
695 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
696
697 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
698 inline_string_calls(true);
699 }
700
701 if (failing()) return;
702
703 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
704
705 // Remove clutter produced by parsing.
706 if (!failing()) {
707 ResourceMark rm;
708 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
709 }
710 }
711
712 // Note: Large methods are capped off in do_one_bytecode().
713 if (failing()) return;
714
715 // After parsing, node notes are no longer automagic.
716 // They must be propagated by register_new_node_with_optimizer(),
717 // clone(), or the like.
718 set_default_node_notes(NULL);
719
720 for (;;) {
721 int successes = Inline_Warm();
722 if (failing()) return;
723 if (successes == 0) break;
724 }
725
726 // Drain the list.
727 Finish_Warm();
728 #ifndef PRODUCT
729 if (_printer && _printer->should_print(1)) {
730 _printer->print_inlining();
731 }
732 #endif
733
734 if (failing()) return;
735 NOT_PRODUCT( verify_graph_edges(); )
736
737 // Now optimize
738 Optimize();
739 if (failing()) return;
740 NOT_PRODUCT( verify_graph_edges(); )
741
742 #ifndef PRODUCT
743 if (print_ideal()) {
744 ttyLocker ttyl; // keep the following output all in one block
745 // This output goes directly to the tty, not the compiler log.
746 // To enable tools to match it up with the compilation activity,
747 // be sure to tag this tty output with the compile ID.
748 if (xtty != NULL) {
749 xtty->head("ideal compile_id='%d'%s", compile_id(),
750 is_osr_compilation() ? " compile_kind='osr'" :
751 "");
752 }
753 root()->dump(9999);
754 if (xtty != NULL) {
755 xtty->tail("ideal");
756 }
757 }
758 #endif
759
760 #ifdef ASSERT
761 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
762 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
763 #endif
764
765 // Dump compilation data to replay it.
766 if (directive->DumpReplayOption) {
767 env()->dump_replay_data(_compile_id);
768 }
769 if (directive->DumpInlineOption && (ilt() != NULL)) {
770 env()->dump_inline_data(_compile_id);
771 }
772
773 // Now that we know the size of all the monitors we can add a fixed slot
774 // for the original deopt pc.
775 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
776 set_fixed_slots(next_slot);
777
778 // Compute when to use implicit null checks. Used by matching trap based
779 // nodes and NullCheck optimization.
780 set_allowed_deopt_reasons();
781
782 // Now generate code
783 Code_Gen();
784 }
785
786 //------------------------------Compile----------------------------------------
787 // Compile a runtime stub
788 Compile::Compile( ciEnv* ci_env,
789 TypeFunc_generator generator,
790 address stub_function,
791 const char *stub_name,
792 int is_fancy_jump,
793 bool pass_tls,
794 bool save_arg_registers,
795 bool return_pc,
796 DirectiveSet* directive)
797 : Phase(Compiler),
798 _compile_id(0),
799 _save_argument_registers(save_arg_registers),
800 _subsume_loads(true),
801 _do_escape_analysis(false),
802 _eliminate_boxing(false),
803 _method(NULL),
804 _entry_bci(InvocationEntryBci),
805 _stub_function(stub_function),
806 _stub_name(stub_name),
807 _stub_entry_point(NULL),
808 _max_node_limit(MaxNodeLimit),
809 _inlining_progress(false),
810 _inlining_incrementally(false),
811 _has_reserved_stack_access(false),
812 #ifndef PRODUCT
813 _trace_opto_output(directive->TraceOptoOutputOption),
814 _print_ideal(directive->PrintIdealOption),
815 #endif
816 _has_method_handle_invokes(false),
817 _clinit_barrier_on_entry(false),
818 _comp_arena(mtCompiler),
819 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
820 _env(ci_env),
821 _directive(directive),
822 _log(ci_env->log()),
823 _failure_reason(NULL),
824 _congraph(NULL),
825 #ifndef PRODUCT
826 _printer(NULL),
827 #endif
828 _dead_node_list(comp_arena()),
829 _dead_node_count(0),
830 _node_arena(mtCompiler),
831 _old_arena(mtCompiler),
832 _mach_constant_base_node(NULL),
833 _Compile_types(mtCompiler),
834 _initial_gvn(NULL),
835 _for_igvn(NULL),
836 _warm_calls(NULL),
837 _number_of_mh_late_inlines(0),
838 _print_inlining_stream(NULL),
839 _print_inlining_list(NULL),
840 _print_inlining_idx(0),
841 _print_inlining_output(NULL),
842 _replay_inline_data(NULL),
843 _java_calls(0),
844 _inner_loops(0),
845 _interpreter_frame_size(0),
846 #ifndef PRODUCT
847 _in_dump_cnt(0),
848 #endif
849 _allowed_reasons(0) {
850 C = this;
851
852 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
853 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
854
855 #ifndef PRODUCT
856 set_print_assembly(PrintFrameConverterAssembly);
857 set_parsed_irreducible_loop(false);
858 #else
859 set_print_assembly(false); // Must initialize.
860 #endif
861 set_has_irreducible_loop(false); // no loops
862
863 CompileWrapper cw(this);
864 Init(/*AliasLevel=*/ 0);
865 init_tf((*generator)());
866
867 {
868 // The following is a dummy for the sake of GraphKit::gen_stub
869 Unique_Node_List for_igvn(comp_arena());
870 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
871 PhaseGVN gvn(Thread::current()->resource_area(),255);
872 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
873 gvn.transform_no_reclaim(top());
874
875 GraphKit kit;
876 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
877 }
878
879 NOT_PRODUCT( verify_graph_edges(); )
880
881 Code_Gen();
882 }
883
884 //------------------------------Init-------------------------------------------
885 // Prepare for a single compilation
886 void Compile::Init(int aliaslevel) {
887 _unique = 0;
888 _regalloc = NULL;
889
890 _tf = NULL; // filled in later
891 _top = NULL; // cached later
892 _matcher = NULL; // filled in later
893 _cfg = NULL; // filled in later
894
895 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
896
897 _node_note_array = NULL;
898 _default_node_notes = NULL;
899 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
900
901 _immutable_memory = NULL; // filled in at first inquiry
902
903 // Globally visible Nodes
904 // First set TOP to NULL to give safe behavior during creation of RootNode
905 set_cached_top_node(NULL);
906 set_root(new RootNode());
907 // Now that you have a Root to point to, create the real TOP
908 set_cached_top_node( new ConNode(Type::TOP) );
909 set_recent_alloc(NULL, NULL);
910
911 // Create Debug Information Recorder to record scopes, oopmaps, etc.
912 env()->set_oop_recorder(new OopRecorder(env()->arena()));
913 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
914 env()->set_dependencies(new Dependencies(env()));
915
916 _fixed_slots = 0;
917 set_has_split_ifs(false);
918 set_has_loops(has_method() && method()->has_loops()); // first approximation
919 set_has_stringbuilder(false);
920 set_has_boxed_value(false);
921 _trap_can_recompile = false; // no traps emitted yet
922 _major_progress = true; // start out assuming good things will happen
923 set_has_unsafe_access(false);
924 set_max_vector_size(0);
925 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
926 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
927 set_decompile_count(0);
928
929 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
930 _loop_opts_cnt = LoopOptsCount;
931 set_do_inlining(Inline);
932 set_max_inline_size(MaxInlineSize);
933 set_freq_inline_size(FreqInlineSize);
934 set_do_scheduling(OptoScheduling);
935 set_do_count_invocations(false);
936 set_do_method_data_update(false);
937
938 set_do_vector_loop(false);
939
940 if (AllowVectorizeOnDemand) {
941 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
942 set_do_vector_loop(true);
943 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
944 } else if (has_method() && method()->name() != 0 &&
945 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
946 set_do_vector_loop(true);
947 }
948 }
949 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
950 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
951
952 set_age_code(has_method() && method()->profile_aging());
953 set_rtm_state(NoRTM); // No RTM lock eliding by default
954 _max_node_limit = _directive->MaxNodeLimitOption;
955
956 #if INCLUDE_RTM_OPT
957 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
958 int rtm_state = method()->method_data()->rtm_state();
959 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
960 // Don't generate RTM lock eliding code.
961 set_rtm_state(NoRTM);
962 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
963 // Generate RTM lock eliding code without abort ratio calculation code.
964 set_rtm_state(UseRTM);
965 } else if (UseRTMDeopt) {
966 // Generate RTM lock eliding code and include abort ratio calculation
967 // code if UseRTMDeopt is on.
968 set_rtm_state(ProfileRTM);
969 }
970 }
971 #endif
972 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
973 set_clinit_barrier_on_entry(true);
974 }
975 if (debug_info()->recording_non_safepoints()) {
976 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
977 (comp_arena(), 8, 0, NULL));
978 set_default_node_notes(Node_Notes::make(this));
979 }
980
981 // // -- Initialize types before each compile --
982 // // Update cached type information
983 // if( _method && _method->constants() )
984 // Type::update_loaded_types(_method, _method->constants());
985
986 // Init alias_type map.
987 if (!_do_escape_analysis && aliaslevel == 3)
988 aliaslevel = 2; // No unique types without escape analysis
989 _AliasLevel = aliaslevel;
990 const int grow_ats = 16;
991 _max_alias_types = grow_ats;
992 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
993 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
994 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
995 {
996 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
997 }
998 // Initialize the first few types.
999 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1000 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1001 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1002 _num_alias_types = AliasIdxRaw+1;
1003 // Zero out the alias type cache.
1004 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1005 // A NULL adr_type hits in the cache right away. Preload the right answer.
1006 probe_alias_cache(NULL)->_index = AliasIdxTop;
1007
1008 _intrinsics = NULL;
1009 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1010 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1011 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1012 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1013 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1014 register_library_intrinsics();
1015 #ifdef ASSERT
1016 _type_verify_symmetry = true;
1017 #endif
1018 }
1019
1020 //---------------------------init_start----------------------------------------
1021 // Install the StartNode on this compile object.
1022 void Compile::init_start(StartNode* s) {
1023 if (failing())
1024 return; // already failing
1025 assert(s == start(), "");
1026 }
1027
1028 /**
1029 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1030 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1031 * the ideal graph.
1032 */
1033 StartNode* Compile::start() const {
1034 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1035 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1036 Node* start = root()->fast_out(i);
1037 if (start->is_Start()) {
1038 return start->as_Start();
1039 }
1040 }
1041 fatal("Did not find Start node!");
1042 return NULL;
1043 }
1044
1045 //-------------------------------immutable_memory-------------------------------------
1046 // Access immutable memory
1047 Node* Compile::immutable_memory() {
1048 if (_immutable_memory != NULL) {
1049 return _immutable_memory;
1050 }
1051 StartNode* s = start();
1052 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1053 Node *p = s->fast_out(i);
1054 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1055 _immutable_memory = p;
1056 return _immutable_memory;
1057 }
1058 }
1059 ShouldNotReachHere();
1060 return NULL;
1061 }
1062
1063 //----------------------set_cached_top_node------------------------------------
1064 // Install the cached top node, and make sure Node::is_top works correctly.
1065 void Compile::set_cached_top_node(Node* tn) {
1066 if (tn != NULL) verify_top(tn);
1067 Node* old_top = _top;
1068 _top = tn;
1069 // Calling Node::setup_is_top allows the nodes the chance to adjust
1070 // their _out arrays.
1071 if (_top != NULL) _top->setup_is_top();
1072 if (old_top != NULL) old_top->setup_is_top();
1073 assert(_top == NULL || top()->is_top(), "");
1074 }
1075
1076 #ifdef ASSERT
1077 uint Compile::count_live_nodes_by_graph_walk() {
1078 Unique_Node_List useful(comp_arena());
1079 // Get useful node list by walking the graph.
1080 identify_useful_nodes(useful);
1081 return useful.size();
1082 }
1083
1084 void Compile::print_missing_nodes() {
1085
1086 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1087 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1088 return;
1089 }
1090
1091 // This is an expensive function. It is executed only when the user
1092 // specifies VerifyIdealNodeCount option or otherwise knows the
1093 // additional work that needs to be done to identify reachable nodes
1094 // by walking the flow graph and find the missing ones using
1095 // _dead_node_list.
1096
1097 Unique_Node_List useful(comp_arena());
1098 // Get useful node list by walking the graph.
1099 identify_useful_nodes(useful);
1100
1101 uint l_nodes = C->live_nodes();
1102 uint l_nodes_by_walk = useful.size();
1103
1104 if (l_nodes != l_nodes_by_walk) {
1105 if (_log != NULL) {
1106 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1107 _log->stamp();
1108 _log->end_head();
1109 }
1110 VectorSet& useful_member_set = useful.member_set();
1111 int last_idx = l_nodes_by_walk;
1112 for (int i = 0; i < last_idx; i++) {
1113 if (useful_member_set.test(i)) {
1114 if (_dead_node_list.test(i)) {
1115 if (_log != NULL) {
1116 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1117 }
1118 if (PrintIdealNodeCount) {
1119 // Print the log message to tty
1120 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1121 useful.at(i)->dump();
1122 }
1123 }
1124 }
1125 else if (! _dead_node_list.test(i)) {
1126 if (_log != NULL) {
1127 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1128 }
1129 if (PrintIdealNodeCount) {
1130 // Print the log message to tty
1131 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1132 }
1133 }
1134 }
1135 if (_log != NULL) {
1136 _log->tail("mismatched_nodes");
1137 }
1138 }
1139 }
1140 void Compile::record_modified_node(Node* n) {
1141 if (_modified_nodes != NULL && !_inlining_incrementally &&
1142 n->outcnt() != 0 && !n->is_Con()) {
1143 _modified_nodes->push(n);
1144 }
1145 }
1146
1147 void Compile::remove_modified_node(Node* n) {
1148 if (_modified_nodes != NULL) {
1149 _modified_nodes->remove(n);
1150 }
1151 }
1152 #endif
1153
1154 #ifndef PRODUCT
1155 void Compile::verify_top(Node* tn) const {
1156 if (tn != NULL) {
1157 assert(tn->is_Con(), "top node must be a constant");
1158 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1159 assert(tn->in(0) != NULL, "must have live top node");
1160 }
1161 }
1162 #endif
1163
1164
1165 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1166
1167 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1168 guarantee(arr != NULL, "");
1169 int num_blocks = arr->length();
1170 if (grow_by < num_blocks) grow_by = num_blocks;
1171 int num_notes = grow_by * _node_notes_block_size;
1172 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1173 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1174 while (num_notes > 0) {
1175 arr->append(notes);
1176 notes += _node_notes_block_size;
1177 num_notes -= _node_notes_block_size;
1178 }
1179 assert(num_notes == 0, "exact multiple, please");
1180 }
1181
1182 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1183 if (source == NULL || dest == NULL) return false;
1184
1185 if (dest->is_Con())
1186 return false; // Do not push debug info onto constants.
1187
1188 #ifdef ASSERT
1189 // Leave a bread crumb trail pointing to the original node:
1190 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1191 dest->set_debug_orig(source);
1192 }
1193 #endif
1194
1195 if (node_note_array() == NULL)
1196 return false; // Not collecting any notes now.
1197
1198 // This is a copy onto a pre-existing node, which may already have notes.
1199 // If both nodes have notes, do not overwrite any pre-existing notes.
1200 Node_Notes* source_notes = node_notes_at(source->_idx);
1201 if (source_notes == NULL || source_notes->is_clear()) return false;
1202 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1203 if (dest_notes == NULL || dest_notes->is_clear()) {
1204 return set_node_notes_at(dest->_idx, source_notes);
1205 }
1206
1207 Node_Notes merged_notes = (*source_notes);
1208 // The order of operations here ensures that dest notes will win...
1209 merged_notes.update_from(dest_notes);
1210 return set_node_notes_at(dest->_idx, &merged_notes);
1211 }
1212
1213
1214 //--------------------------allow_range_check_smearing-------------------------
1215 // Gating condition for coalescing similar range checks.
1216 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1217 // single covering check that is at least as strong as any of them.
1218 // If the optimization succeeds, the simplified (strengthened) range check
1219 // will always succeed. If it fails, we will deopt, and then give up
1220 // on the optimization.
1221 bool Compile::allow_range_check_smearing() const {
1222 // If this method has already thrown a range-check,
1223 // assume it was because we already tried range smearing
1224 // and it failed.
1225 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1226 return !already_trapped;
1227 }
1228
1229
1230 //------------------------------flatten_alias_type-----------------------------
1231 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1232 int offset = tj->offset();
1233 TypePtr::PTR ptr = tj->ptr();
1234
1235 // Known instance (scalarizable allocation) alias only with itself.
1236 bool is_known_inst = tj->isa_oopptr() != NULL &&
1237 tj->is_oopptr()->is_known_instance();
1238
1239 // Process weird unsafe references.
1240 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1241 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1242 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1243 tj = TypeOopPtr::BOTTOM;
1244 ptr = tj->ptr();
1245 offset = tj->offset();
1246 }
1247
1248 // Array pointers need some flattening
1249 const TypeAryPtr *ta = tj->isa_aryptr();
1250 if (ta && ta->is_stable()) {
1251 // Erase stability property for alias analysis.
1252 tj = ta = ta->cast_to_stable(false);
1253 }
1254 if( ta && is_known_inst ) {
1255 if ( offset != Type::OffsetBot &&
1256 offset > arrayOopDesc::length_offset_in_bytes() ) {
1257 offset = Type::OffsetBot; // Flatten constant access into array body only
1258 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1259 }
1260 } else if( ta && _AliasLevel >= 2 ) {
1261 // For arrays indexed by constant indices, we flatten the alias
1262 // space to include all of the array body. Only the header, klass
1263 // and array length can be accessed un-aliased.
1264 if( offset != Type::OffsetBot ) {
1265 if( ta->const_oop() ) { // MethodData* or Method*
1266 offset = Type::OffsetBot; // Flatten constant access into array body
1267 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1268 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1269 // range is OK as-is.
1270 tj = ta = TypeAryPtr::RANGE;
1271 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1272 tj = TypeInstPtr::KLASS; // all klass loads look alike
1273 ta = TypeAryPtr::RANGE; // generic ignored junk
1274 ptr = TypePtr::BotPTR;
1275 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1276 tj = TypeInstPtr::MARK;
1277 ta = TypeAryPtr::RANGE; // generic ignored junk
1278 ptr = TypePtr::BotPTR;
1279 } else { // Random constant offset into array body
1280 offset = Type::OffsetBot; // Flatten constant access into array body
1281 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1282 }
1283 }
1284 // Arrays of fixed size alias with arrays of unknown size.
1285 if (ta->size() != TypeInt::POS) {
1286 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1287 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1288 }
1289 // Arrays of known objects become arrays of unknown objects.
1290 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1291 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1292 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1293 }
1294 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1295 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1296 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1297 }
1298 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1299 // cannot be distinguished by bytecode alone.
1300 if (ta->elem() == TypeInt::BOOL) {
1301 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1302 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1303 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1304 }
1305 // During the 2nd round of IterGVN, NotNull castings are removed.
1306 // Make sure the Bottom and NotNull variants alias the same.
1307 // Also, make sure exact and non-exact variants alias the same.
1308 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1309 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1310 }
1311 }
1312
1313 // Oop pointers need some flattening
1314 const TypeInstPtr *to = tj->isa_instptr();
1315 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1316 ciInstanceKlass *k = to->klass()->as_instance_klass();
1317 if( ptr == TypePtr::Constant ) {
1318 if (to->klass() != ciEnv::current()->Class_klass() ||
1319 offset < k->size_helper() * wordSize) {
1320 // No constant oop pointers (such as Strings); they alias with
1321 // unknown strings.
1322 assert(!is_known_inst, "not scalarizable allocation");
1323 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1324 }
1325 } else if( is_known_inst ) {
1326 tj = to; // Keep NotNull and klass_is_exact for instance type
1327 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1328 // During the 2nd round of IterGVN, NotNull castings are removed.
1329 // Make sure the Bottom and NotNull variants alias the same.
1330 // Also, make sure exact and non-exact variants alias the same.
1331 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1332 }
1333 if (to->speculative() != NULL) {
1334 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1335 }
1336 // Canonicalize the holder of this field
1337 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1338 // First handle header references such as a LoadKlassNode, even if the
1339 // object's klass is unloaded at compile time (4965979).
1340 if (!is_known_inst) { // Do it only for non-instance types
1341 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1342 }
1343 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1344 // Static fields are in the space above the normal instance
1345 // fields in the java.lang.Class instance.
1346 if (to->klass() != ciEnv::current()->Class_klass()) {
1347 to = NULL;
1348 tj = TypeOopPtr::BOTTOM;
1349 offset = tj->offset();
1350 }
1351 } else {
1352 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1353 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1354 if( is_known_inst ) {
1355 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1356 } else {
1357 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1358 }
1359 }
1360 }
1361 }
1362
1363 // Klass pointers to object array klasses need some flattening
1364 const TypeKlassPtr *tk = tj->isa_klassptr();
1365 if( tk ) {
1366 // If we are referencing a field within a Klass, we need
1367 // to assume the worst case of an Object. Both exact and
1368 // inexact types must flatten to the same alias class so
1369 // use NotNull as the PTR.
1370 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1371
1372 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1373 TypeKlassPtr::OBJECT->klass(),
1374 offset);
1375 }
1376
1377 ciKlass* klass = tk->klass();
1378 if( klass->is_obj_array_klass() ) {
1379 ciKlass* k = TypeAryPtr::OOPS->klass();
1380 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1381 k = TypeInstPtr::BOTTOM->klass();
1382 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1383 }
1384
1385 // Check for precise loads from the primary supertype array and force them
1386 // to the supertype cache alias index. Check for generic array loads from
1387 // the primary supertype array and also force them to the supertype cache
1388 // alias index. Since the same load can reach both, we need to merge
1389 // these 2 disparate memories into the same alias class. Since the
1390 // primary supertype array is read-only, there's no chance of confusion
1391 // where we bypass an array load and an array store.
1392 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1393 if (offset == Type::OffsetBot ||
1394 (offset >= primary_supers_offset &&
1395 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1396 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1397 offset = in_bytes(Klass::secondary_super_cache_offset());
1398 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1399 }
1400 }
1401
1402 // Flatten all Raw pointers together.
1403 if (tj->base() == Type::RawPtr)
1404 tj = TypeRawPtr::BOTTOM;
1405
1406 if (tj->base() == Type::AnyPtr)
1407 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1408
1409 // Flatten all to bottom for now
1410 switch( _AliasLevel ) {
1411 case 0:
1412 tj = TypePtr::BOTTOM;
1413 break;
1414 case 1: // Flatten to: oop, static, field or array
1415 switch (tj->base()) {
1416 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1417 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1418 case Type::AryPtr: // do not distinguish arrays at all
1419 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1420 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1421 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1422 default: ShouldNotReachHere();
1423 }
1424 break;
1425 case 2: // No collapsing at level 2; keep all splits
1426 case 3: // No collapsing at level 3; keep all splits
1427 break;
1428 default:
1429 Unimplemented();
1430 }
1431
1432 offset = tj->offset();
1433 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1434
1435 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1436 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1437 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1438 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1439 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1440 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1441 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1442 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1443 assert( tj->ptr() != TypePtr::TopPTR &&
1444 tj->ptr() != TypePtr::AnyNull &&
1445 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1446 // assert( tj->ptr() != TypePtr::Constant ||
1447 // tj->base() == Type::RawPtr ||
1448 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1449
1450 return tj;
1451 }
1452
1453 void Compile::AliasType::Init(int i, const TypePtr* at) {
1454 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1455 _index = i;
1456 _adr_type = at;
1457 _field = NULL;
1458 _element = NULL;
1459 _is_rewritable = true; // default
1460 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1461 if (atoop != NULL && atoop->is_known_instance()) {
1462 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1463 _general_index = Compile::current()->get_alias_index(gt);
1464 } else {
1465 _general_index = 0;
1466 }
1467 }
1468
1469 BasicType Compile::AliasType::basic_type() const {
1470 if (element() != NULL) {
1471 const Type* element = adr_type()->is_aryptr()->elem();
1472 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1473 } if (field() != NULL) {
1474 return field()->layout_type();
1475 } else {
1476 return T_ILLEGAL; // unknown
1477 }
1478 }
1479
1480 //---------------------------------print_on------------------------------------
1481 #ifndef PRODUCT
1482 void Compile::AliasType::print_on(outputStream* st) {
1483 if (index() < 10)
1484 st->print("@ <%d> ", index());
1485 else st->print("@ <%d>", index());
1486 st->print(is_rewritable() ? " " : " RO");
1487 int offset = adr_type()->offset();
1488 if (offset == Type::OffsetBot)
1489 st->print(" +any");
1490 else st->print(" +%-3d", offset);
1491 st->print(" in ");
1492 adr_type()->dump_on(st);
1493 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1494 if (field() != NULL && tjp) {
1495 if (tjp->klass() != field()->holder() ||
1496 tjp->offset() != field()->offset_in_bytes()) {
1497 st->print(" != ");
1498 field()->print();
1499 st->print(" ***");
1500 }
1501 }
1502 }
1503
1504 void print_alias_types() {
1505 Compile* C = Compile::current();
1506 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1507 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1508 C->alias_type(idx)->print_on(tty);
1509 tty->cr();
1510 }
1511 }
1512 #endif
1513
1514
1515 //----------------------------probe_alias_cache--------------------------------
1516 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1517 intptr_t key = (intptr_t) adr_type;
1518 key ^= key >> logAliasCacheSize;
1519 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1520 }
1521
1522
1523 //-----------------------------grow_alias_types--------------------------------
1524 void Compile::grow_alias_types() {
1525 const int old_ats = _max_alias_types; // how many before?
1526 const int new_ats = old_ats; // how many more?
1527 const int grow_ats = old_ats+new_ats; // how many now?
1528 _max_alias_types = grow_ats;
1529 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1530 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1531 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1532 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1533 }
1534
1535
1536 //--------------------------------find_alias_type------------------------------
1537 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1538 if (_AliasLevel == 0)
1539 return alias_type(AliasIdxBot);
1540
1541 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1542 if (ace->_adr_type == adr_type) {
1543 return alias_type(ace->_index);
1544 }
1545
1546 // Handle special cases.
1547 if (adr_type == NULL) return alias_type(AliasIdxTop);
1548 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1549
1550 // Do it the slow way.
1551 const TypePtr* flat = flatten_alias_type(adr_type);
1552
1553 #ifdef ASSERT
1554 {
1555 ResourceMark rm;
1556 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1557 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1558 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1559 Type::str(adr_type));
1560 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1561 const TypeOopPtr* foop = flat->is_oopptr();
1562 // Scalarizable allocations have exact klass always.
1563 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1564 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1565 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1566 Type::str(foop), Type::str(xoop));
1567 }
1568 }
1569 #endif
1570
1571 int idx = AliasIdxTop;
1572 for (int i = 0; i < num_alias_types(); i++) {
1573 if (alias_type(i)->adr_type() == flat) {
1574 idx = i;
1575 break;
1576 }
1577 }
1578
1579 if (idx == AliasIdxTop) {
1580 if (no_create) return NULL;
1581 // Grow the array if necessary.
1582 if (_num_alias_types == _max_alias_types) grow_alias_types();
1583 // Add a new alias type.
1584 idx = _num_alias_types++;
1585 _alias_types[idx]->Init(idx, flat);
1586 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1587 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1588 if (flat->isa_instptr()) {
1589 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1590 && flat->is_instptr()->klass() == env()->Class_klass())
1591 alias_type(idx)->set_rewritable(false);
1592 }
1593 if (flat->isa_aryptr()) {
1594 #ifdef ASSERT
1595 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1596 // (T_BYTE has the weakest alignment and size restrictions...)
1597 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1598 #endif
1599 if (flat->offset() == TypePtr::OffsetBot) {
1600 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1601 }
1602 }
1603 if (flat->isa_klassptr()) {
1604 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1605 alias_type(idx)->set_rewritable(false);
1606 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1607 alias_type(idx)->set_rewritable(false);
1608 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1609 alias_type(idx)->set_rewritable(false);
1610 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1611 alias_type(idx)->set_rewritable(false);
1612 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1613 alias_type(idx)->set_rewritable(false);
1614 }
1615 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1616 // but the base pointer type is not distinctive enough to identify
1617 // references into JavaThread.)
1618
1619 // Check for final fields.
1620 const TypeInstPtr* tinst = flat->isa_instptr();
1621 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1622 ciField* field;
1623 if (tinst->const_oop() != NULL &&
1624 tinst->klass() == ciEnv::current()->Class_klass() &&
1625 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1626 // static field
1627 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1628 field = k->get_field_by_offset(tinst->offset(), true);
1629 } else {
1630 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1631 field = k->get_field_by_offset(tinst->offset(), false);
1632 }
1633 assert(field == NULL ||
1634 original_field == NULL ||
1635 (field->holder() == original_field->holder() &&
1636 field->offset() == original_field->offset() &&
1637 field->is_static() == original_field->is_static()), "wrong field?");
1638 // Set field() and is_rewritable() attributes.
1639 if (field != NULL) alias_type(idx)->set_field(field);
1640 }
1641 }
1642
1643 // Fill the cache for next time.
1644 ace->_adr_type = adr_type;
1645 ace->_index = idx;
1646 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1647
1648 // Might as well try to fill the cache for the flattened version, too.
1649 AliasCacheEntry* face = probe_alias_cache(flat);
1650 if (face->_adr_type == NULL) {
1651 face->_adr_type = flat;
1652 face->_index = idx;
1653 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1654 }
1655
1656 return alias_type(idx);
1657 }
1658
1659
1660 Compile::AliasType* Compile::alias_type(ciField* field) {
1661 const TypeOopPtr* t;
1662 if (field->is_static())
1663 t = TypeInstPtr::make(field->holder()->java_mirror());
1664 else
1665 t = TypeOopPtr::make_from_klass_raw(field->holder());
1666 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1667 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1668 return atp;
1669 }
1670
1671
1672 //------------------------------have_alias_type--------------------------------
1673 bool Compile::have_alias_type(const TypePtr* adr_type) {
1674 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1675 if (ace->_adr_type == adr_type) {
1676 return true;
1677 }
1678
1679 // Handle special cases.
1680 if (adr_type == NULL) return true;
1681 if (adr_type == TypePtr::BOTTOM) return true;
1682
1683 return find_alias_type(adr_type, true, NULL) != NULL;
1684 }
1685
1686 //-----------------------------must_alias--------------------------------------
1687 // True if all values of the given address type are in the given alias category.
1688 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1689 if (alias_idx == AliasIdxBot) return true; // the universal category
1690 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1691 if (alias_idx == AliasIdxTop) return false; // the empty category
1692 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1693
1694 // the only remaining possible overlap is identity
1695 int adr_idx = get_alias_index(adr_type);
1696 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1697 assert(adr_idx == alias_idx ||
1698 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1699 && adr_type != TypeOopPtr::BOTTOM),
1700 "should not be testing for overlap with an unsafe pointer");
1701 return adr_idx == alias_idx;
1702 }
1703
1704 //------------------------------can_alias--------------------------------------
1705 // True if any values of the given address type are in the given alias category.
1706 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1707 if (alias_idx == AliasIdxTop) return false; // the empty category
1708 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1709 // Known instance doesn't alias with bottom memory
1710 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1711 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1712
1713 // the only remaining possible overlap is identity
1714 int adr_idx = get_alias_index(adr_type);
1715 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1716 return adr_idx == alias_idx;
1717 }
1718
1719
1720
1721 //---------------------------pop_warm_call-------------------------------------
1722 WarmCallInfo* Compile::pop_warm_call() {
1723 WarmCallInfo* wci = _warm_calls;
1724 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1725 return wci;
1726 }
1727
1728 //----------------------------Inline_Warm--------------------------------------
1729 int Compile::Inline_Warm() {
1730 // If there is room, try to inline some more warm call sites.
1731 // %%% Do a graph index compaction pass when we think we're out of space?
1732 if (!InlineWarmCalls) return 0;
1733
1734 int calls_made_hot = 0;
1735 int room_to_grow = NodeCountInliningCutoff - unique();
1736 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1737 int amount_grown = 0;
1738 WarmCallInfo* call;
1739 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1740 int est_size = (int)call->size();
1741 if (est_size > (room_to_grow - amount_grown)) {
1742 // This one won't fit anyway. Get rid of it.
1743 call->make_cold();
1744 continue;
1745 }
1746 call->make_hot();
1747 calls_made_hot++;
1748 amount_grown += est_size;
1749 amount_to_grow -= est_size;
1750 }
1751
1752 if (calls_made_hot > 0) set_major_progress();
1753 return calls_made_hot;
1754 }
1755
1756
1757 //----------------------------Finish_Warm--------------------------------------
1758 void Compile::Finish_Warm() {
1759 if (!InlineWarmCalls) return;
1760 if (failing()) return;
1761 if (warm_calls() == NULL) return;
1762
1763 // Clean up loose ends, if we are out of space for inlining.
1764 WarmCallInfo* call;
1765 while ((call = pop_warm_call()) != NULL) {
1766 call->make_cold();
1767 }
1768 }
1769
1770 //---------------------cleanup_loop_predicates-----------------------
1771 // Remove the opaque nodes that protect the predicates so that all unused
1772 // checks and uncommon_traps will be eliminated from the ideal graph
1773 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1774 if (predicate_count()==0) return;
1775 for (int i = predicate_count(); i > 0; i--) {
1776 Node * n = predicate_opaque1_node(i-1);
1777 assert(n->Opcode() == Op_Opaque1, "must be");
1778 igvn.replace_node(n, n->in(1));
1779 }
1780 assert(predicate_count()==0, "should be clean!");
1781 }
1782
1783 void Compile::add_range_check_cast(Node* n) {
1784 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1785 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1786 _range_check_casts->append(n);
1787 }
1788
1789 // Remove all range check dependent CastIINodes.
1790 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1791 for (int i = range_check_cast_count(); i > 0; i--) {
1792 Node* cast = range_check_cast_node(i-1);
1793 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1794 igvn.replace_node(cast, cast->in(1));
1795 }
1796 assert(range_check_cast_count() == 0, "should be empty");
1797 }
1798
1799 void Compile::add_opaque4_node(Node* n) {
1800 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1801 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1802 _opaque4_nodes->append(n);
1803 }
1804
1805 // Remove all Opaque4 nodes.
1806 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1807 for (int i = opaque4_count(); i > 0; i--) {
1808 Node* opaq = opaque4_node(i-1);
1809 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
1810 igvn.replace_node(opaq, opaq->in(2));
1811 }
1812 assert(opaque4_count() == 0, "should be empty");
1813 }
1814
1815 // StringOpts and late inlining of string methods
1816 void Compile::inline_string_calls(bool parse_time) {
1817 {
1818 // remove useless nodes to make the usage analysis simpler
1819 ResourceMark rm;
1820 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1821 }
1822
1823 {
1824 ResourceMark rm;
1825 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1826 PhaseStringOpts pso(initial_gvn(), for_igvn());
1827 print_method(PHASE_AFTER_STRINGOPTS, 3);
1828 }
1829
1830 // now inline anything that we skipped the first time around
1831 if (!parse_time) {
1832 _late_inlines_pos = _late_inlines.length();
1833 }
1834
1835 while (_string_late_inlines.length() > 0) {
1836 CallGenerator* cg = _string_late_inlines.pop();
1837 cg->do_late_inline();
1838 if (failing()) return;
1839 }
1840 _string_late_inlines.trunc_to(0);
1841 }
1842
1843 // Late inlining of boxing methods
1844 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1845 if (_boxing_late_inlines.length() > 0) {
1846 assert(has_boxed_value(), "inconsistent");
1847
1848 PhaseGVN* gvn = initial_gvn();
1849 set_inlining_incrementally(true);
1850
1851 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1852 for_igvn()->clear();
1853 gvn->replace_with(&igvn);
1854
1855 _late_inlines_pos = _late_inlines.length();
1856
1857 while (_boxing_late_inlines.length() > 0) {
1858 CallGenerator* cg = _boxing_late_inlines.pop();
1859 cg->do_late_inline();
1860 if (failing()) return;
1861 }
1862 _boxing_late_inlines.trunc_to(0);
1863
1864 inline_incrementally_cleanup(igvn);
1865
1866 set_inlining_incrementally(false);
1867 }
1868 }
1869
1870 bool Compile::inline_incrementally_one() {
1871 assert(IncrementalInline, "incremental inlining should be on");
1872
1873 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1874 set_inlining_progress(false);
1875 set_do_cleanup(false);
1876 int i = 0;
1877 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1878 CallGenerator* cg = _late_inlines.at(i);
1879 _late_inlines_pos = i+1;
1880 cg->do_late_inline();
1881 if (failing()) return false;
1882 }
1883 int j = 0;
1884 for (; i < _late_inlines.length(); i++, j++) {
1885 _late_inlines.at_put(j, _late_inlines.at(i));
1886 }
1887 _late_inlines.trunc_to(j);
1888 assert(inlining_progress() || _late_inlines.length() == 0, "");
1889
1890 bool needs_cleanup = true;
1891
1892 set_inlining_progress(false);
1893 set_do_cleanup(false);
1894 return (_late_inlines.length() > 0) && !needs_cleanup;
1895 }
1896
1897 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
1898 {
1899 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
1900 ResourceMark rm;
1901 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1902 }
1903 {
1904 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
1905 igvn = PhaseIterGVN(initial_gvn());
1906 igvn.optimize();
1907 }
1908 }
1909
1910 // Perform incremental inlining until bound on number of live nodes is reached
1911 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1912 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
1913
1914 set_inlining_incrementally(true);
1915 uint low_live_nodes = 0;
1916
1917 while (_late_inlines.length() > 0) {
1918 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1919 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1920 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
1921 // PhaseIdealLoop is expensive so we only try it once we are
1922 // out of live nodes and we only try it again if the previous
1923 // helped got the number of nodes down significantly
1924 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
1925 if (failing()) return;
1926 low_live_nodes = live_nodes();
1927 _major_progress = true;
1928 }
1929
1930 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1931 break; // finish
1932 }
1933 }
1934
1935 for_igvn()->clear();
1936 initial_gvn()->replace_with(&igvn);
1937
1938 while (inline_incrementally_one()) {
1939 assert(!failing(), "inconsistent");
1940 }
1941
1942 if (failing()) return;
1943
1944 inline_incrementally_cleanup(igvn);
1945
1946 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
1947
1948
1949 if (failing()) return;
1950 }
1951 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1952
1953 if (_string_late_inlines.length() > 0) {
1954 assert(has_stringbuilder(), "inconsistent");
1955 for_igvn()->clear();
1956 initial_gvn()->replace_with(&igvn);
1957
1958 inline_string_calls(false);
1959
1960 if (failing()) return;
1961
1962 inline_incrementally_cleanup(igvn);
1963 }
1964
1965 set_inlining_incrementally(false);
1966 }
1967
1968
1969 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
1970 if(_loop_opts_cnt > 0) {
1971 debug_only( int cnt = 0; );
1972 while(major_progress() && (_loop_opts_cnt > 0)) {
1973 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
1974 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1975 PhaseIdealLoop::optimize(igvn, mode);
1976 _loop_opts_cnt--;
1977 if (failing()) return false;
1978 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
1979 }
1980 }
1981 return true;
1982 }
1983
1984 // Remove edges from "root" to each SafePoint at a backward branch.
1985 // They were inserted during parsing (see add_safepoint()) to make
1986 // infinite loops without calls or exceptions visible to root, i.e.,
1987 // useful.
1988 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
1989 Node *r = root();
1990 if (r != NULL) {
1991 for (uint i = r->req(); i < r->len(); ++i) {
1992 Node *n = r->in(i);
1993 if (n != NULL && n->is_SafePoint()) {
1994 r->rm_prec(i);
1995 if (n->outcnt() == 0) {
1996 igvn.remove_dead_node(n);
1997 }
1998 --i;
1999 }
2000 }
2001 // Parsing may have added top inputs to the root node (Path
2002 // leading to the Halt node proven dead). Make sure we get a
2003 // chance to clean them up.
2004 igvn._worklist.push(r);
2005 igvn.optimize();
2006 }
2007 }
2008
2009 //------------------------------Optimize---------------------------------------
2010 // Given a graph, optimize it.
2011 void Compile::Optimize() {
2012 TracePhase tp("optimizer", &timers[_t_optimizer]);
2013
2014 #ifndef PRODUCT
2015 if (_directive->BreakAtCompileOption) {
2016 BREAKPOINT;
2017 }
2018
2019 #endif
2020
2021 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2022 #ifdef ASSERT
2023 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2024 #endif
2025
2026 ResourceMark rm;
2027
2028 print_inlining_reinit();
2029
2030 NOT_PRODUCT( verify_graph_edges(); )
2031
2032 print_method(PHASE_AFTER_PARSING);
2033
2034 {
2035 // Iterative Global Value Numbering, including ideal transforms
2036 // Initialize IterGVN with types and values from parse-time GVN
2037 PhaseIterGVN igvn(initial_gvn());
2038 #ifdef ASSERT
2039 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2040 #endif
2041 {
2042 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2043 igvn.optimize();
2044 }
2045
2046 if (failing()) return;
2047
2048 print_method(PHASE_ITER_GVN1, 2);
2049
2050 inline_incrementally(igvn);
2051
2052 print_method(PHASE_INCREMENTAL_INLINE, 2);
2053
2054 if (failing()) return;
2055
2056 if (eliminate_boxing()) {
2057 // Inline valueOf() methods now.
2058 inline_boxing_calls(igvn);
2059
2060 if (AlwaysIncrementalInline) {
2061 inline_incrementally(igvn);
2062 }
2063 if (failing()) return;
2064
2065 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2066 }
2067
2068 // Now that all inlining is over, cut edge from root to loop
2069 // safepoints
2070 remove_root_to_sfpts_edges(igvn);
2071
2072 // Remove the speculative part of types and clean up the graph from
2073 // the extra CastPP nodes whose only purpose is to carry them. Do
2074 // that early so that optimizations are not disrupted by the extra
2075 // CastPP nodes.
2076 remove_speculative_types(igvn);
2077
2078 // No more new expensive nodes will be added to the list from here
2079 // so keep only the actual candidates for optimizations.
2080 cleanup_expensive_nodes(igvn);
2081
2082 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2083 if (EnableVectorSupport && has_vbox_nodes()) {
2084 TracePhase tp("", &timers[_t_vector]);
2085 PhaseVector pv(igvn);
2086 pv.optimize_vector_boxes();
2087
2088 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2089 }
2090 assert(!has_vbox_nodes(), "sanity");
2091
2092 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2093 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2094 initial_gvn()->replace_with(&igvn);
2095 for_igvn()->clear();
2096 Unique_Node_List new_worklist(C->comp_arena());
2097 {
2098 ResourceMark rm;
2099 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2100 }
2101 set_for_igvn(&new_worklist);
2102 igvn = PhaseIterGVN(initial_gvn());
2103 igvn.optimize();
2104 }
2105
2106 // FIXME for_igvn() is corrupted from here: new_worklist which is set_for_ignv() was allocated on stack.
2107
2108 // Perform escape analysis
2109 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2110 if (has_loops()) {
2111 // Cleanup graph (remove dead nodes).
2112 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2113 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2114 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2115 if (failing()) return;
2116 }
2117 ConnectionGraph::do_analysis(this, &igvn);
2118
2119 if (failing()) return;
2120
2121 // Optimize out fields loads from scalar replaceable allocations.
2122 igvn.optimize();
2123 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2124
2125 if (failing()) return;
2126
2127 if (congraph() != NULL && macro_count() > 0) {
2128 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2129 PhaseMacroExpand mexp(igvn);
2130 mexp.eliminate_macro_nodes();
2131 igvn.set_delay_transform(false);
2132
2133 igvn.optimize();
2134 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2135
2136 if (failing()) return;
2137 }
2138 }
2139
2140 // Loop transforms on the ideal graph. Range Check Elimination,
2141 // peeling, unrolling, etc.
2142
2143 // Set loop opts counter
2144 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2145 {
2146 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2147 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2148 _loop_opts_cnt--;
2149 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2150 if (failing()) return;
2151 }
2152 // Loop opts pass if partial peeling occurred in previous pass
2153 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2154 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2155 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2156 _loop_opts_cnt--;
2157 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2158 if (failing()) return;
2159 }
2160 // Loop opts pass for loop-unrolling before CCP
2161 if(major_progress() && (_loop_opts_cnt > 0)) {
2162 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2163 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2164 _loop_opts_cnt--;
2165 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2166 }
2167 if (!failing()) {
2168 // Verify that last round of loop opts produced a valid graph
2169 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2170 PhaseIdealLoop::verify(igvn);
2171 }
2172 }
2173 if (failing()) return;
2174
2175 // Conditional Constant Propagation;
2176 PhaseCCP ccp( &igvn );
2177 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2178 {
2179 TracePhase tp("ccp", &timers[_t_ccp]);
2180 ccp.do_transform();
2181 }
2182 print_method(PHASE_CPP1, 2);
2183
2184 assert( true, "Break here to ccp.dump_old2new_map()");
2185
2186 // Iterative Global Value Numbering, including ideal transforms
2187 {
2188 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2189 igvn = ccp;
2190 igvn.optimize();
2191 }
2192 print_method(PHASE_ITER_GVN2, 2);
2193
2194 if (failing()) return;
2195
2196 // Loop transforms on the ideal graph. Range Check Elimination,
2197 // peeling, unrolling, etc.
2198 if (!optimize_loops(igvn, LoopOptsDefault)) {
2199 return;
2200 }
2201
2202 if (failing()) return;
2203
2204 // Ensure that major progress is now clear
2205 C->clear_major_progress();
2206
2207 {
2208 // Verify that all previous optimizations produced a valid graph
2209 // at least to this point, even if no loop optimizations were done.
2210 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2211 PhaseIdealLoop::verify(igvn);
2212 }
2213
2214 if (range_check_cast_count() > 0) {
2215 // No more loop optimizations. Remove all range check dependent CastIINodes.
2216 C->remove_range_check_casts(igvn);
2217 igvn.optimize();
2218 }
2219
2220 #ifdef ASSERT
2221 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2222 #endif
2223
2224 {
2225 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2226 PhaseMacroExpand mex(igvn);
2227 if (mex.expand_macro_nodes()) {
2228 assert(failing(), "must bail out w/ explicit message");
2229 return;
2230 }
2231 print_method(PHASE_MACRO_EXPANSION, 2);
2232 }
2233
2234 {
2235 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2236 if (bs->expand_barriers(this, igvn)) {
2237 assert(failing(), "must bail out w/ explicit message");
2238 return;
2239 }
2240 print_method(PHASE_BARRIER_EXPANSION, 2);
2241 }
2242
2243 if (opaque4_count() > 0) {
2244 C->remove_opaque4_nodes(igvn);
2245 igvn.optimize();
2246 }
2247
2248 DEBUG_ONLY( _modified_nodes = NULL; )
2249 } // (End scope of igvn; run destructor if necessary for asserts.)
2250
2251 process_print_inlining();
2252 // A method with only infinite loops has no edges entering loops from root
2253 {
2254 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2255 if (final_graph_reshaping()) {
2256 assert(failing(), "must bail out w/ explicit message");
2257 return;
2258 }
2259 }
2260
2261 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2262 }
2263
2264 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
2265 ResourceMark rm;
2266 stringStream ss;
2267 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
2268 if (n != NULL) {
2269 #ifndef PRODUCT
2270 ss.print(": %s %d", n->Name(), n->_idx);
2271 #else
2272 ss.print(": %d %d", n->Opcode(), n->_idx);
2273 #endif // !PRODUCT
2274 } else {
2275 ss.print_raw(": NULL");
2276 }
2277 C->print_method(cpt, ss.as_string(), level);
2278 }
2279
2280 void Compile::inline_vector_reboxing_calls() {
2281 if (C->_vector_reboxing_late_inlines.length() > 0) {
2282 PhaseGVN* gvn = C->initial_gvn();
2283
2284 _late_inlines_pos = C->_late_inlines.length();
2285 while (_vector_reboxing_late_inlines.length() > 0) {
2286 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2287 cg->do_late_inline();
2288 if (failing()) return;
2289 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2290 }
2291 _vector_reboxing_late_inlines.trunc_to(0);
2292 }
2293 }
2294
2295 bool Compile::has_vbox_nodes() {
2296 if (C->_vector_reboxing_late_inlines.length() > 0) {
2297 return true;
2298 }
2299 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2300 Node * n = C->macro_node(macro_idx);
2301 assert(n->is_macro(), "only macro nodes expected here");
2302 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2303 return true;
2304 }
2305 }
2306 return false;
2307 }
2308
2309 //------------------------------Code_Gen---------------------------------------
2310 // Given a graph, generate code for it
2311 void Compile::Code_Gen() {
2312 if (failing()) {
2313 return;
2314 }
2315
2316 // Perform instruction selection. You might think we could reclaim Matcher
2317 // memory PDQ, but actually the Matcher is used in generating spill code.
2318 // Internals of the Matcher (including some VectorSets) must remain live
2319 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2320 // set a bit in reclaimed memory.
2321
2322 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2323 // nodes. Mapping is only valid at the root of each matched subtree.
2324 NOT_PRODUCT( verify_graph_edges(); )
2325
2326 Matcher matcher;
2327 _matcher = &matcher;
2328 {
2329 TracePhase tp("matcher", &timers[_t_matcher]);
2330 matcher.match();
2331 if (failing()) {
2332 return;
2333 }
2334 print_method(PHASE_AFTER_MATCHING, 3);
2335 }
2336 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2337 // nodes. Mapping is only valid at the root of each matched subtree.
2338 NOT_PRODUCT( verify_graph_edges(); )
2339
2340 // If you have too many nodes, or if matching has failed, bail out
2341 check_node_count(0, "out of nodes matching instructions");
2342 if (failing()) {
2343 return;
2344 }
2345
2346 print_method(PHASE_MATCHING, 2);
2347
2348 // Build a proper-looking CFG
2349 PhaseCFG cfg(node_arena(), root(), matcher);
2350 _cfg = &cfg;
2351 {
2352 TracePhase tp("scheduler", &timers[_t_scheduler]);
2353 bool success = cfg.do_global_code_motion();
2354 if (!success) {
2355 return;
2356 }
2357
2358 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2359 NOT_PRODUCT( verify_graph_edges(); )
2360 debug_only( cfg.verify(); )
2361 }
2362
2363 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2364 _regalloc = ®alloc;
2365 {
2366 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2367 // Perform register allocation. After Chaitin, use-def chains are
2368 // no longer accurate (at spill code) and so must be ignored.
2369 // Node->LRG->reg mappings are still accurate.
2370 _regalloc->Register_Allocate();
2371
2372 // Bail out if the allocator builds too many nodes
2373 if (failing()) {
2374 return;
2375 }
2376 }
2377
2378 // Prior to register allocation we kept empty basic blocks in case the
2379 // the allocator needed a place to spill. After register allocation we
2380 // are not adding any new instructions. If any basic block is empty, we
2381 // can now safely remove it.
2382 {
2383 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2384 cfg.remove_empty_blocks();
2385 if (do_freq_based_layout()) {
2386 PhaseBlockLayout layout(cfg);
2387 } else {
2388 cfg.set_loop_alignment();
2389 }
2390 cfg.fixup_flow();
2391 }
2392
2393 // Apply peephole optimizations
2394 if( OptoPeephole ) {
2395 TracePhase tp("peephole", &timers[_t_peephole]);
2396 PhasePeephole peep( _regalloc, cfg);
2397 peep.do_transform();
2398 }
2399
2400 // Do late expand if CPU requires this.
2401 if (Matcher::require_postalloc_expand) {
2402 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2403 cfg.postalloc_expand(_regalloc);
2404 }
2405
2406 // Convert Nodes to instruction bits in a buffer
2407 {
2408 TracePhase tp("output", &timers[_t_output]);
2409 PhaseOutput output;
2410 output.Output();
2411 if (failing()) return;
2412 output.install();
2413 }
2414
2415 print_method(PHASE_FINAL_CODE);
2416
2417 // He's dead, Jim.
2418 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2419 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2420 }
2421
2422 //------------------------------Final_Reshape_Counts---------------------------
2423 // This class defines counters to help identify when a method
2424 // may/must be executed using hardware with only 24-bit precision.
2425 struct Final_Reshape_Counts : public StackObj {
2426 int _call_count; // count non-inlined 'common' calls
2427 int _float_count; // count float ops requiring 24-bit precision
2428 int _double_count; // count double ops requiring more precision
2429 int _java_call_count; // count non-inlined 'java' calls
2430 int _inner_loop_count; // count loops which need alignment
2431 VectorSet _visited; // Visitation flags
2432 Node_List _tests; // Set of IfNodes & PCTableNodes
2433
2434 Final_Reshape_Counts() :
2435 _call_count(0), _float_count(0), _double_count(0),
2436 _java_call_count(0), _inner_loop_count(0),
2437 _visited( Thread::current()->resource_area() ) { }
2438
2439 void inc_call_count () { _call_count ++; }
2440 void inc_float_count () { _float_count ++; }
2441 void inc_double_count() { _double_count++; }
2442 void inc_java_call_count() { _java_call_count++; }
2443 void inc_inner_loop_count() { _inner_loop_count++; }
2444
2445 int get_call_count () const { return _call_count ; }
2446 int get_float_count () const { return _float_count ; }
2447 int get_double_count() const { return _double_count; }
2448 int get_java_call_count() const { return _java_call_count; }
2449 int get_inner_loop_count() const { return _inner_loop_count; }
2450 };
2451
2452 #ifdef ASSERT
2453 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2454 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2455 // Make sure the offset goes inside the instance layout.
2456 return k->contains_field_offset(tp->offset());
2457 // Note that OffsetBot and OffsetTop are very negative.
2458 }
2459 #endif
2460
2461 // Eliminate trivially redundant StoreCMs and accumulate their
2462 // precedence edges.
2463 void Compile::eliminate_redundant_card_marks(Node* n) {
2464 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2465 if (n->in(MemNode::Address)->outcnt() > 1) {
2466 // There are multiple users of the same address so it might be
2467 // possible to eliminate some of the StoreCMs
2468 Node* mem = n->in(MemNode::Memory);
2469 Node* adr = n->in(MemNode::Address);
2470 Node* val = n->in(MemNode::ValueIn);
2471 Node* prev = n;
2472 bool done = false;
2473 // Walk the chain of StoreCMs eliminating ones that match. As
2474 // long as it's a chain of single users then the optimization is
2475 // safe. Eliminating partially redundant StoreCMs would require
2476 // cloning copies down the other paths.
2477 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2478 if (adr == mem->in(MemNode::Address) &&
2479 val == mem->in(MemNode::ValueIn)) {
2480 // redundant StoreCM
2481 if (mem->req() > MemNode::OopStore) {
2482 // Hasn't been processed by this code yet.
2483 n->add_prec(mem->in(MemNode::OopStore));
2484 } else {
2485 // Already converted to precedence edge
2486 for (uint i = mem->req(); i < mem->len(); i++) {
2487 // Accumulate any precedence edges
2488 if (mem->in(i) != NULL) {
2489 n->add_prec(mem->in(i));
2490 }
2491 }
2492 // Everything above this point has been processed.
2493 done = true;
2494 }
2495 // Eliminate the previous StoreCM
2496 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2497 assert(mem->outcnt() == 0, "should be dead");
2498 mem->disconnect_inputs(NULL, this);
2499 } else {
2500 prev = mem;
2501 }
2502 mem = prev->in(MemNode::Memory);
2503 }
2504 }
2505 }
2506
2507 //------------------------------final_graph_reshaping_impl----------------------
2508 // Implement items 1-5 from final_graph_reshaping below.
2509 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2510
2511 if ( n->outcnt() == 0 ) return; // dead node
2512 uint nop = n->Opcode();
2513
2514 // Check for 2-input instruction with "last use" on right input.
2515 // Swap to left input. Implements item (2).
2516 if( n->req() == 3 && // two-input instruction
2517 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2518 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2519 n->in(2)->outcnt() == 1 &&// right use IS a last use
2520 !n->in(2)->is_Con() ) { // right use is not a constant
2521 // Check for commutative opcode
2522 switch( nop ) {
2523 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2524 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
2525 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
2526 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2527 case Op_AndL: case Op_XorL: case Op_OrL:
2528 case Op_AndI: case Op_XorI: case Op_OrI: {
2529 // Move "last use" input to left by swapping inputs
2530 n->swap_edges(1, 2);
2531 break;
2532 }
2533 default:
2534 break;
2535 }
2536 }
2537
2538 #ifdef ASSERT
2539 if( n->is_Mem() ) {
2540 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2541 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2542 // oop will be recorded in oop map if load crosses safepoint
2543 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2544 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2545 "raw memory operations should have control edge");
2546 }
2547 if (n->is_MemBar()) {
2548 MemBarNode* mb = n->as_MemBar();
2549 if (mb->trailing_store() || mb->trailing_load_store()) {
2550 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2551 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2552 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2553 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2554 } else if (mb->leading()) {
2555 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2556 }
2557 }
2558 #endif
2559 // Count FPU ops and common calls, implements item (3)
2560 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2561 if (!gc_handled) {
2562 final_graph_reshaping_main_switch(n, frc, nop);
2563 }
2564
2565 // Collect CFG split points
2566 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2567 frc._tests.push(n);
2568 }
2569 }
2570
2571 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2572 switch( nop ) {
2573 // Count all float operations that may use FPU
2574 case Op_AddF:
2575 case Op_SubF:
2576 case Op_MulF:
2577 case Op_DivF:
2578 case Op_NegF:
2579 case Op_ModF:
2580 case Op_ConvI2F:
2581 case Op_ConF:
2582 case Op_CmpF:
2583 case Op_CmpF3:
2584 // case Op_ConvL2F: // longs are split into 32-bit halves
2585 frc.inc_float_count();
2586 break;
2587
2588 case Op_ConvF2D:
2589 case Op_ConvD2F:
2590 frc.inc_float_count();
2591 frc.inc_double_count();
2592 break;
2593
2594 // Count all double operations that may use FPU
2595 case Op_AddD:
2596 case Op_SubD:
2597 case Op_MulD:
2598 case Op_DivD:
2599 case Op_NegD:
2600 case Op_ModD:
2601 case Op_ConvI2D:
2602 case Op_ConvD2I:
2603 // case Op_ConvL2D: // handled by leaf call
2604 // case Op_ConvD2L: // handled by leaf call
2605 case Op_ConD:
2606 case Op_CmpD:
2607 case Op_CmpD3:
2608 frc.inc_double_count();
2609 break;
2610 case Op_Opaque1: // Remove Opaque Nodes before matching
2611 case Op_Opaque2: // Remove Opaque Nodes before matching
2612 case Op_Opaque3:
2613 n->subsume_by(n->in(1), this);
2614 break;
2615 case Op_CallStaticJava:
2616 case Op_CallJava:
2617 case Op_CallDynamicJava:
2618 frc.inc_java_call_count(); // Count java call site;
2619 case Op_CallRuntime:
2620 case Op_CallLeaf:
2621 case Op_CallLeafNoFP: {
2622 assert (n->is_Call(), "");
2623 CallNode *call = n->as_Call();
2624 // Count call sites where the FP mode bit would have to be flipped.
2625 // Do not count uncommon runtime calls:
2626 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2627 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2628 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2629 frc.inc_call_count(); // Count the call site
2630 } else { // See if uncommon argument is shared
2631 Node *n = call->in(TypeFunc::Parms);
2632 int nop = n->Opcode();
2633 // Clone shared simple arguments to uncommon calls, item (1).
2634 if (n->outcnt() > 1 &&
2635 !n->is_Proj() &&
2636 nop != Op_CreateEx &&
2637 nop != Op_CheckCastPP &&
2638 nop != Op_DecodeN &&
2639 nop != Op_DecodeNKlass &&
2640 !n->is_Mem() &&
2641 !n->is_Phi()) {
2642 Node *x = n->clone();
2643 call->set_req(TypeFunc::Parms, x);
2644 }
2645 }
2646 break;
2647 }
2648
2649 case Op_StoreD:
2650 case Op_LoadD:
2651 case Op_LoadD_unaligned:
2652 frc.inc_double_count();
2653 goto handle_mem;
2654 case Op_StoreF:
2655 case Op_LoadF:
2656 frc.inc_float_count();
2657 goto handle_mem;
2658
2659 case Op_StoreCM:
2660 {
2661 // Convert OopStore dependence into precedence edge
2662 Node* prec = n->in(MemNode::OopStore);
2663 n->del_req(MemNode::OopStore);
2664 n->add_prec(prec);
2665 eliminate_redundant_card_marks(n);
2666 }
2667
2668 // fall through
2669
2670 case Op_StoreB:
2671 case Op_StoreC:
2672 case Op_StorePConditional:
2673 case Op_StoreI:
2674 case Op_StoreL:
2675 case Op_StoreIConditional:
2676 case Op_StoreLConditional:
2677 case Op_CompareAndSwapB:
2678 case Op_CompareAndSwapS:
2679 case Op_CompareAndSwapI:
2680 case Op_CompareAndSwapL:
2681 case Op_CompareAndSwapP:
2682 case Op_CompareAndSwapN:
2683 case Op_WeakCompareAndSwapB:
2684 case Op_WeakCompareAndSwapS:
2685 case Op_WeakCompareAndSwapI:
2686 case Op_WeakCompareAndSwapL:
2687 case Op_WeakCompareAndSwapP:
2688 case Op_WeakCompareAndSwapN:
2689 case Op_CompareAndExchangeB:
2690 case Op_CompareAndExchangeS:
2691 case Op_CompareAndExchangeI:
2692 case Op_CompareAndExchangeL:
2693 case Op_CompareAndExchangeP:
2694 case Op_CompareAndExchangeN:
2695 case Op_GetAndAddS:
2696 case Op_GetAndAddB:
2697 case Op_GetAndAddI:
2698 case Op_GetAndAddL:
2699 case Op_GetAndSetS:
2700 case Op_GetAndSetB:
2701 case Op_GetAndSetI:
2702 case Op_GetAndSetL:
2703 case Op_GetAndSetP:
2704 case Op_GetAndSetN:
2705 case Op_StoreP:
2706 case Op_StoreN:
2707 case Op_StoreNKlass:
2708 case Op_LoadB:
2709 case Op_LoadUB:
2710 case Op_LoadUS:
2711 case Op_LoadI:
2712 case Op_LoadKlass:
2713 case Op_LoadNKlass:
2714 case Op_LoadL:
2715 case Op_LoadL_unaligned:
2716 case Op_LoadPLocked:
2717 case Op_LoadP:
2718 case Op_LoadN:
2719 case Op_LoadRange:
2720 case Op_LoadS: {
2721 handle_mem:
2722 #ifdef ASSERT
2723 if( VerifyOptoOopOffsets ) {
2724 MemNode* mem = n->as_Mem();
2725 // Check to see if address types have grounded out somehow.
2726 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2727 assert( !tp || oop_offset_is_sane(tp), "" );
2728 }
2729 #endif
2730 break;
2731 }
2732
2733 case Op_AddP: { // Assert sane base pointers
2734 Node *addp = n->in(AddPNode::Address);
2735 assert( !addp->is_AddP() ||
2736 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2737 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2738 "Base pointers must match (addp %u)", addp->_idx );
2739 #ifdef _LP64
2740 if ((UseCompressedOops || UseCompressedClassPointers) &&
2741 addp->Opcode() == Op_ConP &&
2742 addp == n->in(AddPNode::Base) &&
2743 n->in(AddPNode::Offset)->is_Con()) {
2744 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2745 // on the platform and on the compressed oops mode.
2746 // Use addressing with narrow klass to load with offset on x86.
2747 // Some platforms can use the constant pool to load ConP.
2748 // Do this transformation here since IGVN will convert ConN back to ConP.
2749 const Type* t = addp->bottom_type();
2750 bool is_oop = t->isa_oopptr() != NULL;
2751 bool is_klass = t->isa_klassptr() != NULL;
2752
2753 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2754 (is_klass && Matcher::const_klass_prefer_decode())) {
2755 Node* nn = NULL;
2756
2757 int op = is_oop ? Op_ConN : Op_ConNKlass;
2758
2759 // Look for existing ConN node of the same exact type.
2760 Node* r = root();
2761 uint cnt = r->outcnt();
2762 for (uint i = 0; i < cnt; i++) {
2763 Node* m = r->raw_out(i);
2764 if (m!= NULL && m->Opcode() == op &&
2765 m->bottom_type()->make_ptr() == t) {
2766 nn = m;
2767 break;
2768 }
2769 }
2770 if (nn != NULL) {
2771 // Decode a narrow oop to match address
2772 // [R12 + narrow_oop_reg<<3 + offset]
2773 if (is_oop) {
2774 nn = new DecodeNNode(nn, t);
2775 } else {
2776 nn = new DecodeNKlassNode(nn, t);
2777 }
2778 // Check for succeeding AddP which uses the same Base.
2779 // Otherwise we will run into the assertion above when visiting that guy.
2780 for (uint i = 0; i < n->outcnt(); ++i) {
2781 Node *out_i = n->raw_out(i);
2782 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
2783 out_i->set_req(AddPNode::Base, nn);
2784 #ifdef ASSERT
2785 for (uint j = 0; j < out_i->outcnt(); ++j) {
2786 Node *out_j = out_i->raw_out(j);
2787 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
2788 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
2789 }
2790 #endif
2791 }
2792 }
2793 n->set_req(AddPNode::Base, nn);
2794 n->set_req(AddPNode::Address, nn);
2795 if (addp->outcnt() == 0) {
2796 addp->disconnect_inputs(NULL, this);
2797 }
2798 }
2799 }
2800 }
2801 #endif
2802 // platform dependent reshaping of the address expression
2803 reshape_address(n->as_AddP());
2804 break;
2805 }
2806
2807 case Op_CastPP: {
2808 // Remove CastPP nodes to gain more freedom during scheduling but
2809 // keep the dependency they encode as control or precedence edges
2810 // (if control is set already) on memory operations. Some CastPP
2811 // nodes don't have a control (don't carry a dependency): skip
2812 // those.
2813 if (n->in(0) != NULL) {
2814 ResourceMark rm;
2815 Unique_Node_List wq;
2816 wq.push(n);
2817 for (uint next = 0; next < wq.size(); ++next) {
2818 Node *m = wq.at(next);
2819 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2820 Node* use = m->fast_out(i);
2821 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2822 use->ensure_control_or_add_prec(n->in(0));
2823 } else {
2824 switch(use->Opcode()) {
2825 case Op_AddP:
2826 case Op_DecodeN:
2827 case Op_DecodeNKlass:
2828 case Op_CheckCastPP:
2829 case Op_CastPP:
2830 wq.push(use);
2831 break;
2832 }
2833 }
2834 }
2835 }
2836 }
2837 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2838 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2839 Node* in1 = n->in(1);
2840 const Type* t = n->bottom_type();
2841 Node* new_in1 = in1->clone();
2842 new_in1->as_DecodeN()->set_type(t);
2843
2844 if (!Matcher::narrow_oop_use_complex_address()) {
2845 //
2846 // x86, ARM and friends can handle 2 adds in addressing mode
2847 // and Matcher can fold a DecodeN node into address by using
2848 // a narrow oop directly and do implicit NULL check in address:
2849 //
2850 // [R12 + narrow_oop_reg<<3 + offset]
2851 // NullCheck narrow_oop_reg
2852 //
2853 // On other platforms (Sparc) we have to keep new DecodeN node and
2854 // use it to do implicit NULL check in address:
2855 //
2856 // decode_not_null narrow_oop_reg, base_reg
2857 // [base_reg + offset]
2858 // NullCheck base_reg
2859 //
2860 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2861 // to keep the information to which NULL check the new DecodeN node
2862 // corresponds to use it as value in implicit_null_check().
2863 //
2864 new_in1->set_req(0, n->in(0));
2865 }
2866
2867 n->subsume_by(new_in1, this);
2868 if (in1->outcnt() == 0) {
2869 in1->disconnect_inputs(NULL, this);
2870 }
2871 } else {
2872 n->subsume_by(n->in(1), this);
2873 if (n->outcnt() == 0) {
2874 n->disconnect_inputs(NULL, this);
2875 }
2876 }
2877 break;
2878 }
2879 #ifdef _LP64
2880 case Op_CmpP:
2881 // Do this transformation here to preserve CmpPNode::sub() and
2882 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2883 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2884 Node* in1 = n->in(1);
2885 Node* in2 = n->in(2);
2886 if (!in1->is_DecodeNarrowPtr()) {
2887 in2 = in1;
2888 in1 = n->in(2);
2889 }
2890 assert(in1->is_DecodeNarrowPtr(), "sanity");
2891
2892 Node* new_in2 = NULL;
2893 if (in2->is_DecodeNarrowPtr()) {
2894 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2895 new_in2 = in2->in(1);
2896 } else if (in2->Opcode() == Op_ConP) {
2897 const Type* t = in2->bottom_type();
2898 if (t == TypePtr::NULL_PTR) {
2899 assert(in1->is_DecodeN(), "compare klass to null?");
2900 // Don't convert CmpP null check into CmpN if compressed
2901 // oops implicit null check is not generated.
2902 // This will allow to generate normal oop implicit null check.
2903 if (Matcher::gen_narrow_oop_implicit_null_checks())
2904 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
2905 //
2906 // This transformation together with CastPP transformation above
2907 // will generated code for implicit NULL checks for compressed oops.
2908 //
2909 // The original code after Optimize()
2910 //
2911 // LoadN memory, narrow_oop_reg
2912 // decode narrow_oop_reg, base_reg
2913 // CmpP base_reg, NULL
2914 // CastPP base_reg // NotNull
2915 // Load [base_reg + offset], val_reg
2916 //
2917 // after these transformations will be
2918 //
2919 // LoadN memory, narrow_oop_reg
2920 // CmpN narrow_oop_reg, NULL
2921 // decode_not_null narrow_oop_reg, base_reg
2922 // Load [base_reg + offset], val_reg
2923 //
2924 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2925 // since narrow oops can be used in debug info now (see the code in
2926 // final_graph_reshaping_walk()).
2927 //
2928 // At the end the code will be matched to
2929 // on x86:
2930 //
2931 // Load_narrow_oop memory, narrow_oop_reg
2932 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2933 // NullCheck narrow_oop_reg
2934 //
2935 // and on sparc:
2936 //
2937 // Load_narrow_oop memory, narrow_oop_reg
2938 // decode_not_null narrow_oop_reg, base_reg
2939 // Load [base_reg + offset], val_reg
2940 // NullCheck base_reg
2941 //
2942 } else if (t->isa_oopptr()) {
2943 new_in2 = ConNode::make(t->make_narrowoop());
2944 } else if (t->isa_klassptr()) {
2945 new_in2 = ConNode::make(t->make_narrowklass());
2946 }
2947 }
2948 if (new_in2 != NULL) {
2949 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
2950 n->subsume_by(cmpN, this);
2951 if (in1->outcnt() == 0) {
2952 in1->disconnect_inputs(NULL, this);
2953 }
2954 if (in2->outcnt() == 0) {
2955 in2->disconnect_inputs(NULL, this);
2956 }
2957 }
2958 }
2959 break;
2960
2961 case Op_DecodeN:
2962 case Op_DecodeNKlass:
2963 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2964 // DecodeN could be pinned when it can't be fold into
2965 // an address expression, see the code for Op_CastPP above.
2966 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2967 break;
2968
2969 case Op_EncodeP:
2970 case Op_EncodePKlass: {
2971 Node* in1 = n->in(1);
2972 if (in1->is_DecodeNarrowPtr()) {
2973 n->subsume_by(in1->in(1), this);
2974 } else if (in1->Opcode() == Op_ConP) {
2975 const Type* t = in1->bottom_type();
2976 if (t == TypePtr::NULL_PTR) {
2977 assert(t->isa_oopptr(), "null klass?");
2978 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
2979 } else if (t->isa_oopptr()) {
2980 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
2981 } else if (t->isa_klassptr()) {
2982 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
2983 }
2984 }
2985 if (in1->outcnt() == 0) {
2986 in1->disconnect_inputs(NULL, this);
2987 }
2988 break;
2989 }
2990
2991 case Op_Proj: {
2992 if (OptimizeStringConcat) {
2993 ProjNode* p = n->as_Proj();
2994 if (p->_is_io_use) {
2995 // Separate projections were used for the exception path which
2996 // are normally removed by a late inline. If it wasn't inlined
2997 // then they will hang around and should just be replaced with
2998 // the original one.
2999 Node* proj = NULL;
3000 // Replace with just one
3001 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3002 Node *use = i.get();
3003 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3004 proj = use;
3005 break;
3006 }
3007 }
3008 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3009 if (proj != NULL) {
3010 p->subsume_by(proj, this);
3011 }
3012 }
3013 }
3014 break;
3015 }
3016
3017 case Op_Phi:
3018 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3019 // The EncodeP optimization may create Phi with the same edges
3020 // for all paths. It is not handled well by Register Allocator.
3021 Node* unique_in = n->in(1);
3022 assert(unique_in != NULL, "");
3023 uint cnt = n->req();
3024 for (uint i = 2; i < cnt; i++) {
3025 Node* m = n->in(i);
3026 assert(m != NULL, "");
3027 if (unique_in != m)
3028 unique_in = NULL;
3029 }
3030 if (unique_in != NULL) {
3031 n->subsume_by(unique_in, this);
3032 }
3033 }
3034 break;
3035
3036 #endif
3037
3038 #ifdef ASSERT
3039 case Op_CastII:
3040 // Verify that all range check dependent CastII nodes were removed.
3041 if (n->isa_CastII()->has_range_check()) {
3042 n->dump(3);
3043 assert(false, "Range check dependent CastII node was not removed");
3044 }
3045 break;
3046 #endif
3047
3048 case Op_ModI:
3049 if (UseDivMod) {
3050 // Check if a%b and a/b both exist
3051 Node* d = n->find_similar(Op_DivI);
3052 if (d) {
3053 // Replace them with a fused divmod if supported
3054 if (Matcher::has_match_rule(Op_DivModI)) {
3055 DivModINode* divmod = DivModINode::make(n);
3056 d->subsume_by(divmod->div_proj(), this);
3057 n->subsume_by(divmod->mod_proj(), this);
3058 } else {
3059 // replace a%b with a-((a/b)*b)
3060 Node* mult = new MulINode(d, d->in(2));
3061 Node* sub = new SubINode(d->in(1), mult);
3062 n->subsume_by(sub, this);
3063 }
3064 }
3065 }
3066 break;
3067
3068 case Op_ModL:
3069 if (UseDivMod) {
3070 // Check if a%b and a/b both exist
3071 Node* d = n->find_similar(Op_DivL);
3072 if (d) {
3073 // Replace them with a fused divmod if supported
3074 if (Matcher::has_match_rule(Op_DivModL)) {
3075 DivModLNode* divmod = DivModLNode::make(n);
3076 d->subsume_by(divmod->div_proj(), this);
3077 n->subsume_by(divmod->mod_proj(), this);
3078 } else {
3079 // replace a%b with a-((a/b)*b)
3080 Node* mult = new MulLNode(d, d->in(2));
3081 Node* sub = new SubLNode(d->in(1), mult);
3082 n->subsume_by(sub, this);
3083 }
3084 }
3085 }
3086 break;
3087
3088 case Op_LoadVector:
3089 case Op_StoreVector:
3090 case Op_LoadVectorGather:
3091 case Op_StoreVectorScatter:
3092 break;
3093
3094 case Op_AddReductionVI:
3095 case Op_AddReductionVL:
3096 case Op_AddReductionVF:
3097 case Op_AddReductionVD:
3098 case Op_MulReductionVI:
3099 case Op_MulReductionVL:
3100 case Op_MulReductionVF:
3101 case Op_MulReductionVD:
3102 case Op_MinReductionV:
3103 case Op_MaxReductionV:
3104 case Op_AndReductionV:
3105 case Op_OrReductionV:
3106 case Op_XorReductionV:
3107 break;
3108
3109 case Op_PackB:
3110 case Op_PackS:
3111 case Op_PackI:
3112 case Op_PackF:
3113 case Op_PackL:
3114 case Op_PackD:
3115 if (n->req()-1 > 2) {
3116 // Replace many operand PackNodes with a binary tree for matching
3117 PackNode* p = (PackNode*) n;
3118 Node* btp = p->binary_tree_pack(1, n->req());
3119 n->subsume_by(btp, this);
3120 }
3121 break;
3122 case Op_Loop:
3123 case Op_CountedLoop:
3124 case Op_OuterStripMinedLoop:
3125 if (n->as_Loop()->is_inner_loop()) {
3126 frc.inc_inner_loop_count();
3127 }
3128 n->as_Loop()->verify_strip_mined(0);
3129 break;
3130 case Op_LShiftI:
3131 case Op_RShiftI:
3132 case Op_URShiftI:
3133 case Op_LShiftL:
3134 case Op_RShiftL:
3135 case Op_URShiftL:
3136 if (Matcher::need_masked_shift_count) {
3137 // The cpu's shift instructions don't restrict the count to the
3138 // lower 5/6 bits. We need to do the masking ourselves.
3139 Node* in2 = n->in(2);
3140 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3141 const TypeInt* t = in2->find_int_type();
3142 if (t != NULL && t->is_con()) {
3143 juint shift = t->get_con();
3144 if (shift > mask) { // Unsigned cmp
3145 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3146 }
3147 } else {
3148 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3149 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3150 n->set_req(2, shift);
3151 }
3152 }
3153 if (in2->outcnt() == 0) { // Remove dead node
3154 in2->disconnect_inputs(NULL, this);
3155 }
3156 }
3157 break;
3158 case Op_MemBarStoreStore:
3159 case Op_MemBarRelease:
3160 // Break the link with AllocateNode: it is no longer useful and
3161 // confuses register allocation.
3162 if (n->req() > MemBarNode::Precedent) {
3163 n->set_req(MemBarNode::Precedent, top());
3164 }
3165 break;
3166 case Op_MemBarAcquire: {
3167 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3168 // At parse time, the trailing MemBarAcquire for a volatile load
3169 // is created with an edge to the load. After optimizations,
3170 // that input may be a chain of Phis. If those phis have no
3171 // other use, then the MemBarAcquire keeps them alive and
3172 // register allocation can be confused.
3173 ResourceMark rm;
3174 Unique_Node_List wq;
3175 wq.push(n->in(MemBarNode::Precedent));
3176 n->set_req(MemBarNode::Precedent, top());
3177 while (wq.size() > 0) {
3178 Node* m = wq.pop();
3179 if (m->outcnt() == 0) {
3180 for (uint j = 0; j < m->req(); j++) {
3181 Node* in = m->in(j);
3182 if (in != NULL) {
3183 wq.push(in);
3184 }
3185 }
3186 m->disconnect_inputs(NULL, this);
3187 }
3188 }
3189 }
3190 break;
3191 }
3192 case Op_RangeCheck: {
3193 RangeCheckNode* rc = n->as_RangeCheck();
3194 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3195 n->subsume_by(iff, this);
3196 frc._tests.push(iff);
3197 break;
3198 }
3199 case Op_ConvI2L: {
3200 if (!Matcher::convi2l_type_required) {
3201 // Code generation on some platforms doesn't need accurate
3202 // ConvI2L types. Widening the type can help remove redundant
3203 // address computations.
3204 n->as_Type()->set_type(TypeLong::INT);
3205 ResourceMark rm;
3206 Unique_Node_List wq;
3207 wq.push(n);
3208 for (uint next = 0; next < wq.size(); next++) {
3209 Node *m = wq.at(next);
3210
3211 for(;;) {
3212 // Loop over all nodes with identical inputs edges as m
3213 Node* k = m->find_similar(m->Opcode());
3214 if (k == NULL) {
3215 break;
3216 }
3217 // Push their uses so we get a chance to remove node made
3218 // redundant
3219 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3220 Node* u = k->fast_out(i);
3221 if (u->Opcode() == Op_LShiftL ||
3222 u->Opcode() == Op_AddL ||
3223 u->Opcode() == Op_SubL ||
3224 u->Opcode() == Op_AddP) {
3225 wq.push(u);
3226 }
3227 }
3228 // Replace all nodes with identical edges as m with m
3229 k->subsume_by(m, this);
3230 }
3231 }
3232 }
3233 break;
3234 }
3235 case Op_CmpUL: {
3236 if (!Matcher::has_match_rule(Op_CmpUL)) {
3237 // No support for unsigned long comparisons
3238 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3239 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3240 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3241 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3242 Node* andl = new AndLNode(orl, remove_sign_mask);
3243 Node* cmp = new CmpLNode(andl, n->in(2));
3244 n->subsume_by(cmp, this);
3245 }
3246 break;
3247 }
3248 default:
3249 assert(!n->is_Call(), "");
3250 assert(!n->is_Mem(), "");
3251 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3252 break;
3253 }
3254 }
3255
3256 //------------------------------final_graph_reshaping_walk---------------------
3257 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3258 // requires that the walk visits a node's inputs before visiting the node.
3259 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3260 ResourceArea *area = Thread::current()->resource_area();
3261 Unique_Node_List sfpt(area);
3262
3263 frc._visited.set(root->_idx); // first, mark node as visited
3264 uint cnt = root->req();
3265 Node *n = root;
3266 uint i = 0;
3267 while (true) {
3268 if (i < cnt) {
3269 // Place all non-visited non-null inputs onto stack
3270 Node* m = n->in(i);
3271 ++i;
3272 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3273 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3274 // compute worst case interpreter size in case of a deoptimization
3275 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3276
3277 sfpt.push(m);
3278 }
3279 cnt = m->req();
3280 nstack.push(n, i); // put on stack parent and next input's index
3281 n = m;
3282 i = 0;
3283 }
3284 } else {
3285 // Now do post-visit work
3286 final_graph_reshaping_impl( n, frc );
3287 if (nstack.is_empty())
3288 break; // finished
3289 n = nstack.node(); // Get node from stack
3290 cnt = n->req();
3291 i = nstack.index();
3292 nstack.pop(); // Shift to the next node on stack
3293 }
3294 }
3295
3296 // Skip next transformation if compressed oops are not used.
3297 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3298 (!UseCompressedOops && !UseCompressedClassPointers))
3299 return;
3300
3301 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3302 // It could be done for an uncommon traps or any safepoints/calls
3303 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3304 while (sfpt.size() > 0) {
3305 n = sfpt.pop();
3306 JVMState *jvms = n->as_SafePoint()->jvms();
3307 assert(jvms != NULL, "sanity");
3308 int start = jvms->debug_start();
3309 int end = n->req();
3310 bool is_uncommon = (n->is_CallStaticJava() &&
3311 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3312 for (int j = start; j < end; j++) {
3313 Node* in = n->in(j);
3314 if (in->is_DecodeNarrowPtr()) {
3315 bool safe_to_skip = true;
3316 if (!is_uncommon ) {
3317 // Is it safe to skip?
3318 for (uint i = 0; i < in->outcnt(); i++) {
3319 Node* u = in->raw_out(i);
3320 if (!u->is_SafePoint() ||
3321 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3322 safe_to_skip = false;
3323 }
3324 }
3325 }
3326 if (safe_to_skip) {
3327 n->set_req(j, in->in(1));
3328 }
3329 if (in->outcnt() == 0) {
3330 in->disconnect_inputs(NULL, this);
3331 }
3332 }
3333 }
3334 }
3335 }
3336
3337 //------------------------------final_graph_reshaping--------------------------
3338 // Final Graph Reshaping.
3339 //
3340 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3341 // and not commoned up and forced early. Must come after regular
3342 // optimizations to avoid GVN undoing the cloning. Clone constant
3343 // inputs to Loop Phis; these will be split by the allocator anyways.
3344 // Remove Opaque nodes.
3345 // (2) Move last-uses by commutative operations to the left input to encourage
3346 // Intel update-in-place two-address operations and better register usage
3347 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3348 // calls canonicalizing them back.
3349 // (3) Count the number of double-precision FP ops, single-precision FP ops
3350 // and call sites. On Intel, we can get correct rounding either by
3351 // forcing singles to memory (requires extra stores and loads after each
3352 // FP bytecode) or we can set a rounding mode bit (requires setting and
3353 // clearing the mode bit around call sites). The mode bit is only used
3354 // if the relative frequency of single FP ops to calls is low enough.
3355 // This is a key transform for SPEC mpeg_audio.
3356 // (4) Detect infinite loops; blobs of code reachable from above but not
3357 // below. Several of the Code_Gen algorithms fail on such code shapes,
3358 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3359 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3360 // Detection is by looking for IfNodes where only 1 projection is
3361 // reachable from below or CatchNodes missing some targets.
3362 // (5) Assert for insane oop offsets in debug mode.
3363
3364 bool Compile::final_graph_reshaping() {
3365 // an infinite loop may have been eliminated by the optimizer,
3366 // in which case the graph will be empty.
3367 if (root()->req() == 1) {
3368 record_method_not_compilable("trivial infinite loop");
3369 return true;
3370 }
3371
3372 // Expensive nodes have their control input set to prevent the GVN
3373 // from freely commoning them. There's no GVN beyond this point so
3374 // no need to keep the control input. We want the expensive nodes to
3375 // be freely moved to the least frequent code path by gcm.
3376 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3377 for (int i = 0; i < expensive_count(); i++) {
3378 _expensive_nodes->at(i)->set_req(0, NULL);
3379 }
3380
3381 Final_Reshape_Counts frc;
3382
3383 // Visit everybody reachable!
3384 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3385 Node_Stack nstack(live_nodes() >> 1);
3386 final_graph_reshaping_walk(nstack, root(), frc);
3387
3388 // Check for unreachable (from below) code (i.e., infinite loops).
3389 for( uint i = 0; i < frc._tests.size(); i++ ) {
3390 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3391 // Get number of CFG targets.
3392 // Note that PCTables include exception targets after calls.
3393 uint required_outcnt = n->required_outcnt();
3394 if (n->outcnt() != required_outcnt) {
3395 // Check for a few special cases. Rethrow Nodes never take the
3396 // 'fall-thru' path, so expected kids is 1 less.
3397 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3398 if (n->in(0)->in(0)->is_Call()) {
3399 CallNode *call = n->in(0)->in(0)->as_Call();
3400 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3401 required_outcnt--; // Rethrow always has 1 less kid
3402 } else if (call->req() > TypeFunc::Parms &&
3403 call->is_CallDynamicJava()) {
3404 // Check for null receiver. In such case, the optimizer has
3405 // detected that the virtual call will always result in a null
3406 // pointer exception. The fall-through projection of this CatchNode
3407 // will not be populated.
3408 Node *arg0 = call->in(TypeFunc::Parms);
3409 if (arg0->is_Type() &&
3410 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3411 required_outcnt--;
3412 }
3413 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3414 call->req() > TypeFunc::Parms+1 &&
3415 call->is_CallStaticJava()) {
3416 // Check for negative array length. In such case, the optimizer has
3417 // detected that the allocation attempt will always result in an
3418 // exception. There is no fall-through projection of this CatchNode .
3419 Node *arg1 = call->in(TypeFunc::Parms+1);
3420 if (arg1->is_Type() &&
3421 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3422 required_outcnt--;
3423 }
3424 }
3425 }
3426 }
3427 // Recheck with a better notion of 'required_outcnt'
3428 if (n->outcnt() != required_outcnt) {
3429 record_method_not_compilable("malformed control flow");
3430 return true; // Not all targets reachable!
3431 }
3432 }
3433 // Check that I actually visited all kids. Unreached kids
3434 // must be infinite loops.
3435 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3436 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3437 record_method_not_compilable("infinite loop");
3438 return true; // Found unvisited kid; must be unreach
3439 }
3440
3441 // Here so verification code in final_graph_reshaping_walk()
3442 // always see an OuterStripMinedLoopEnd
3443 if (n->is_OuterStripMinedLoopEnd()) {
3444 IfNode* init_iff = n->as_If();
3445 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3446 n->subsume_by(iff, this);
3447 }
3448 }
3449
3450 #ifdef IA32
3451 // If original bytecodes contained a mixture of floats and doubles
3452 // check if the optimizer has made it homogenous, item (3).
3453 if (UseSSE == 0 &&
3454 frc.get_float_count() > 32 &&
3455 frc.get_double_count() == 0 &&
3456 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3457 set_24_bit_selection_and_mode(false, true);
3458 }
3459 #endif // IA32
3460
3461 set_java_calls(frc.get_java_call_count());
3462 set_inner_loops(frc.get_inner_loop_count());
3463
3464 // No infinite loops, no reason to bail out.
3465 return false;
3466 }
3467
3468 //-----------------------------too_many_traps----------------------------------
3469 // Report if there are too many traps at the current method and bci.
3470 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3471 bool Compile::too_many_traps(ciMethod* method,
3472 int bci,
3473 Deoptimization::DeoptReason reason) {
3474 ciMethodData* md = method->method_data();
3475 if (md->is_empty()) {
3476 // Assume the trap has not occurred, or that it occurred only
3477 // because of a transient condition during start-up in the interpreter.
3478 return false;
3479 }
3480 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3481 if (md->has_trap_at(bci, m, reason) != 0) {
3482 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3483 // Also, if there are multiple reasons, or if there is no per-BCI record,
3484 // assume the worst.
3485 if (log())
3486 log()->elem("observe trap='%s' count='%d'",
3487 Deoptimization::trap_reason_name(reason),
3488 md->trap_count(reason));
3489 return true;
3490 } else {
3491 // Ignore method/bci and see if there have been too many globally.
3492 return too_many_traps(reason, md);
3493 }
3494 }
3495
3496 // Less-accurate variant which does not require a method and bci.
3497 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3498 ciMethodData* logmd) {
3499 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3500 // Too many traps globally.
3501 // Note that we use cumulative trap_count, not just md->trap_count.
3502 if (log()) {
3503 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3504 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3505 Deoptimization::trap_reason_name(reason),
3506 mcount, trap_count(reason));
3507 }
3508 return true;
3509 } else {
3510 // The coast is clear.
3511 return false;
3512 }
3513 }
3514
3515 //--------------------------too_many_recompiles--------------------------------
3516 // Report if there are too many recompiles at the current method and bci.
3517 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3518 // Is not eager to return true, since this will cause the compiler to use
3519 // Action_none for a trap point, to avoid too many recompilations.
3520 bool Compile::too_many_recompiles(ciMethod* method,
3521 int bci,
3522 Deoptimization::DeoptReason reason) {
3523 ciMethodData* md = method->method_data();
3524 if (md->is_empty()) {
3525 // Assume the trap has not occurred, or that it occurred only
3526 // because of a transient condition during start-up in the interpreter.
3527 return false;
3528 }
3529 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3530 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3531 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3532 Deoptimization::DeoptReason per_bc_reason
3533 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3534 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3535 if ((per_bc_reason == Deoptimization::Reason_none
3536 || md->has_trap_at(bci, m, reason) != 0)
3537 // The trap frequency measure we care about is the recompile count:
3538 && md->trap_recompiled_at(bci, m)
3539 && md->overflow_recompile_count() >= bc_cutoff) {
3540 // Do not emit a trap here if it has already caused recompilations.
3541 // Also, if there are multiple reasons, or if there is no per-BCI record,
3542 // assume the worst.
3543 if (log())
3544 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3545 Deoptimization::trap_reason_name(reason),
3546 md->trap_count(reason),
3547 md->overflow_recompile_count());
3548 return true;
3549 } else if (trap_count(reason) != 0
3550 && decompile_count() >= m_cutoff) {
3551 // Too many recompiles globally, and we have seen this sort of trap.
3552 // Use cumulative decompile_count, not just md->decompile_count.
3553 if (log())
3554 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3555 Deoptimization::trap_reason_name(reason),
3556 md->trap_count(reason), trap_count(reason),
3557 md->decompile_count(), decompile_count());
3558 return true;
3559 } else {
3560 // The coast is clear.
3561 return false;
3562 }
3563 }
3564
3565 // Compute when not to trap. Used by matching trap based nodes and
3566 // NullCheck optimization.
3567 void Compile::set_allowed_deopt_reasons() {
3568 _allowed_reasons = 0;
3569 if (is_method_compilation()) {
3570 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3571 assert(rs < BitsPerInt, "recode bit map");
3572 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3573 _allowed_reasons |= nth_bit(rs);
3574 }
3575 }
3576 }
3577 }
3578
3579 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3580 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3581 }
3582
3583 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3584 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3585 }
3586
3587 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3588 if (holder->is_initialized()) {
3589 return false;
3590 }
3591 if (holder->is_being_initialized()) {
3592 if (accessing_method->holder() == holder) {
3593 // Access inside a class. The barrier can be elided when access happens in <clinit>,
3594 // <init>, or a static method. In all those cases, there was an initialization
3595 // barrier on the holder klass passed.
3596 if (accessing_method->is_static_initializer() ||
3597 accessing_method->is_object_initializer() ||
3598 accessing_method->is_static()) {
3599 return false;
3600 }
3601 } else if (accessing_method->holder()->is_subclass_of(holder)) {
3602 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3603 // In case of <init> or a static method, the barrier is on the subclass is not enough:
3604 // child class can become fully initialized while its parent class is still being initialized.
3605 if (accessing_method->is_static_initializer()) {
3606 return false;
3607 }
3608 }
3609 ciMethod* root = method(); // the root method of compilation
3610 if (root != accessing_method) {
3611 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3612 }
3613 }
3614 return true;
3615 }
3616
3617 #ifndef PRODUCT
3618 //------------------------------verify_graph_edges---------------------------
3619 // Walk the Graph and verify that there is a one-to-one correspondence
3620 // between Use-Def edges and Def-Use edges in the graph.
3621 void Compile::verify_graph_edges(bool no_dead_code) {
3622 if (VerifyGraphEdges) {
3623 ResourceArea *area = Thread::current()->resource_area();
3624 Unique_Node_List visited(area);
3625 // Call recursive graph walk to check edges
3626 _root->verify_edges(visited);
3627 if (no_dead_code) {
3628 // Now make sure that no visited node is used by an unvisited node.
3629 bool dead_nodes = false;
3630 Unique_Node_List checked(area);
3631 while (visited.size() > 0) {
3632 Node* n = visited.pop();
3633 checked.push(n);
3634 for (uint i = 0; i < n->outcnt(); i++) {
3635 Node* use = n->raw_out(i);
3636 if (checked.member(use)) continue; // already checked
3637 if (visited.member(use)) continue; // already in the graph
3638 if (use->is_Con()) continue; // a dead ConNode is OK
3639 // At this point, we have found a dead node which is DU-reachable.
3640 if (!dead_nodes) {
3641 tty->print_cr("*** Dead nodes reachable via DU edges:");
3642 dead_nodes = true;
3643 }
3644 use->dump(2);
3645 tty->print_cr("---");
3646 checked.push(use); // No repeats; pretend it is now checked.
3647 }
3648 }
3649 assert(!dead_nodes, "using nodes must be reachable from root");
3650 }
3651 }
3652 }
3653 #endif
3654
3655 // The Compile object keeps track of failure reasons separately from the ciEnv.
3656 // This is required because there is not quite a 1-1 relation between the
3657 // ciEnv and its compilation task and the Compile object. Note that one
3658 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3659 // to backtrack and retry without subsuming loads. Other than this backtracking
3660 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3661 // by the logic in C2Compiler.
3662 void Compile::record_failure(const char* reason) {
3663 if (log() != NULL) {
3664 log()->elem("failure reason='%s' phase='compile'", reason);
3665 }
3666 if (_failure_reason == NULL) {
3667 // Record the first failure reason.
3668 _failure_reason = reason;
3669 }
3670
3671 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3672 C->print_method(PHASE_FAILURE);
3673 }
3674 _root = NULL; // flush the graph, too
3675 }
3676
3677 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3678 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3679 _phase_name(name), _dolog(CITimeVerbose)
3680 {
3681 if (_dolog) {
3682 C = Compile::current();
3683 _log = C->log();
3684 } else {
3685 C = NULL;
3686 _log = NULL;
3687 }
3688 if (_log != NULL) {
3689 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3690 _log->stamp();
3691 _log->end_head();
3692 }
3693 }
3694
3695 Compile::TracePhase::~TracePhase() {
3696
3697 C = Compile::current();
3698 if (_dolog) {
3699 _log = C->log();
3700 } else {
3701 _log = NULL;
3702 }
3703
3704 #ifdef ASSERT
3705 if (PrintIdealNodeCount) {
3706 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3707 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3708 }
3709
3710 if (VerifyIdealNodeCount) {
3711 Compile::current()->print_missing_nodes();
3712 }
3713 #endif
3714
3715 if (_log != NULL) {
3716 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3717 }
3718 }
3719
3720 //----------------------------static_subtype_check-----------------------------
3721 // Shortcut important common cases when superklass is exact:
3722 // (0) superklass is java.lang.Object (can occur in reflective code)
3723 // (1) subklass is already limited to a subtype of superklass => always ok
3724 // (2) subklass does not overlap with superklass => always fail
3725 // (3) superklass has NO subtypes and we can check with a simple compare.
3726 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3727 if (StressReflectiveCode) {
3728 return SSC_full_test; // Let caller generate the general case.
3729 }
3730
3731 if (superk == env()->Object_klass()) {
3732 return SSC_always_true; // (0) this test cannot fail
3733 }
3734
3735 ciType* superelem = superk;
3736 if (superelem->is_array_klass())
3737 superelem = superelem->as_array_klass()->base_element_type();
3738
3739 if (!subk->is_interface()) { // cannot trust static interface types yet
3740 if (subk->is_subtype_of(superk)) {
3741 return SSC_always_true; // (1) false path dead; no dynamic test needed
3742 }
3743 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
3744 !superk->is_subtype_of(subk)) {
3745 return SSC_always_false;
3746 }
3747 }
3748
3749 // If casting to an instance klass, it must have no subtypes
3750 if (superk->is_interface()) {
3751 // Cannot trust interfaces yet.
3752 // %%% S.B. superk->nof_implementors() == 1
3753 } else if (superelem->is_instance_klass()) {
3754 ciInstanceKlass* ik = superelem->as_instance_klass();
3755 if (!ik->has_subklass() && !ik->is_interface()) {
3756 if (!ik->is_final()) {
3757 // Add a dependency if there is a chance of a later subclass.
3758 dependencies()->assert_leaf_type(ik);
3759 }
3760 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3761 }
3762 } else {
3763 // A primitive array type has no subtypes.
3764 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3765 }
3766
3767 return SSC_full_test;
3768 }
3769
3770 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
3771 #ifdef _LP64
3772 // The scaled index operand to AddP must be a clean 64-bit value.
3773 // Java allows a 32-bit int to be incremented to a negative
3774 // value, which appears in a 64-bit register as a large
3775 // positive number. Using that large positive number as an
3776 // operand in pointer arithmetic has bad consequences.
3777 // On the other hand, 32-bit overflow is rare, and the possibility
3778 // can often be excluded, if we annotate the ConvI2L node with
3779 // a type assertion that its value is known to be a small positive
3780 // number. (The prior range check has ensured this.)
3781 // This assertion is used by ConvI2LNode::Ideal.
3782 int index_max = max_jint - 1; // array size is max_jint, index is one less
3783 if (sizetype != NULL) index_max = sizetype->_hi - 1;
3784 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
3785 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
3786 #endif
3787 return idx;
3788 }
3789
3790 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
3791 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
3792 if (ctrl != NULL) {
3793 // Express control dependency by a CastII node with a narrow type.
3794 value = new CastIINode(value, itype, false, true /* range check dependency */);
3795 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
3796 // node from floating above the range check during loop optimizations. Otherwise, the
3797 // ConvI2L node may be eliminated independently of the range check, causing the data path
3798 // to become TOP while the control path is still there (although it's unreachable).
3799 value->set_req(0, ctrl);
3800 // Save CastII node to remove it after loop optimizations.
3801 phase->C->add_range_check_cast(value);
3802 value = phase->transform(value);
3803 }
3804 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
3805 return phase->transform(new ConvI2LNode(value, ltype));
3806 }
3807
3808 void Compile::print_inlining_stream_free() {
3809 if (_print_inlining_stream != NULL) {
3810 _print_inlining_stream->~stringStream();
3811 _print_inlining_stream = NULL;
3812 }
3813 }
3814
3815 // The message about the current inlining is accumulated in
3816 // _print_inlining_stream and transfered into the _print_inlining_list
3817 // once we know whether inlining succeeds or not. For regular
3818 // inlining, messages are appended to the buffer pointed by
3819 // _print_inlining_idx in the _print_inlining_list. For late inlining,
3820 // a new buffer is added after _print_inlining_idx in the list. This
3821 // way we can update the inlining message for late inlining call site
3822 // when the inlining is attempted again.
3823 void Compile::print_inlining_init() {
3824 if (print_inlining() || print_intrinsics()) {
3825 // print_inlining_init is actually called several times.
3826 print_inlining_stream_free();
3827 _print_inlining_stream = new stringStream();
3828 // Watch out: The memory initialized by the constructor call PrintInliningBuffer()
3829 // will be copied into the only initial element. The default destructor of
3830 // PrintInliningBuffer will be called when leaving the scope here. If it
3831 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss
3832 // would be destructed, too!
3833 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
3834 }
3835 }
3836
3837 void Compile::print_inlining_reinit() {
3838 if (print_inlining() || print_intrinsics()) {
3839 print_inlining_stream_free();
3840 // Re allocate buffer when we change ResourceMark
3841 _print_inlining_stream = new stringStream();
3842 }
3843 }
3844
3845 void Compile::print_inlining_reset() {
3846 _print_inlining_stream->reset();
3847 }
3848
3849 void Compile::print_inlining_commit() {
3850 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
3851 // Transfer the message from _print_inlining_stream to the current
3852 // _print_inlining_list buffer and clear _print_inlining_stream.
3853 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
3854 print_inlining_reset();
3855 }
3856
3857 void Compile::print_inlining_push() {
3858 // Add new buffer to the _print_inlining_list at current position
3859 _print_inlining_idx++;
3860 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
3861 }
3862
3863 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
3864 return _print_inlining_list->at(_print_inlining_idx);
3865 }
3866
3867 void Compile::print_inlining_update(CallGenerator* cg) {
3868 if (print_inlining() || print_intrinsics()) {
3869 if (!cg->is_late_inline()) {
3870 if (print_inlining_current().cg() != NULL) {
3871 print_inlining_push();
3872 }
3873 print_inlining_commit();
3874 } else {
3875 if (print_inlining_current().cg() != cg &&
3876 (print_inlining_current().cg() != NULL ||
3877 print_inlining_current().ss()->size() != 0)) {
3878 print_inlining_push();
3879 }
3880 print_inlining_commit();
3881 print_inlining_current().set_cg(cg);
3882 }
3883 }
3884 }
3885
3886 void Compile::print_inlining_move_to(CallGenerator* cg) {
3887 // We resume inlining at a late inlining call site. Locate the
3888 // corresponding inlining buffer so that we can update it.
3889 if (print_inlining()) {
3890 for (int i = 0; i < _print_inlining_list->length(); i++) {
3891 if (_print_inlining_list->adr_at(i)->cg() == cg) {
3892 _print_inlining_idx = i;
3893 return;
3894 }
3895 }
3896 ShouldNotReachHere();
3897 }
3898 }
3899
3900 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
3901 if (print_inlining()) {
3902 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
3903 assert(print_inlining_current().cg() == cg, "wrong entry");
3904 // replace message with new message
3905 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
3906 print_inlining_commit();
3907 print_inlining_current().set_cg(cg);
3908 }
3909 }
3910
3911 void Compile::print_inlining_assert_ready() {
3912 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
3913 }
3914
3915 void Compile::process_print_inlining() {
3916 bool do_print_inlining = print_inlining() || print_intrinsics();
3917 if (do_print_inlining || log() != NULL) {
3918 // Print inlining message for candidates that we couldn't inline
3919 // for lack of space
3920 for (int i = 0; i < _late_inlines.length(); i++) {
3921 CallGenerator* cg = _late_inlines.at(i);
3922 if (!cg->is_mh_late_inline()) {
3923 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
3924 if (do_print_inlining) {
3925 cg->print_inlining_late(msg);
3926 }
3927 log_late_inline_failure(cg, msg);
3928 }
3929 }
3930 }
3931 if (do_print_inlining) {
3932 ResourceMark rm;
3933 stringStream ss;
3934 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
3935 for (int i = 0; i < _print_inlining_list->length(); i++) {
3936 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3937 _print_inlining_list->at(i).freeStream();
3938 }
3939 // Reset _print_inlining_list, it only contains destructed objects.
3940 // It is on the arena, so it will be freed when the arena is reset.
3941 _print_inlining_list = NULL;
3942 // _print_inlining_stream won't be used anymore, either.
3943 print_inlining_stream_free();
3944 size_t end = ss.size();
3945 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
3946 strncpy(_print_inlining_output, ss.base(), end+1);
3947 _print_inlining_output[end] = 0;
3948 }
3949 }
3950
3951 void Compile::dump_print_inlining() {
3952 if (_print_inlining_output != NULL) {
3953 tty->print_raw(_print_inlining_output);
3954 }
3955 }
3956
3957 void Compile::log_late_inline(CallGenerator* cg) {
3958 if (log() != NULL) {
3959 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
3960 cg->unique_id());
3961 JVMState* p = cg->call_node()->jvms();
3962 while (p != NULL) {
3963 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
3964 p = p->caller();
3965 }
3966 log()->tail("late_inline");
3967 }
3968 }
3969
3970 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
3971 log_late_inline(cg);
3972 if (log() != NULL) {
3973 log()->inline_fail(msg);
3974 }
3975 }
3976
3977 void Compile::log_inline_id(CallGenerator* cg) {
3978 if (log() != NULL) {
3979 // The LogCompilation tool needs a unique way to identify late
3980 // inline call sites. This id must be unique for this call site in
3981 // this compilation. Try to have it unique across compilations as
3982 // well because it can be convenient when grepping through the log
3983 // file.
3984 // Distinguish OSR compilations from others in case CICountOSR is
3985 // on.
3986 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
3987 cg->set_unique_id(id);
3988 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
3989 }
3990 }
3991
3992 void Compile::log_inline_failure(const char* msg) {
3993 if (C->log() != NULL) {
3994 C->log()->inline_fail(msg);
3995 }
3996 }
3997
3998
3999 // Dump inlining replay data to the stream.
4000 // Don't change thread state and acquire any locks.
4001 void Compile::dump_inline_data(outputStream* out) {
4002 InlineTree* inl_tree = ilt();
4003 if (inl_tree != NULL) {
4004 out->print(" inline %d", inl_tree->count());
4005 inl_tree->dump_replay_data(out);
4006 }
4007 }
4008
4009 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4010 if (n1->Opcode() < n2->Opcode()) return -1;
4011 else if (n1->Opcode() > n2->Opcode()) return 1;
4012
4013 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4014 for (uint i = 1; i < n1->req(); i++) {
4015 if (n1->in(i) < n2->in(i)) return -1;
4016 else if (n1->in(i) > n2->in(i)) return 1;
4017 }
4018
4019 return 0;
4020 }
4021
4022 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4023 Node* n1 = *n1p;
4024 Node* n2 = *n2p;
4025
4026 return cmp_expensive_nodes(n1, n2);
4027 }
4028
4029 void Compile::sort_expensive_nodes() {
4030 if (!expensive_nodes_sorted()) {
4031 _expensive_nodes->sort(cmp_expensive_nodes);
4032 }
4033 }
4034
4035 bool Compile::expensive_nodes_sorted() const {
4036 for (int i = 1; i < _expensive_nodes->length(); i++) {
4037 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4038 return false;
4039 }
4040 }
4041 return true;
4042 }
4043
4044 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4045 if (_expensive_nodes->length() == 0) {
4046 return false;
4047 }
4048
4049 assert(OptimizeExpensiveOps, "optimization off?");
4050
4051 // Take this opportunity to remove dead nodes from the list
4052 int j = 0;
4053 for (int i = 0; i < _expensive_nodes->length(); i++) {
4054 Node* n = _expensive_nodes->at(i);
4055 if (!n->is_unreachable(igvn)) {
4056 assert(n->is_expensive(), "should be expensive");
4057 _expensive_nodes->at_put(j, n);
4058 j++;
4059 }
4060 }
4061 _expensive_nodes->trunc_to(j);
4062
4063 // Then sort the list so that similar nodes are next to each other
4064 // and check for at least two nodes of identical kind with same data
4065 // inputs.
4066 sort_expensive_nodes();
4067
4068 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4069 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4070 return true;
4071 }
4072 }
4073
4074 return false;
4075 }
4076
4077 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4078 if (_expensive_nodes->length() == 0) {
4079 return;
4080 }
4081
4082 assert(OptimizeExpensiveOps, "optimization off?");
4083
4084 // Sort to bring similar nodes next to each other and clear the
4085 // control input of nodes for which there's only a single copy.
4086 sort_expensive_nodes();
4087
4088 int j = 0;
4089 int identical = 0;
4090 int i = 0;
4091 bool modified = false;
4092 for (; i < _expensive_nodes->length()-1; i++) {
4093 assert(j <= i, "can't write beyond current index");
4094 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4095 identical++;
4096 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4097 continue;
4098 }
4099 if (identical > 0) {
4100 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4101 identical = 0;
4102 } else {
4103 Node* n = _expensive_nodes->at(i);
4104 igvn.replace_input_of(n, 0, NULL);
4105 igvn.hash_insert(n);
4106 modified = true;
4107 }
4108 }
4109 if (identical > 0) {
4110 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4111 } else if (_expensive_nodes->length() >= 1) {
4112 Node* n = _expensive_nodes->at(i);
4113 igvn.replace_input_of(n, 0, NULL);
4114 igvn.hash_insert(n);
4115 modified = true;
4116 }
4117 _expensive_nodes->trunc_to(j);
4118 if (modified) {
4119 igvn.optimize();
4120 }
4121 }
4122
4123 void Compile::add_expensive_node(Node * n) {
4124 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4125 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4126 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4127 if (OptimizeExpensiveOps) {
4128 _expensive_nodes->append(n);
4129 } else {
4130 // Clear control input and let IGVN optimize expensive nodes if
4131 // OptimizeExpensiveOps is off.
4132 n->set_req(0, NULL);
4133 }
4134 }
4135
4136 /**
4137 * Remove the speculative part of types and clean up the graph
4138 */
4139 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4140 if (UseTypeSpeculation) {
4141 Unique_Node_List worklist;
4142 worklist.push(root());
4143 int modified = 0;
4144 // Go over all type nodes that carry a speculative type, drop the
4145 // speculative part of the type and enqueue the node for an igvn
4146 // which may optimize it out.
4147 for (uint next = 0; next < worklist.size(); ++next) {
4148 Node *n = worklist.at(next);
4149 if (n->is_Type()) {
4150 TypeNode* tn = n->as_Type();
4151 const Type* t = tn->type();
4152 const Type* t_no_spec = t->remove_speculative();
4153 if (t_no_spec != t) {
4154 bool in_hash = igvn.hash_delete(n);
4155 assert(in_hash, "node should be in igvn hash table");
4156 tn->set_type(t_no_spec);
4157 igvn.hash_insert(n);
4158 igvn._worklist.push(n); // give it a chance to go away
4159 modified++;
4160 }
4161 }
4162 uint max = n->len();
4163 for( uint i = 0; i < max; ++i ) {
4164 Node *m = n->in(i);
4165 if (not_a_node(m)) continue;
4166 worklist.push(m);
4167 }
4168 }
4169 // Drop the speculative part of all types in the igvn's type table
4170 igvn.remove_speculative_types();
4171 if (modified > 0) {
4172 igvn.optimize();
4173 }
4174 #ifdef ASSERT
4175 // Verify that after the IGVN is over no speculative type has resurfaced
4176 worklist.clear();
4177 worklist.push(root());
4178 for (uint next = 0; next < worklist.size(); ++next) {
4179 Node *n = worklist.at(next);
4180 const Type* t = igvn.type_or_null(n);
4181 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4182 if (n->is_Type()) {
4183 t = n->as_Type()->type();
4184 assert(t == t->remove_speculative(), "no more speculative types");
4185 }
4186 uint max = n->len();
4187 for( uint i = 0; i < max; ++i ) {
4188 Node *m = n->in(i);
4189 if (not_a_node(m)) continue;
4190 worklist.push(m);
4191 }
4192 }
4193 igvn.check_no_speculative_types();
4194 #endif
4195 }
4196 }
4197
4198 // Auxiliary method to support randomized stressing/fuzzing.
4199 //
4200 // This method can be called the arbitrary number of times, with current count
4201 // as the argument. The logic allows selecting a single candidate from the
4202 // running list of candidates as follows:
4203 // int count = 0;
4204 // Cand* selected = null;
4205 // while(cand = cand->next()) {
4206 // if (randomized_select(++count)) {
4207 // selected = cand;
4208 // }
4209 // }
4210 //
4211 // Including count equalizes the chances any candidate is "selected".
4212 // This is useful when we don't have the complete list of candidates to choose
4213 // from uniformly. In this case, we need to adjust the randomicity of the
4214 // selection, or else we will end up biasing the selection towards the latter
4215 // candidates.
4216 //
4217 // Quick back-envelope calculation shows that for the list of n candidates
4218 // the equal probability for the candidate to persist as "best" can be
4219 // achieved by replacing it with "next" k-th candidate with the probability
4220 // of 1/k. It can be easily shown that by the end of the run, the
4221 // probability for any candidate is converged to 1/n, thus giving the
4222 // uniform distribution among all the candidates.
4223 //
4224 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4225 #define RANDOMIZED_DOMAIN_POW 29
4226 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4227 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4228 bool Compile::randomized_select(int count) {
4229 assert(count > 0, "only positive");
4230 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4231 }
4232
4233 CloneMap& Compile::clone_map() { return _clone_map; }
4234 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4235
4236 void NodeCloneInfo::dump() const {
4237 tty->print(" {%d:%d} ", idx(), gen());
4238 }
4239
4240 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4241 uint64_t val = value(old->_idx);
4242 NodeCloneInfo cio(val);
4243 assert(val != 0, "old node should be in the map");
4244 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4245 insert(nnn->_idx, cin.get());
4246 #ifndef PRODUCT
4247 if (is_debug()) {
4248 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4249 }
4250 #endif
4251 }
4252
4253 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4254 NodeCloneInfo cio(value(old->_idx));
4255 if (cio.get() == 0) {
4256 cio.set(old->_idx, 0);
4257 insert(old->_idx, cio.get());
4258 #ifndef PRODUCT
4259 if (is_debug()) {
4260 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4261 }
4262 #endif
4263 }
4264 clone(old, nnn, gen);
4265 }
4266
4267 int CloneMap::max_gen() const {
4268 int g = 0;
4269 DictI di(_dict);
4270 for(; di.test(); ++di) {
4271 int t = gen(di._key);
4272 if (g < t) {
4273 g = t;
4274 #ifndef PRODUCT
4275 if (is_debug()) {
4276 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4277 }
4278 #endif
4279 }
4280 }
4281 return g;
4282 }
4283
4284 void CloneMap::dump(node_idx_t key) const {
4285 uint64_t val = value(key);
4286 if (val != 0) {
4287 NodeCloneInfo ni(val);
4288 ni.dump();
4289 }
4290 }