1 /*
2 * Copyright (c) 1997, 2022, 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 #include "precompiled.hpp"
25 #include "jvm_io.h"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/javaClasses.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 "jfr/jfrEvents.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "opto/addnode.hpp"
41 #include "opto/block.hpp"
42 #include "opto/c2compiler.hpp"
43 #include "opto/callGenerator.hpp"
44 #include "opto/callnode.hpp"
45 #include "opto/castnode.hpp"
46 #include "opto/cfgnode.hpp"
47 #include "opto/chaitin.hpp"
48 #include "opto/compile.hpp"
49 #include "opto/connode.hpp"
50 #include "opto/convertnode.hpp"
51 #include "opto/divnode.hpp"
52 #include "opto/escape.hpp"
53 #include "opto/idealGraphPrinter.hpp"
54 #include "opto/loopnode.hpp"
55 #include "opto/machnode.hpp"
56 #include "opto/macro.hpp"
57 #include "opto/matcher.hpp"
58 #include "opto/mathexactnode.hpp"
59 #include "opto/memnode.hpp"
60 #include "opto/mulnode.hpp"
61 #include "opto/narrowptrnode.hpp"
62 #include "opto/node.hpp"
63 #include "opto/opcodes.hpp"
64 #include "opto/output.hpp"
65 #include "opto/parse.hpp"
66 #include "opto/phaseX.hpp"
67 #include "opto/rootnode.hpp"
68 #include "opto/runtime.hpp"
69 #include "opto/stringopts.hpp"
70 #include "opto/type.hpp"
71 #include "opto/vector.hpp"
72 #include "opto/vectornode.hpp"
73 #include "runtime/globals_extension.hpp"
74 #include "runtime/sharedRuntime.hpp"
75 #include "runtime/signature.hpp"
76 #include "runtime/stubRoutines.hpp"
77 #include "runtime/timer.hpp"
78 #include "utilities/align.hpp"
79 #include "utilities/copy.hpp"
80 #include "utilities/macros.hpp"
81 #include "utilities/resourceHash.hpp"
82
83
84 // -------------------- Compile::mach_constant_base_node -----------------------
85 // Constant table base node singleton.
86 MachConstantBaseNode* Compile::mach_constant_base_node() {
87 if (_mach_constant_base_node == NULL) {
88 _mach_constant_base_node = new MachConstantBaseNode();
89 _mach_constant_base_node->add_req(C->root());
90 }
91 return _mach_constant_base_node;
92 }
93
94
95 /// Support for intrinsics.
96
97 // Return the index at which m must be inserted (or already exists).
98 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
99 class IntrinsicDescPair {
100 private:
101 ciMethod* _m;
102 bool _is_virtual;
103 public:
104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
106 ciMethod* m= elt->method();
107 ciMethod* key_m = key->_m;
108 if (key_m < m) return -1;
109 else if (key_m > m) return 1;
110 else {
111 bool is_virtual = elt->is_virtual();
112 bool key_virtual = key->_is_virtual;
113 if (key_virtual < is_virtual) return -1;
114 else if (key_virtual > is_virtual) return 1;
115 else return 0;
116 }
117 }
118 };
119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
120 #ifdef ASSERT
121 for (int i = 1; i < _intrinsics.length(); i++) {
122 CallGenerator* cg1 = _intrinsics.at(i-1);
123 CallGenerator* cg2 = _intrinsics.at(i);
124 assert(cg1->method() != cg2->method()
125 ? cg1->method() < cg2->method()
126 : cg1->is_virtual() < cg2->is_virtual(),
127 "compiler intrinsics list must stay sorted");
128 }
129 #endif
130 IntrinsicDescPair pair(m, is_virtual);
131 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
132 }
133
134 void Compile::register_intrinsic(CallGenerator* cg) {
135 bool found = false;
136 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
137 assert(!found, "registering twice");
138 _intrinsics.insert_before(index, cg);
139 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
140 }
141
142 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
143 assert(m->is_loaded(), "don't try this on unloaded methods");
144 if (_intrinsics.length() > 0) {
145 bool found = false;
146 int index = intrinsic_insertion_index(m, is_virtual, found);
147 if (found) {
148 return _intrinsics.at(index);
149 }
150 }
151 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
152 if (m->intrinsic_id() != vmIntrinsics::_none &&
153 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
154 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
155 if (cg != NULL) {
156 // Save it for next time:
157 register_intrinsic(cg);
158 return cg;
159 } else {
160 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
161 }
162 }
163 return NULL;
164 }
165
166 // Compile::make_vm_intrinsic is defined in library_call.cpp.
167
168 #ifndef PRODUCT
169 // statistics gathering...
170
171 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
172 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
173
174 inline int as_int(vmIntrinsics::ID id) {
175 return vmIntrinsics::as_int(id);
176 }
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[as_int(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[as_int(id)] += 1);
188 if (count == 1) {
189 changed = true; // first time
190 }
191 // increment the overall count also:
192 _intrinsic_hist_count[as_int(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[as_int(id)] = (jubyte) (oflags | flags);
209 }
210 // update the overall flags also:
211 _intrinsic_hist_flags[as_int(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[as_int(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 (auto id : EnumRange<vmIntrinsicID>{}) {
237 int flags = _intrinsic_hist_flags[as_int(id)];
238 juint count = _intrinsic_hist_count[as_int(id)];
239 if ((flags | count) != 0) {
240 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
241 }
242 }
243 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
244 if (xtty != NULL) xtty->tail("statistics");
245 }
246
247 void Compile::print_statistics() {
248 { ttyLocker ttyl;
249 if (xtty != NULL) xtty->head("statistics type='opto'");
250 Parse::print_statistics();
251 PhaseStringOpts::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[as_int(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 // Identify all nodes that are reachable from below, useful.
291 // Use breadth-first pass that records state in a Unique_Node_List,
292 // recursive traversal is slower.
293 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
294 int estimated_worklist_size = live_nodes();
295 useful.map( estimated_worklist_size, NULL ); // preallocate space
296
297 // Initialize worklist
298 if (root() != NULL) { useful.push(root()); }
299 // If 'top' is cached, declare it useful to preserve cached node
300 if( cached_top_node() ) { useful.push(cached_top_node()); }
301
302 // Push all useful nodes onto the list, breadthfirst
303 for( uint next = 0; next < useful.size(); ++next ) {
304 assert( next < unique(), "Unique useful nodes < total nodes");
305 Node *n = useful.at(next);
306 uint max = n->len();
307 for( uint i = 0; i < max; ++i ) {
308 Node *m = n->in(i);
309 if (not_a_node(m)) continue;
310 useful.push(m);
311 }
312 }
313 }
314
315 // Update dead_node_list with any missing dead nodes using useful
316 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
317 void Compile::update_dead_node_list(Unique_Node_List &useful) {
318 uint max_idx = unique();
319 VectorSet& useful_node_set = useful.member_set();
320
321 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
322 // If node with index node_idx is not in useful set,
323 // mark it as dead in dead node list.
324 if (!useful_node_set.test(node_idx)) {
325 record_dead_node(node_idx);
326 }
327 }
328 }
329
330 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
331 int shift = 0;
332 for (int i = 0; i < inlines->length(); i++) {
333 CallGenerator* cg = inlines->at(i);
334 if (useful.member(cg->call_node())) {
335 if (shift > 0) {
336 inlines->at_put(i - shift, cg);
337 }
338 } else {
339 shift++; // skip over the dead element
340 }
341 }
342 if (shift > 0) {
343 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
344 }
345 }
346
347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
348 assert(dead != NULL && dead->is_Call(), "sanity");
349 int found = 0;
350 for (int i = 0; i < inlines->length(); i++) {
351 if (inlines->at(i)->call_node() == dead) {
352 inlines->remove_at(i);
353 found++;
354 NOT_DEBUG( break; ) // elements are unique, so exit early
355 }
356 }
357 assert(found <= 1, "not unique");
358 }
359
360 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
361 for (int i = node_list.length() - 1; i >= 0; i--) {
362 Node* n = node_list.at(i);
363 if (!useful.member(n)) {
364 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
365 }
366 }
367 }
368
369 void Compile::remove_useless_node(Node* dead) {
370 remove_modified_node(dead);
371
372 // Constant node that has no out-edges and has only one in-edge from
373 // root is usually dead. However, sometimes reshaping walk makes
374 // it reachable by adding use edges. So, we will NOT count Con nodes
375 // as dead to be conservative about the dead node count at any
376 // given time.
377 if (!dead->is_Con()) {
378 record_dead_node(dead->_idx);
379 }
380 if (dead->is_macro()) {
381 remove_macro_node(dead);
382 }
383 if (dead->is_expensive()) {
384 remove_expensive_node(dead);
385 }
386 if (dead->Opcode() == Op_Opaque4) {
387 remove_skeleton_predicate_opaq(dead);
388 }
389 if (dead->for_post_loop_opts_igvn()) {
390 remove_from_post_loop_opts_igvn(dead);
391 }
392 if (dead->is_Call()) {
393 remove_useless_late_inlines( &_late_inlines, dead);
394 remove_useless_late_inlines( &_string_late_inlines, dead);
395 remove_useless_late_inlines( &_boxing_late_inlines, dead);
396 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
397 }
398 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
399 bs->unregister_potential_barrier_node(dead);
400 }
401
402 // Disconnect all useless nodes by disconnecting those at the boundary.
403 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
404 uint next = 0;
405 while (next < useful.size()) {
406 Node *n = useful.at(next++);
407 if (n->is_SafePoint()) {
408 // We're done with a parsing phase. Replaced nodes are not valid
409 // beyond that point.
410 n->as_SafePoint()->delete_replaced_nodes();
411 }
412 // Use raw traversal of out edges since this code removes out edges
413 int max = n->outcnt();
414 for (int j = 0; j < max; ++j) {
415 Node* child = n->raw_out(j);
416 if (!useful.member(child)) {
417 assert(!child->is_top() || child != top(),
418 "If top is cached in Compile object it is in useful list");
419 // Only need to remove this out-edge to the useless node
420 n->raw_del_out(j);
421 --j;
422 --max;
423 }
424 }
425 if (n->outcnt() == 1 && n->has_special_unique_user()) {
426 record_for_igvn(n->unique_out());
427 }
428 }
429
430 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
431 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
432 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
433 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
434 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
435 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
436
437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
438 bs->eliminate_useless_gc_barriers(useful, this);
439 // clean up the late inline lists
440 remove_useless_late_inlines( &_late_inlines, useful);
441 remove_useless_late_inlines( &_string_late_inlines, useful);
442 remove_useless_late_inlines( &_boxing_late_inlines, useful);
443 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
444 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
445 }
446
447 // ============================================================================
448 //------------------------------CompileWrapper---------------------------------
449 class CompileWrapper : public StackObj {
450 Compile *const _compile;
451 public:
452 CompileWrapper(Compile* compile);
453
454 ~CompileWrapper();
455 };
456
457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
458 // the Compile* pointer is stored in the current ciEnv:
459 ciEnv* env = compile->env();
460 assert(env == ciEnv::current(), "must already be a ciEnv active");
461 assert(env->compiler_data() == NULL, "compile already active?");
462 env->set_compiler_data(compile);
463 assert(compile == Compile::current(), "sanity");
464
465 compile->set_type_dict(NULL);
466 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
467 compile->clone_map().set_clone_idx(0);
468 compile->set_type_last_size(0);
469 compile->set_last_tf(NULL, NULL);
470 compile->set_indexSet_arena(NULL);
471 compile->set_indexSet_free_block_list(NULL);
472 compile->init_type_arena();
473 Type::Initialize(compile);
474 _compile->begin_method();
475 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
476 }
477 CompileWrapper::~CompileWrapper() {
478 // simulate crash during compilation
479 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
480
481 _compile->end_method();
482 _compile->env()->set_compiler_data(NULL);
483 }
484
485
486 //----------------------------print_compile_messages---------------------------
487 void Compile::print_compile_messages() {
488 #ifndef PRODUCT
489 // Check if recompiling
490 if (!subsume_loads() && PrintOpto) {
491 // Recompiling without allowing machine instructions to subsume loads
492 tty->print_cr("*********************************************************");
493 tty->print_cr("** Bailout: Recompile without subsuming loads **");
494 tty->print_cr("*********************************************************");
495 }
496 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
497 // Recompiling without escape analysis
498 tty->print_cr("*********************************************************");
499 tty->print_cr("** Bailout: Recompile without escape analysis **");
500 tty->print_cr("*********************************************************");
501 }
502 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
503 // Recompiling without iterative escape analysis
504 tty->print_cr("*********************************************************");
505 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
506 tty->print_cr("*********************************************************");
507 }
508 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
509 // Recompiling without boxing elimination
510 tty->print_cr("*********************************************************");
511 tty->print_cr("** Bailout: Recompile without boxing elimination **");
512 tty->print_cr("*********************************************************");
513 }
514 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
515 // Recompiling without locks coarsening
516 tty->print_cr("*********************************************************");
517 tty->print_cr("** Bailout: Recompile without locks coarsening **");
518 tty->print_cr("*********************************************************");
519 }
520 if (env()->break_at_compile()) {
521 // Open the debugger when compiling this method.
522 tty->print("### Breaking when compiling: ");
523 method()->print_short_name();
524 tty->cr();
525 BREAKPOINT;
526 }
527
528 if( PrintOpto ) {
529 if (is_osr_compilation()) {
530 tty->print("[OSR]%3d", _compile_id);
531 } else {
532 tty->print("%3d", _compile_id);
533 }
534 }
535 #endif
536 }
537
538 #ifndef PRODUCT
539 void Compile::print_ideal_ir(const char* phase_name) {
540 ttyLocker ttyl;
541 // keep the following output all in one block
542 // This output goes directly to the tty, not the compiler log.
543 // To enable tools to match it up with the compilation activity,
544 // be sure to tag this tty output with the compile ID.
545 if (xtty != NULL) {
546 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
547 compile_id(),
548 is_osr_compilation() ? " compile_kind='osr'" : "",
549 phase_name);
550 }
551 if (_output == nullptr) {
552 root()->dump(9999);
553 } else {
554 // Dump the node blockwise if we have a scheduling
555 _output->print_scheduling();
556 }
557
558 if (xtty != NULL) {
559 xtty->tail("ideal");
560 }
561 }
562 #endif
563
564 // ============================================================================
565 //------------------------------Compile standard-------------------------------
566 debug_only( int Compile::_debug_idx = 100000; )
567
568 // Compile a method. entry_bci is -1 for normal compilations and indicates
569 // the continuation bci for on stack replacement.
570
571
572 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
573 Options options, DirectiveSet* directive)
574 : Phase(Compiler),
575 _compile_id(ci_env->compile_id()),
576 _options(options),
577 _method(target),
578 _entry_bci(osr_bci),
579 _ilt(NULL),
580 _stub_function(NULL),
581 _stub_name(NULL),
582 _stub_entry_point(NULL),
583 _max_node_limit(MaxNodeLimit),
584 _post_loop_opts_phase(false),
585 _inlining_progress(false),
586 _inlining_incrementally(false),
587 _do_cleanup(false),
588 _has_reserved_stack_access(target->has_reserved_stack_access()),
589 #ifndef PRODUCT
590 _igv_idx(0),
591 _trace_opto_output(directive->TraceOptoOutputOption),
592 #endif
593 _has_method_handle_invokes(false),
594 _clinit_barrier_on_entry(false),
595 _stress_seed(0),
596 _comp_arena(mtCompiler),
597 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
598 _env(ci_env),
599 _directive(directive),
600 _log(ci_env->log()),
601 _failure_reason(NULL),
602 _intrinsics (comp_arena(), 0, 0, NULL),
603 _macro_nodes (comp_arena(), 8, 0, NULL),
604 _predicate_opaqs (comp_arena(), 8, 0, NULL),
605 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
606 _expensive_nodes (comp_arena(), 8, 0, NULL),
607 _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
608 _coarsened_locks (comp_arena(), 8, 0, NULL),
609 _congraph(NULL),
610 NOT_PRODUCT(_igv_printer(NULL) COMMA)
611 _dead_node_list(comp_arena()),
612 _dead_node_count(0),
613 _node_arena(mtCompiler),
614 _old_arena(mtCompiler),
615 _mach_constant_base_node(NULL),
616 _Compile_types(mtCompiler),
617 _initial_gvn(NULL),
618 _for_igvn(NULL),
619 _late_inlines(comp_arena(), 2, 0, NULL),
620 _string_late_inlines(comp_arena(), 2, 0, NULL),
621 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
622 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
623 _late_inlines_pos(0),
624 _number_of_mh_late_inlines(0),
625 _native_invokers(comp_arena(), 1, 0, NULL),
626 _print_inlining_stream(NULL),
627 _print_inlining_list(NULL),
628 _print_inlining_idx(0),
629 _print_inlining_output(NULL),
630 _replay_inline_data(NULL),
631 _java_calls(0),
632 _inner_loops(0),
633 _interpreter_frame_size(0),
634 _output(NULL)
635 #ifndef PRODUCT
636 , _in_dump_cnt(0)
637 #endif
638 {
639 C = this;
640 CompileWrapper cw(this);
641
642 if (CITimeVerbose) {
643 tty->print(" ");
644 target->holder()->name()->print();
645 tty->print(".");
646 target->print_short_name();
647 tty->print(" ");
648 }
649 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
650 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
651
652 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
653 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
654 // We can always print a disassembly, either abstract (hex dump) or
655 // with the help of a suitable hsdis library. Thus, we should not
656 // couple print_assembly and print_opto_assembly controls.
657 // But: always print opto and regular assembly on compile command 'print'.
658 bool print_assembly = directive->PrintAssemblyOption;
659 set_print_assembly(print_opto_assembly || print_assembly);
660 #else
661 set_print_assembly(false); // must initialize.
662 #endif
663
664 #ifndef PRODUCT
665 set_parsed_irreducible_loop(false);
666
667 if (directive->ReplayInlineOption) {
668 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
669 }
670 #endif
671 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
672 set_print_intrinsics(directive->PrintIntrinsicsOption);
673 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
674
675 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
676 // Make sure the method being compiled gets its own MDO,
677 // so we can at least track the decompile_count().
678 // Need MDO to record RTM code generation state.
679 method()->ensure_method_data();
680 }
681
682 Init(::AliasLevel);
683
684
685 print_compile_messages();
686
687 _ilt = InlineTree::build_inline_tree_root();
688
689 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
690 assert(num_alias_types() >= AliasIdxRaw, "");
691
692 #define MINIMUM_NODE_HASH 1023
693 // Node list that Iterative GVN will start with
694 Unique_Node_List for_igvn(comp_arena());
695 set_for_igvn(&for_igvn);
696
697 // GVN that will be run immediately on new nodes
698 uint estimated_size = method()->code_size()*4+64;
699 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
700 PhaseGVN gvn(node_arena(), estimated_size);
701 set_initial_gvn(&gvn);
702
703 print_inlining_init();
704 { // Scope for timing the parser
705 TracePhase tp("parse", &timers[_t_parser]);
706
707 // Put top into the hash table ASAP.
708 initial_gvn()->transform_no_reclaim(top());
709
710 // Set up tf(), start(), and find a CallGenerator.
711 CallGenerator* cg = NULL;
712 if (is_osr_compilation()) {
713 const TypeTuple *domain = StartOSRNode::osr_domain();
714 const TypeTuple *range = TypeTuple::make_range(method()->signature());
715 init_tf(TypeFunc::make(domain, range));
716 StartNode* s = new StartOSRNode(root(), domain);
717 initial_gvn()->set_type_bottom(s);
718 init_start(s);
719 cg = CallGenerator::for_osr(method(), entry_bci());
720 } else {
721 // Normal case.
722 init_tf(TypeFunc::make(method()));
723 StartNode* s = new StartNode(root(), tf()->domain());
724 initial_gvn()->set_type_bottom(s);
725 init_start(s);
726 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
727 // With java.lang.ref.reference.get() we must go through the
728 // intrinsic - even when get() is the root
729 // method of the compile - so that, if necessary, the value in
730 // the referent field of the reference object gets recorded by
731 // the pre-barrier code.
732 cg = find_intrinsic(method(), false);
733 }
734 if (cg == NULL) {
735 float past_uses = method()->interpreter_invocation_count();
736 float expected_uses = past_uses;
737 cg = CallGenerator::for_inline(method(), expected_uses);
738 }
739 }
740 if (failing()) return;
741 if (cg == NULL) {
742 record_method_not_compilable("cannot parse method");
743 return;
744 }
745 JVMState* jvms = build_start_state(start(), tf());
746 if ((jvms = cg->generate(jvms)) == NULL) {
747 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
748 record_method_not_compilable("method parse failed");
749 }
750 return;
751 }
752 GraphKit kit(jvms);
753
754 if (!kit.stopped()) {
755 // Accept return values, and transfer control we know not where.
756 // This is done by a special, unique ReturnNode bound to root.
757 return_values(kit.jvms());
758 }
759
760 if (kit.has_exceptions()) {
761 // Any exceptions that escape from this call must be rethrown
762 // to whatever caller is dynamically above us on the stack.
763 // This is done by a special, unique RethrowNode bound to root.
764 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
765 }
766
767 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
768
769 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
770 inline_string_calls(true);
771 }
772
773 if (failing()) return;
774
775 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
776
777 // Remove clutter produced by parsing.
778 if (!failing()) {
779 ResourceMark rm;
780 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
781 }
782 }
783
784 // Note: Large methods are capped off in do_one_bytecode().
785 if (failing()) return;
786
787 // After parsing, node notes are no longer automagic.
788 // They must be propagated by register_new_node_with_optimizer(),
789 // clone(), or the like.
790 set_default_node_notes(NULL);
791
792 #ifndef PRODUCT
793 if (should_print_igv(1)) {
794 _igv_printer->print_inlining();
795 }
796 #endif
797
798 if (failing()) return;
799 NOT_PRODUCT( verify_graph_edges(); )
800
801 // If any phase is randomized for stress testing, seed random number
802 // generation and log the seed for repeatability.
803 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
804 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
805 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
806 FLAG_SET_ERGO(StressSeed, _stress_seed);
807 } else {
808 _stress_seed = StressSeed;
809 }
810 if (_log != NULL) {
811 _log->elem("stress_test seed='%u'", _stress_seed);
812 }
813 }
814
815 // Now optimize
816 Optimize();
817 if (failing()) return;
818 NOT_PRODUCT( verify_graph_edges(); )
819
820 #ifndef PRODUCT
821 if (should_print_ideal()) {
822 print_ideal_ir("print_ideal");
823 }
824 #endif
825
826 #ifdef ASSERT
827 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
828 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
829 #endif
830
831 // Dump compilation data to replay it.
832 if (directive->DumpReplayOption) {
833 env()->dump_replay_data(_compile_id);
834 }
835 if (directive->DumpInlineOption && (ilt() != NULL)) {
836 env()->dump_inline_data(_compile_id);
837 }
838
839 // Now that we know the size of all the monitors we can add a fixed slot
840 // for the original deopt pc.
841 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
842 set_fixed_slots(next_slot);
843
844 // Compute when to use implicit null checks. Used by matching trap based
845 // nodes and NullCheck optimization.
846 set_allowed_deopt_reasons();
847
848 // Now generate code
849 Code_Gen();
850 }
851
852 //------------------------------Compile----------------------------------------
853 // Compile a runtime stub
854 Compile::Compile( ciEnv* ci_env,
855 TypeFunc_generator generator,
856 address stub_function,
857 const char *stub_name,
858 int is_fancy_jump,
859 bool pass_tls,
860 bool return_pc,
861 DirectiveSet* directive)
862 : Phase(Compiler),
863 _compile_id(0),
864 _options(Options::for_runtime_stub()),
865 _method(NULL),
866 _entry_bci(InvocationEntryBci),
867 _stub_function(stub_function),
868 _stub_name(stub_name),
869 _stub_entry_point(NULL),
870 _max_node_limit(MaxNodeLimit),
871 _post_loop_opts_phase(false),
872 _inlining_progress(false),
873 _inlining_incrementally(false),
874 _has_reserved_stack_access(false),
875 #ifndef PRODUCT
876 _igv_idx(0),
877 _trace_opto_output(directive->TraceOptoOutputOption),
878 #endif
879 _has_method_handle_invokes(false),
880 _clinit_barrier_on_entry(false),
881 _stress_seed(0),
882 _comp_arena(mtCompiler),
883 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
884 _env(ci_env),
885 _directive(directive),
886 _log(ci_env->log()),
887 _failure_reason(NULL),
888 _congraph(NULL),
889 NOT_PRODUCT(_igv_printer(NULL) COMMA)
890 _dead_node_list(comp_arena()),
891 _dead_node_count(0),
892 _node_arena(mtCompiler),
893 _old_arena(mtCompiler),
894 _mach_constant_base_node(NULL),
895 _Compile_types(mtCompiler),
896 _initial_gvn(NULL),
897 _for_igvn(NULL),
898 _number_of_mh_late_inlines(0),
899 _native_invokers(),
900 _print_inlining_stream(NULL),
901 _print_inlining_list(NULL),
902 _print_inlining_idx(0),
903 _print_inlining_output(NULL),
904 _replay_inline_data(NULL),
905 _java_calls(0),
906 _inner_loops(0),
907 _interpreter_frame_size(0),
908 _output(NULL),
909 #ifndef PRODUCT
910 _in_dump_cnt(0),
911 #endif
912 _allowed_reasons(0) {
913 C = this;
914
915 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
916 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
917
918 #ifndef PRODUCT
919 set_print_assembly(PrintFrameConverterAssembly);
920 set_parsed_irreducible_loop(false);
921 #else
922 set_print_assembly(false); // Must initialize.
923 #endif
924 set_has_irreducible_loop(false); // no loops
925
926 CompileWrapper cw(this);
927 Init(/*AliasLevel=*/ 0);
928 init_tf((*generator)());
929
930 {
931 // The following is a dummy for the sake of GraphKit::gen_stub
932 Unique_Node_List for_igvn(comp_arena());
933 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
934 PhaseGVN gvn(Thread::current()->resource_area(),255);
935 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
936 gvn.transform_no_reclaim(top());
937
938 GraphKit kit;
939 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
940 }
941
942 NOT_PRODUCT( verify_graph_edges(); )
943
944 Code_Gen();
945 }
946
947 //------------------------------Init-------------------------------------------
948 // Prepare for a single compilation
949 void Compile::Init(int aliaslevel) {
950 _unique = 0;
951 _regalloc = NULL;
952
953 _tf = NULL; // filled in later
954 _top = NULL; // cached later
955 _matcher = NULL; // filled in later
956 _cfg = NULL; // filled in later
957
958 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
959
960 _node_note_array = NULL;
961 _default_node_notes = NULL;
962 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
963
964 _immutable_memory = NULL; // filled in at first inquiry
965
966 // Globally visible Nodes
967 // First set TOP to NULL to give safe behavior during creation of RootNode
968 set_cached_top_node(NULL);
969 set_root(new RootNode());
970 // Now that you have a Root to point to, create the real TOP
971 set_cached_top_node( new ConNode(Type::TOP) );
972 set_recent_alloc(NULL, NULL);
973
974 // Create Debug Information Recorder to record scopes, oopmaps, etc.
975 env()->set_oop_recorder(new OopRecorder(env()->arena()));
976 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
977 env()->set_dependencies(new Dependencies(env()));
978
979 _fixed_slots = 0;
980 set_has_split_ifs(false);
981 set_has_loops(false); // first approximation
982 set_has_stringbuilder(false);
983 set_has_boxed_value(false);
984 _trap_can_recompile = false; // no traps emitted yet
985 _major_progress = true; // start out assuming good things will happen
986 set_has_unsafe_access(false);
987 set_max_vector_size(0);
988 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
989 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
990 set_decompile_count(0);
991
992 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
993 _loop_opts_cnt = LoopOptsCount;
994 set_do_inlining(Inline);
995 set_max_inline_size(MaxInlineSize);
996 set_freq_inline_size(FreqInlineSize);
997 set_do_scheduling(OptoScheduling);
998
999 set_do_vector_loop(false);
1000
1001 if (AllowVectorizeOnDemand) {
1002 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1003 set_do_vector_loop(true);
1004 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());})
1005 } else if (has_method() && method()->name() != 0 &&
1006 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1007 set_do_vector_loop(true);
1008 }
1009 }
1010 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1011 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());})
1012
1013 set_age_code(has_method() && method()->profile_aging());
1014 set_rtm_state(NoRTM); // No RTM lock eliding by default
1015 _max_node_limit = _directive->MaxNodeLimitOption;
1016
1017 #if INCLUDE_RTM_OPT
1018 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1019 int rtm_state = method()->method_data()->rtm_state();
1020 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1021 // Don't generate RTM lock eliding code.
1022 set_rtm_state(NoRTM);
1023 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1024 // Generate RTM lock eliding code without abort ratio calculation code.
1025 set_rtm_state(UseRTM);
1026 } else if (UseRTMDeopt) {
1027 // Generate RTM lock eliding code and include abort ratio calculation
1028 // code if UseRTMDeopt is on.
1029 set_rtm_state(ProfileRTM);
1030 }
1031 }
1032 #endif
1033 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1034 set_clinit_barrier_on_entry(true);
1035 }
1036 if (debug_info()->recording_non_safepoints()) {
1037 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1038 (comp_arena(), 8, 0, NULL));
1039 set_default_node_notes(Node_Notes::make(this));
1040 }
1041
1042 // // -- Initialize types before each compile --
1043 // // Update cached type information
1044 // if( _method && _method->constants() )
1045 // Type::update_loaded_types(_method, _method->constants());
1046
1047 // Init alias_type map.
1048 if (!do_escape_analysis() && aliaslevel == 3) {
1049 aliaslevel = 2; // No unique types without escape analysis
1050 }
1051 _AliasLevel = aliaslevel;
1052 const int grow_ats = 16;
1053 _max_alias_types = grow_ats;
1054 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1055 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1056 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1057 {
1058 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1059 }
1060 // Initialize the first few types.
1061 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1062 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1063 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1064 _num_alias_types = AliasIdxRaw+1;
1065 // Zero out the alias type cache.
1066 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1067 // A NULL adr_type hits in the cache right away. Preload the right answer.
1068 probe_alias_cache(NULL)->_index = AliasIdxTop;
1069
1070 #ifdef ASSERT
1071 _type_verify_symmetry = true;
1072 _phase_optimize_finished = false;
1073 _exception_backedge = false;
1074 #endif
1075 }
1076
1077 //---------------------------init_start----------------------------------------
1078 // Install the StartNode on this compile object.
1079 void Compile::init_start(StartNode* s) {
1080 if (failing())
1081 return; // already failing
1082 assert(s == start(), "");
1083 }
1084
1085 /**
1086 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1087 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1088 * the ideal graph.
1089 */
1090 StartNode* Compile::start() const {
1091 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1092 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1093 Node* start = root()->fast_out(i);
1094 if (start->is_Start()) {
1095 return start->as_Start();
1096 }
1097 }
1098 fatal("Did not find Start node!");
1099 return NULL;
1100 }
1101
1102 //-------------------------------immutable_memory-------------------------------------
1103 // Access immutable memory
1104 Node* Compile::immutable_memory() {
1105 if (_immutable_memory != NULL) {
1106 return _immutable_memory;
1107 }
1108 StartNode* s = start();
1109 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1110 Node *p = s->fast_out(i);
1111 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1112 _immutable_memory = p;
1113 return _immutable_memory;
1114 }
1115 }
1116 ShouldNotReachHere();
1117 return NULL;
1118 }
1119
1120 //----------------------set_cached_top_node------------------------------------
1121 // Install the cached top node, and make sure Node::is_top works correctly.
1122 void Compile::set_cached_top_node(Node* tn) {
1123 if (tn != NULL) verify_top(tn);
1124 Node* old_top = _top;
1125 _top = tn;
1126 // Calling Node::setup_is_top allows the nodes the chance to adjust
1127 // their _out arrays.
1128 if (_top != NULL) _top->setup_is_top();
1129 if (old_top != NULL) old_top->setup_is_top();
1130 assert(_top == NULL || top()->is_top(), "");
1131 }
1132
1133 #ifdef ASSERT
1134 uint Compile::count_live_nodes_by_graph_walk() {
1135 Unique_Node_List useful(comp_arena());
1136 // Get useful node list by walking the graph.
1137 identify_useful_nodes(useful);
1138 return useful.size();
1139 }
1140
1141 void Compile::print_missing_nodes() {
1142
1143 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1144 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1145 return;
1146 }
1147
1148 // This is an expensive function. It is executed only when the user
1149 // specifies VerifyIdealNodeCount option or otherwise knows the
1150 // additional work that needs to be done to identify reachable nodes
1151 // by walking the flow graph and find the missing ones using
1152 // _dead_node_list.
1153
1154 Unique_Node_List useful(comp_arena());
1155 // Get useful node list by walking the graph.
1156 identify_useful_nodes(useful);
1157
1158 uint l_nodes = C->live_nodes();
1159 uint l_nodes_by_walk = useful.size();
1160
1161 if (l_nodes != l_nodes_by_walk) {
1162 if (_log != NULL) {
1163 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1164 _log->stamp();
1165 _log->end_head();
1166 }
1167 VectorSet& useful_member_set = useful.member_set();
1168 int last_idx = l_nodes_by_walk;
1169 for (int i = 0; i < last_idx; i++) {
1170 if (useful_member_set.test(i)) {
1171 if (_dead_node_list.test(i)) {
1172 if (_log != NULL) {
1173 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1174 }
1175 if (PrintIdealNodeCount) {
1176 // Print the log message to tty
1177 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1178 useful.at(i)->dump();
1179 }
1180 }
1181 }
1182 else if (! _dead_node_list.test(i)) {
1183 if (_log != NULL) {
1184 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1185 }
1186 if (PrintIdealNodeCount) {
1187 // Print the log message to tty
1188 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1189 }
1190 }
1191 }
1192 if (_log != NULL) {
1193 _log->tail("mismatched_nodes");
1194 }
1195 }
1196 }
1197 void Compile::record_modified_node(Node* n) {
1198 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1199 _modified_nodes->push(n);
1200 }
1201 }
1202
1203 void Compile::remove_modified_node(Node* n) {
1204 if (_modified_nodes != NULL) {
1205 _modified_nodes->remove(n);
1206 }
1207 }
1208 #endif
1209
1210 #ifndef PRODUCT
1211 void Compile::verify_top(Node* tn) const {
1212 if (tn != NULL) {
1213 assert(tn->is_Con(), "top node must be a constant");
1214 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1215 assert(tn->in(0) != NULL, "must have live top node");
1216 }
1217 }
1218 #endif
1219
1220
1221 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1222
1223 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1224 guarantee(arr != NULL, "");
1225 int num_blocks = arr->length();
1226 if (grow_by < num_blocks) grow_by = num_blocks;
1227 int num_notes = grow_by * _node_notes_block_size;
1228 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1229 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1230 while (num_notes > 0) {
1231 arr->append(notes);
1232 notes += _node_notes_block_size;
1233 num_notes -= _node_notes_block_size;
1234 }
1235 assert(num_notes == 0, "exact multiple, please");
1236 }
1237
1238 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1239 if (source == NULL || dest == NULL) return false;
1240
1241 if (dest->is_Con())
1242 return false; // Do not push debug info onto constants.
1243
1244 #ifdef ASSERT
1245 // Leave a bread crumb trail pointing to the original node:
1246 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1247 dest->set_debug_orig(source);
1248 }
1249 #endif
1250
1251 if (node_note_array() == NULL)
1252 return false; // Not collecting any notes now.
1253
1254 // This is a copy onto a pre-existing node, which may already have notes.
1255 // If both nodes have notes, do not overwrite any pre-existing notes.
1256 Node_Notes* source_notes = node_notes_at(source->_idx);
1257 if (source_notes == NULL || source_notes->is_clear()) return false;
1258 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1259 if (dest_notes == NULL || dest_notes->is_clear()) {
1260 return set_node_notes_at(dest->_idx, source_notes);
1261 }
1262
1263 Node_Notes merged_notes = (*source_notes);
1264 // The order of operations here ensures that dest notes will win...
1265 merged_notes.update_from(dest_notes);
1266 return set_node_notes_at(dest->_idx, &merged_notes);
1267 }
1268
1269
1270 //--------------------------allow_range_check_smearing-------------------------
1271 // Gating condition for coalescing similar range checks.
1272 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1273 // single covering check that is at least as strong as any of them.
1274 // If the optimization succeeds, the simplified (strengthened) range check
1275 // will always succeed. If it fails, we will deopt, and then give up
1276 // on the optimization.
1277 bool Compile::allow_range_check_smearing() const {
1278 // If this method has already thrown a range-check,
1279 // assume it was because we already tried range smearing
1280 // and it failed.
1281 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1282 return !already_trapped;
1283 }
1284
1285
1286 //------------------------------flatten_alias_type-----------------------------
1287 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1288 int offset = tj->offset();
1289 TypePtr::PTR ptr = tj->ptr();
1290
1291 // Known instance (scalarizable allocation) alias only with itself.
1292 bool is_known_inst = tj->isa_oopptr() != NULL &&
1293 tj->is_oopptr()->is_known_instance();
1294
1295 // Process weird unsafe references.
1296 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1297 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1298 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1299 tj = TypeOopPtr::BOTTOM;
1300 ptr = tj->ptr();
1301 offset = tj->offset();
1302 }
1303
1304 // Array pointers need some flattening
1305 const TypeAryPtr *ta = tj->isa_aryptr();
1306 if (ta && ta->is_stable()) {
1307 // Erase stability property for alias analysis.
1308 tj = ta = ta->cast_to_stable(false);
1309 }
1310 if( ta && is_known_inst ) {
1311 if ( offset != Type::OffsetBot &&
1312 offset > arrayOopDesc::length_offset_in_bytes() ) {
1313 offset = Type::OffsetBot; // Flatten constant access into array body only
1314 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1315 }
1316 } else if( ta && _AliasLevel >= 2 ) {
1317 // For arrays indexed by constant indices, we flatten the alias
1318 // space to include all of the array body. Only the header, klass
1319 // and array length can be accessed un-aliased.
1320 if( offset != Type::OffsetBot ) {
1321 if( ta->const_oop() ) { // MethodData* or Method*
1322 offset = Type::OffsetBot; // Flatten constant access into array body
1323 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1324 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1325 // range is OK as-is.
1326 tj = ta = TypeAryPtr::RANGE;
1327 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1328 tj = TypeInstPtr::KLASS; // all klass loads look alike
1329 ta = TypeAryPtr::RANGE; // generic ignored junk
1330 ptr = TypePtr::BotPTR;
1331 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1332 tj = TypeInstPtr::MARK;
1333 ta = TypeAryPtr::RANGE; // generic ignored junk
1334 ptr = TypePtr::BotPTR;
1335 } else { // Random constant offset into array body
1336 offset = Type::OffsetBot; // Flatten constant access into array body
1337 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1338 }
1339 }
1340 // Arrays of fixed size alias with arrays of unknown size.
1341 if (ta->size() != TypeInt::POS) {
1342 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1343 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1344 }
1345 // Arrays of known objects become arrays of unknown objects.
1346 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1347 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1348 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1349 }
1350 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1351 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1352 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1353 }
1354 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1355 // cannot be distinguished by bytecode alone.
1356 if (ta->elem() == TypeInt::BOOL) {
1357 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1358 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1359 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1360 }
1361 // During the 2nd round of IterGVN, NotNull castings are removed.
1362 // Make sure the Bottom and NotNull variants alias the same.
1363 // Also, make sure exact and non-exact variants alias the same.
1364 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1365 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1366 }
1367 }
1368
1369 // Oop pointers need some flattening
1370 const TypeInstPtr *to = tj->isa_instptr();
1371 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1372 ciInstanceKlass *k = to->klass()->as_instance_klass();
1373 if( ptr == TypePtr::Constant ) {
1374 if (to->klass() != ciEnv::current()->Class_klass() ||
1375 offset < k->layout_helper_size_in_bytes()) {
1376 // No constant oop pointers (such as Strings); they alias with
1377 // unknown strings.
1378 assert(!is_known_inst, "not scalarizable allocation");
1379 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1380 }
1381 } else if( is_known_inst ) {
1382 tj = to; // Keep NotNull and klass_is_exact for instance type
1383 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1384 // During the 2nd round of IterGVN, NotNull castings are removed.
1385 // Make sure the Bottom and NotNull variants alias the same.
1386 // Also, make sure exact and non-exact variants alias the same.
1387 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1388 }
1389 if (to->speculative() != NULL) {
1390 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1391 }
1392 // Canonicalize the holder of this field
1393 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1394 // First handle header references such as a LoadKlassNode, even if the
1395 // object's klass is unloaded at compile time (4965979).
1396 if (!is_known_inst) { // Do it only for non-instance types
1397 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1398 }
1399 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) {
1400 // Static fields are in the space above the normal instance
1401 // fields in the java.lang.Class instance.
1402 if (to->klass() != ciEnv::current()->Class_klass()) {
1403 to = NULL;
1404 tj = TypeOopPtr::BOTTOM;
1405 offset = tj->offset();
1406 }
1407 } else {
1408 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1409 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1410 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1411 if( is_known_inst ) {
1412 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1413 } else {
1414 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1415 }
1416 }
1417 }
1418 }
1419
1420 // Klass pointers to object array klasses need some flattening
1421 const TypeKlassPtr *tk = tj->isa_klassptr();
1422 if( tk ) {
1423 // If we are referencing a field within a Klass, we need
1424 // to assume the worst case of an Object. Both exact and
1425 // inexact types must flatten to the same alias class so
1426 // use NotNull as the PTR.
1427 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1428
1429 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1430 TypeInstKlassPtr::OBJECT->klass(),
1431 offset);
1432 }
1433
1434 ciKlass* klass = tk->klass();
1435 if( klass->is_obj_array_klass() ) {
1436 ciKlass* k = TypeAryPtr::OOPS->klass();
1437 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1438 k = TypeInstPtr::BOTTOM->klass();
1439 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1440 }
1441
1442 // Check for precise loads from the primary supertype array and force them
1443 // to the supertype cache alias index. Check for generic array loads from
1444 // the primary supertype array and also force them to the supertype cache
1445 // alias index. Since the same load can reach both, we need to merge
1446 // these 2 disparate memories into the same alias class. Since the
1447 // primary supertype array is read-only, there's no chance of confusion
1448 // where we bypass an array load and an array store.
1449 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1450 if (offset == Type::OffsetBot ||
1451 (offset >= primary_supers_offset &&
1452 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1453 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1454 offset = in_bytes(Klass::secondary_super_cache_offset());
1455 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1456 }
1457 }
1458
1459 // Flatten all Raw pointers together.
1460 if (tj->base() == Type::RawPtr)
1461 tj = TypeRawPtr::BOTTOM;
1462
1463 if (tj->base() == Type::AnyPtr)
1464 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1465
1466 // Flatten all to bottom for now
1467 switch( _AliasLevel ) {
1468 case 0:
1469 tj = TypePtr::BOTTOM;
1470 break;
1471 case 1: // Flatten to: oop, static, field or array
1472 switch (tj->base()) {
1473 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1474 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1475 case Type::AryPtr: // do not distinguish arrays at all
1476 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1477 case Type::KlassPtr:
1478 case Type::AryKlassPtr:
1479 case Type::InstKlassPtr: tj = TypeInstKlassPtr::OBJECT; break;
1480 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1481 default: ShouldNotReachHere();
1482 }
1483 break;
1484 case 2: // No collapsing at level 2; keep all splits
1485 case 3: // No collapsing at level 3; keep all splits
1486 break;
1487 default:
1488 Unimplemented();
1489 }
1490
1491 offset = tj->offset();
1492 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1493
1494 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1495 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1496 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1497 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1498 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1499 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1500 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1501 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1502 assert( tj->ptr() != TypePtr::TopPTR &&
1503 tj->ptr() != TypePtr::AnyNull &&
1504 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1505 // assert( tj->ptr() != TypePtr::Constant ||
1506 // tj->base() == Type::RawPtr ||
1507 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1508
1509 return tj;
1510 }
1511
1512 void Compile::AliasType::Init(int i, const TypePtr* at) {
1513 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1514 _index = i;
1515 _adr_type = at;
1516 _field = NULL;
1517 _element = NULL;
1518 _is_rewritable = true; // default
1519 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1520 if (atoop != NULL && atoop->is_known_instance()) {
1521 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1522 _general_index = Compile::current()->get_alias_index(gt);
1523 } else {
1524 _general_index = 0;
1525 }
1526 }
1527
1528 BasicType Compile::AliasType::basic_type() const {
1529 if (element() != NULL) {
1530 const Type* element = adr_type()->is_aryptr()->elem();
1531 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1532 } if (field() != NULL) {
1533 return field()->layout_type();
1534 } else {
1535 return T_ILLEGAL; // unknown
1536 }
1537 }
1538
1539 //---------------------------------print_on------------------------------------
1540 #ifndef PRODUCT
1541 void Compile::AliasType::print_on(outputStream* st) {
1542 if (index() < 10)
1543 st->print("@ <%d> ", index());
1544 else st->print("@ <%d>", index());
1545 st->print(is_rewritable() ? " " : " RO");
1546 int offset = adr_type()->offset();
1547 if (offset == Type::OffsetBot)
1548 st->print(" +any");
1549 else st->print(" +%-3d", offset);
1550 st->print(" in ");
1551 adr_type()->dump_on(st);
1552 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1553 if (field() != NULL && tjp) {
1554 if (tjp->klass() != field()->holder() ||
1555 tjp->offset() != field()->offset_in_bytes()) {
1556 st->print(" != ");
1557 field()->print();
1558 st->print(" ***");
1559 }
1560 }
1561 }
1562
1563 void print_alias_types() {
1564 Compile* C = Compile::current();
1565 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1566 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1567 C->alias_type(idx)->print_on(tty);
1568 tty->cr();
1569 }
1570 }
1571 #endif
1572
1573
1574 //----------------------------probe_alias_cache--------------------------------
1575 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1576 intptr_t key = (intptr_t) adr_type;
1577 key ^= key >> logAliasCacheSize;
1578 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1579 }
1580
1581
1582 //-----------------------------grow_alias_types--------------------------------
1583 void Compile::grow_alias_types() {
1584 const int old_ats = _max_alias_types; // how many before?
1585 const int new_ats = old_ats; // how many more?
1586 const int grow_ats = old_ats+new_ats; // how many now?
1587 _max_alias_types = grow_ats;
1588 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1589 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1590 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1591 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1592 }
1593
1594
1595 //--------------------------------find_alias_type------------------------------
1596 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1597 if (_AliasLevel == 0)
1598 return alias_type(AliasIdxBot);
1599
1600 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1601 if (ace->_adr_type == adr_type) {
1602 return alias_type(ace->_index);
1603 }
1604
1605 // Handle special cases.
1606 if (adr_type == NULL) return alias_type(AliasIdxTop);
1607 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1608
1609 // Do it the slow way.
1610 const TypePtr* flat = flatten_alias_type(adr_type);
1611
1612 #ifdef ASSERT
1613 {
1614 ResourceMark rm;
1615 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1616 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1617 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1618 Type::str(adr_type));
1619 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1620 const TypeOopPtr* foop = flat->is_oopptr();
1621 // Scalarizable allocations have exact klass always.
1622 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1623 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1624 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1625 Type::str(foop), Type::str(xoop));
1626 }
1627 }
1628 #endif
1629
1630 int idx = AliasIdxTop;
1631 for (int i = 0; i < num_alias_types(); i++) {
1632 if (alias_type(i)->adr_type() == flat) {
1633 idx = i;
1634 break;
1635 }
1636 }
1637
1638 if (idx == AliasIdxTop) {
1639 if (no_create) return NULL;
1640 // Grow the array if necessary.
1641 if (_num_alias_types == _max_alias_types) grow_alias_types();
1642 // Add a new alias type.
1643 idx = _num_alias_types++;
1644 _alias_types[idx]->Init(idx, flat);
1645 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1646 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1647 if (flat->isa_instptr()) {
1648 if (flat->offset() == java_lang_Class::klass_offset()
1649 && flat->is_instptr()->klass() == env()->Class_klass())
1650 alias_type(idx)->set_rewritable(false);
1651 }
1652 if (flat->isa_aryptr()) {
1653 #ifdef ASSERT
1654 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1655 // (T_BYTE has the weakest alignment and size restrictions...)
1656 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1657 #endif
1658 if (flat->offset() == TypePtr::OffsetBot) {
1659 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1660 }
1661 }
1662 if (flat->isa_klassptr()) {
1663 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1664 alias_type(idx)->set_rewritable(false);
1665 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1666 alias_type(idx)->set_rewritable(false);
1667 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1668 alias_type(idx)->set_rewritable(false);
1669 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1670 alias_type(idx)->set_rewritable(false);
1671 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1672 alias_type(idx)->set_rewritable(false);
1673 }
1674 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1675 // but the base pointer type is not distinctive enough to identify
1676 // references into JavaThread.)
1677
1678 // Check for final fields.
1679 const TypeInstPtr* tinst = flat->isa_instptr();
1680 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1681 ciField* field;
1682 if (tinst->const_oop() != NULL &&
1683 tinst->klass() == ciEnv::current()->Class_klass() &&
1684 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) {
1685 // static field
1686 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1687 field = k->get_field_by_offset(tinst->offset(), true);
1688 } else {
1689 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1690 field = k->get_field_by_offset(tinst->offset(), false);
1691 }
1692 assert(field == NULL ||
1693 original_field == NULL ||
1694 (field->holder() == original_field->holder() &&
1695 field->offset() == original_field->offset() &&
1696 field->is_static() == original_field->is_static()), "wrong field?");
1697 // Set field() and is_rewritable() attributes.
1698 if (field != NULL) alias_type(idx)->set_field(field);
1699 }
1700 }
1701
1702 // Fill the cache for next time.
1703 ace->_adr_type = adr_type;
1704 ace->_index = idx;
1705 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1706
1707 // Might as well try to fill the cache for the flattened version, too.
1708 AliasCacheEntry* face = probe_alias_cache(flat);
1709 if (face->_adr_type == NULL) {
1710 face->_adr_type = flat;
1711 face->_index = idx;
1712 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1713 }
1714
1715 return alias_type(idx);
1716 }
1717
1718
1719 Compile::AliasType* Compile::alias_type(ciField* field) {
1720 const TypeOopPtr* t;
1721 if (field->is_static())
1722 t = TypeInstPtr::make(field->holder()->java_mirror());
1723 else
1724 t = TypeOopPtr::make_from_klass_raw(field->holder());
1725 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1726 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1727 return atp;
1728 }
1729
1730
1731 //------------------------------have_alias_type--------------------------------
1732 bool Compile::have_alias_type(const TypePtr* adr_type) {
1733 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1734 if (ace->_adr_type == adr_type) {
1735 return true;
1736 }
1737
1738 // Handle special cases.
1739 if (adr_type == NULL) return true;
1740 if (adr_type == TypePtr::BOTTOM) return true;
1741
1742 return find_alias_type(adr_type, true, NULL) != NULL;
1743 }
1744
1745 //-----------------------------must_alias--------------------------------------
1746 // True if all values of the given address type are in the given alias category.
1747 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1748 if (alias_idx == AliasIdxBot) return true; // the universal category
1749 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1750 if (alias_idx == AliasIdxTop) return false; // the empty category
1751 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1752
1753 // the only remaining possible overlap is identity
1754 int adr_idx = get_alias_index(adr_type);
1755 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1756 assert(adr_idx == alias_idx ||
1757 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1758 && adr_type != TypeOopPtr::BOTTOM),
1759 "should not be testing for overlap with an unsafe pointer");
1760 return adr_idx == alias_idx;
1761 }
1762
1763 //------------------------------can_alias--------------------------------------
1764 // True if any values of the given address type are in the given alias category.
1765 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1766 if (alias_idx == AliasIdxTop) return false; // the empty category
1767 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1768 // Known instance doesn't alias with bottom memory
1769 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1770 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1771
1772 // the only remaining possible overlap is identity
1773 int adr_idx = get_alias_index(adr_type);
1774 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1775 return adr_idx == alias_idx;
1776 }
1777
1778 //---------------------cleanup_loop_predicates-----------------------
1779 // Remove the opaque nodes that protect the predicates so that all unused
1780 // checks and uncommon_traps will be eliminated from the ideal graph
1781 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1782 if (predicate_count()==0) return;
1783 for (int i = predicate_count(); i > 0; i--) {
1784 Node * n = predicate_opaque1_node(i-1);
1785 assert(n->Opcode() == Op_Opaque1, "must be");
1786 igvn.replace_node(n, n->in(1));
1787 }
1788 assert(predicate_count()==0, "should be clean!");
1789 }
1790
1791 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1792 if (!n->for_post_loop_opts_igvn()) {
1793 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1794 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1795 _for_post_loop_igvn.append(n);
1796 }
1797 }
1798
1799 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1800 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1801 _for_post_loop_igvn.remove(n);
1802 }
1803
1804 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1805 // Verify that all previous optimizations produced a valid graph
1806 // at least to this point, even if no loop optimizations were done.
1807 PhaseIdealLoop::verify(igvn);
1808
1809 C->set_post_loop_opts_phase(); // no more loop opts allowed
1810
1811 assert(!C->major_progress(), "not cleared");
1812
1813 if (_for_post_loop_igvn.length() > 0) {
1814 while (_for_post_loop_igvn.length() > 0) {
1815 Node* n = _for_post_loop_igvn.pop();
1816 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1817 igvn._worklist.push(n);
1818 }
1819 igvn.optimize();
1820 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1821
1822 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1823 if (C->major_progress()) {
1824 C->clear_major_progress(); // ensure that major progress is now clear
1825 }
1826 }
1827 }
1828
1829 // StringOpts and late inlining of string methods
1830 void Compile::inline_string_calls(bool parse_time) {
1831 {
1832 // remove useless nodes to make the usage analysis simpler
1833 ResourceMark rm;
1834 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1835 }
1836
1837 {
1838 ResourceMark rm;
1839 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1840 PhaseStringOpts pso(initial_gvn());
1841 print_method(PHASE_AFTER_STRINGOPTS, 3);
1842 }
1843
1844 // now inline anything that we skipped the first time around
1845 if (!parse_time) {
1846 _late_inlines_pos = _late_inlines.length();
1847 }
1848
1849 while (_string_late_inlines.length() > 0) {
1850 CallGenerator* cg = _string_late_inlines.pop();
1851 cg->do_late_inline();
1852 if (failing()) return;
1853 }
1854 _string_late_inlines.trunc_to(0);
1855 }
1856
1857 // Late inlining of boxing methods
1858 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1859 if (_boxing_late_inlines.length() > 0) {
1860 assert(has_boxed_value(), "inconsistent");
1861
1862 PhaseGVN* gvn = initial_gvn();
1863 set_inlining_incrementally(true);
1864
1865 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1866 for_igvn()->clear();
1867 gvn->replace_with(&igvn);
1868
1869 _late_inlines_pos = _late_inlines.length();
1870
1871 while (_boxing_late_inlines.length() > 0) {
1872 CallGenerator* cg = _boxing_late_inlines.pop();
1873 cg->do_late_inline();
1874 if (failing()) return;
1875 }
1876 _boxing_late_inlines.trunc_to(0);
1877
1878 inline_incrementally_cleanup(igvn);
1879
1880 set_inlining_incrementally(false);
1881 }
1882 }
1883
1884 bool Compile::inline_incrementally_one() {
1885 assert(IncrementalInline, "incremental inlining should be on");
1886
1887 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1888
1889 set_inlining_progress(false);
1890 set_do_cleanup(false);
1891
1892 for (int i = 0; i < _late_inlines.length(); i++) {
1893 _late_inlines_pos = i+1;
1894 CallGenerator* cg = _late_inlines.at(i);
1895 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
1896 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
1897 cg->do_late_inline();
1898 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
1899 if (failing()) {
1900 return false;
1901 } else if (inlining_progress()) {
1902 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
1903 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
1904 break; // process one call site at a time
1905 }
1906 } else {
1907 // Ignore late inline direct calls when inlining is not allowed.
1908 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
1909 }
1910 }
1911 // Remove processed elements.
1912 _late_inlines.remove_till(_late_inlines_pos);
1913 _late_inlines_pos = 0;
1914
1915 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
1916
1917 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
1918
1919 set_inlining_progress(false);
1920 set_do_cleanup(false);
1921
1922 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
1923 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
1924 }
1925
1926 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
1927 {
1928 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
1929 ResourceMark rm;
1930 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1931 }
1932 {
1933 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
1934 igvn = PhaseIterGVN(initial_gvn());
1935 igvn.optimize();
1936 }
1937 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
1938 }
1939
1940 // Perform incremental inlining until bound on number of live nodes is reached
1941 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1942 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
1943
1944 set_inlining_incrementally(true);
1945 uint low_live_nodes = 0;
1946
1947 while (_late_inlines.length() > 0) {
1948 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1949 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1950 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
1951 // PhaseIdealLoop is expensive so we only try it once we are
1952 // out of live nodes and we only try it again if the previous
1953 // helped got the number of nodes down significantly
1954 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
1955 if (failing()) return;
1956 low_live_nodes = live_nodes();
1957 _major_progress = true;
1958 }
1959
1960 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1961 bool do_print_inlining = print_inlining() || print_intrinsics();
1962 if (do_print_inlining || log() != NULL) {
1963 // Print inlining message for candidates that we couldn't inline for lack of space.
1964 for (int i = 0; i < _late_inlines.length(); i++) {
1965 CallGenerator* cg = _late_inlines.at(i);
1966 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
1967 if (do_print_inlining) {
1968 cg->print_inlining_late(msg);
1969 }
1970 log_late_inline_failure(cg, msg);
1971 }
1972 }
1973 break; // finish
1974 }
1975 }
1976
1977 for_igvn()->clear();
1978 initial_gvn()->replace_with(&igvn);
1979
1980 while (inline_incrementally_one()) {
1981 assert(!failing(), "inconsistent");
1982 }
1983 if (failing()) return;
1984
1985 inline_incrementally_cleanup(igvn);
1986
1987 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
1988
1989 if (failing()) return;
1990
1991 if (_late_inlines.length() == 0) {
1992 break; // no more progress
1993 }
1994 }
1995 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1996
1997 if (_string_late_inlines.length() > 0) {
1998 assert(has_stringbuilder(), "inconsistent");
1999 for_igvn()->clear();
2000 initial_gvn()->replace_with(&igvn);
2001
2002 inline_string_calls(false);
2003
2004 if (failing()) return;
2005
2006 inline_incrementally_cleanup(igvn);
2007 }
2008
2009 set_inlining_incrementally(false);
2010 }
2011
2012 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2013 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2014 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2015 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2016 // as if "inlining_incrementally() == true" were set.
2017 assert(inlining_incrementally() == false, "not allowed");
2018 assert(_modified_nodes == NULL, "not allowed");
2019 assert(_late_inlines.length() > 0, "sanity");
2020
2021 while (_late_inlines.length() > 0) {
2022 for_igvn()->clear();
2023 initial_gvn()->replace_with(&igvn);
2024
2025 while (inline_incrementally_one()) {
2026 assert(!failing(), "inconsistent");
2027 }
2028 if (failing()) return;
2029
2030 inline_incrementally_cleanup(igvn);
2031 }
2032 }
2033
2034 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2035 if (_loop_opts_cnt > 0) {
2036 debug_only( int cnt = 0; );
2037 while (major_progress() && (_loop_opts_cnt > 0)) {
2038 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2039 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2040 PhaseIdealLoop::optimize(igvn, mode);
2041 _loop_opts_cnt--;
2042 if (failing()) return false;
2043 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2044 }
2045 }
2046 return true;
2047 }
2048
2049 // Remove edges from "root" to each SafePoint at a backward branch.
2050 // They were inserted during parsing (see add_safepoint()) to make
2051 // infinite loops without calls or exceptions visible to root, i.e.,
2052 // useful.
2053 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2054 Node *r = root();
2055 if (r != NULL) {
2056 for (uint i = r->req(); i < r->len(); ++i) {
2057 Node *n = r->in(i);
2058 if (n != NULL && n->is_SafePoint()) {
2059 r->rm_prec(i);
2060 if (n->outcnt() == 0) {
2061 igvn.remove_dead_node(n);
2062 }
2063 --i;
2064 }
2065 }
2066 // Parsing may have added top inputs to the root node (Path
2067 // leading to the Halt node proven dead). Make sure we get a
2068 // chance to clean them up.
2069 igvn._worklist.push(r);
2070 igvn.optimize();
2071 }
2072 }
2073
2074 //------------------------------Optimize---------------------------------------
2075 // Given a graph, optimize it.
2076 void Compile::Optimize() {
2077 TracePhase tp("optimizer", &timers[_t_optimizer]);
2078
2079 #ifndef PRODUCT
2080 if (env()->break_at_compile()) {
2081 BREAKPOINT;
2082 }
2083
2084 #endif
2085
2086 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2087 #ifdef ASSERT
2088 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2089 #endif
2090
2091 ResourceMark rm;
2092
2093 print_inlining_reinit();
2094
2095 NOT_PRODUCT( verify_graph_edges(); )
2096
2097 print_method(PHASE_AFTER_PARSING, 1);
2098
2099 {
2100 // Iterative Global Value Numbering, including ideal transforms
2101 // Initialize IterGVN with types and values from parse-time GVN
2102 PhaseIterGVN igvn(initial_gvn());
2103 #ifdef ASSERT
2104 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2105 #endif
2106 {
2107 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2108 igvn.optimize();
2109 }
2110
2111 if (failing()) return;
2112
2113 print_method(PHASE_ITER_GVN1, 2);
2114
2115 inline_incrementally(igvn);
2116
2117 print_method(PHASE_INCREMENTAL_INLINE, 2);
2118
2119 if (failing()) return;
2120
2121 if (eliminate_boxing()) {
2122 // Inline valueOf() methods now.
2123 inline_boxing_calls(igvn);
2124
2125 if (AlwaysIncrementalInline) {
2126 inline_incrementally(igvn);
2127 }
2128
2129 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2130
2131 if (failing()) return;
2132 }
2133
2134 // Remove the speculative part of types and clean up the graph from
2135 // the extra CastPP nodes whose only purpose is to carry them. Do
2136 // that early so that optimizations are not disrupted by the extra
2137 // CastPP nodes.
2138 remove_speculative_types(igvn);
2139
2140 // No more new expensive nodes will be added to the list from here
2141 // so keep only the actual candidates for optimizations.
2142 cleanup_expensive_nodes(igvn);
2143
2144 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2145 if (EnableVectorSupport && has_vbox_nodes()) {
2146 TracePhase tp("", &timers[_t_vector]);
2147 PhaseVector pv(igvn);
2148 pv.optimize_vector_boxes();
2149
2150 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2151 }
2152 assert(!has_vbox_nodes(), "sanity");
2153
2154 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2155 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2156 initial_gvn()->replace_with(&igvn);
2157 Unique_Node_List* old_worklist = for_igvn();
2158 old_worklist->clear();
2159 Unique_Node_List new_worklist(C->comp_arena());
2160 {
2161 ResourceMark rm;
2162 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2163 }
2164 Unique_Node_List* save_for_igvn = for_igvn();
2165 set_for_igvn(&new_worklist);
2166 igvn = PhaseIterGVN(initial_gvn());
2167 igvn.optimize();
2168 set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2169 }
2170
2171 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2172 // safepoints
2173 remove_root_to_sfpts_edges(igvn);
2174
2175 // Perform escape analysis
2176 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2177 if (has_loops()) {
2178 // Cleanup graph (remove dead nodes).
2179 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2180 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2181 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2182 if (failing()) return;
2183 }
2184 bool progress;
2185 do {
2186 ConnectionGraph::do_analysis(this, &igvn);
2187
2188 if (failing()) return;
2189
2190 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2191
2192 // Optimize out fields loads from scalar replaceable allocations.
2193 igvn.optimize();
2194 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2195
2196 if (failing()) return;
2197
2198 if (congraph() != NULL && macro_count() > 0) {
2199 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2200 PhaseMacroExpand mexp(igvn);
2201 mexp.eliminate_macro_nodes();
2202 igvn.set_delay_transform(false);
2203
2204 igvn.optimize();
2205 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2206
2207 if (failing()) return;
2208 }
2209 progress = do_iterative_escape_analysis() &&
2210 (macro_count() < mcount) &&
2211 ConnectionGraph::has_candidates(this);
2212 // Try again if candidates exist and made progress
2213 // by removing some allocations and/or locks.
2214 } while (progress);
2215 }
2216
2217 // Loop transforms on the ideal graph. Range Check Elimination,
2218 // peeling, unrolling, etc.
2219
2220 // Set loop opts counter
2221 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2222 {
2223 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2224 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2225 _loop_opts_cnt--;
2226 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2227 if (failing()) return;
2228 }
2229 // Loop opts pass if partial peeling occurred in previous pass
2230 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2231 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2232 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2233 _loop_opts_cnt--;
2234 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2235 if (failing()) return;
2236 }
2237 // Loop opts pass for loop-unrolling before CCP
2238 if(major_progress() && (_loop_opts_cnt > 0)) {
2239 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2240 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2241 _loop_opts_cnt--;
2242 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2243 }
2244 if (!failing()) {
2245 // Verify that last round of loop opts produced a valid graph
2246 PhaseIdealLoop::verify(igvn);
2247 }
2248 }
2249 if (failing()) return;
2250
2251 // Conditional Constant Propagation;
2252 PhaseCCP ccp( &igvn );
2253 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2254 {
2255 TracePhase tp("ccp", &timers[_t_ccp]);
2256 ccp.do_transform();
2257 }
2258 print_method(PHASE_CCP1, 2);
2259
2260 assert( true, "Break here to ccp.dump_old2new_map()");
2261
2262 // Iterative Global Value Numbering, including ideal transforms
2263 {
2264 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2265 igvn = ccp;
2266 igvn.optimize();
2267 }
2268 print_method(PHASE_ITER_GVN2, 2);
2269
2270 if (failing()) return;
2271
2272 // Loop transforms on the ideal graph. Range Check Elimination,
2273 // peeling, unrolling, etc.
2274 if (!optimize_loops(igvn, LoopOptsDefault)) {
2275 return;
2276 }
2277
2278 if (failing()) return;
2279
2280 C->clear_major_progress(); // ensure that major progress is now clear
2281
2282 process_for_post_loop_opts_igvn(igvn);
2283
2284 #ifdef ASSERT
2285 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2286 #endif
2287
2288 {
2289 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2290 PhaseMacroExpand mex(igvn);
2291 if (mex.expand_macro_nodes()) {
2292 assert(failing(), "must bail out w/ explicit message");
2293 return;
2294 }
2295 print_method(PHASE_MACRO_EXPANSION, 2);
2296 }
2297
2298 {
2299 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2300 if (bs->expand_barriers(this, igvn)) {
2301 assert(failing(), "must bail out w/ explicit message");
2302 return;
2303 }
2304 print_method(PHASE_BARRIER_EXPANSION, 2);
2305 }
2306
2307 if (C->max_vector_size() > 0) {
2308 C->optimize_logic_cones(igvn);
2309 igvn.optimize();
2310 }
2311
2312 DEBUG_ONLY( _modified_nodes = NULL; )
2313
2314 assert(igvn._worklist.size() == 0, "not empty");
2315
2316 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2317
2318 if (_late_inlines.length() > 0) {
2319 // More opportunities to optimize virtual and MH calls.
2320 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2321 process_late_inline_calls_no_inline(igvn);
2322 }
2323 } // (End scope of igvn; run destructor if necessary for asserts.)
2324
2325 check_no_dead_use();
2326
2327 process_print_inlining();
2328
2329 // A method with only infinite loops has no edges entering loops from root
2330 {
2331 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2332 if (final_graph_reshaping()) {
2333 assert(failing(), "must bail out w/ explicit message");
2334 return;
2335 }
2336 }
2337
2338 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2339 DEBUG_ONLY(set_phase_optimize_finished();)
2340 }
2341
2342 #ifdef ASSERT
2343 void Compile::check_no_dead_use() const {
2344 ResourceMark rm;
2345 Unique_Node_List wq;
2346 wq.push(root());
2347 for (uint i = 0; i < wq.size(); ++i) {
2348 Node* n = wq.at(i);
2349 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2350 Node* u = n->fast_out(j);
2351 if (u->outcnt() == 0 && !u->is_Con()) {
2352 u->dump();
2353 fatal("no reachable node should have no use");
2354 }
2355 wq.push(u);
2356 }
2357 }
2358 }
2359 #endif
2360
2361 void Compile::inline_vector_reboxing_calls() {
2362 if (C->_vector_reboxing_late_inlines.length() > 0) {
2363 _late_inlines_pos = C->_late_inlines.length();
2364 while (_vector_reboxing_late_inlines.length() > 0) {
2365 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2366 cg->do_late_inline();
2367 if (failing()) return;
2368 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2369 }
2370 _vector_reboxing_late_inlines.trunc_to(0);
2371 }
2372 }
2373
2374 bool Compile::has_vbox_nodes() {
2375 if (C->_vector_reboxing_late_inlines.length() > 0) {
2376 return true;
2377 }
2378 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2379 Node * n = C->macro_node(macro_idx);
2380 assert(n->is_macro(), "only macro nodes expected here");
2381 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2382 return true;
2383 }
2384 }
2385 return false;
2386 }
2387
2388 //---------------------------- Bitwise operation packing optimization ---------------------------
2389
2390 static bool is_vector_unary_bitwise_op(Node* n) {
2391 return n->Opcode() == Op_XorV &&
2392 VectorNode::is_vector_bitwise_not_pattern(n);
2393 }
2394
2395 static bool is_vector_binary_bitwise_op(Node* n) {
2396 switch (n->Opcode()) {
2397 case Op_AndV:
2398 case Op_OrV:
2399 return true;
2400
2401 case Op_XorV:
2402 return !is_vector_unary_bitwise_op(n);
2403
2404 default:
2405 return false;
2406 }
2407 }
2408
2409 static bool is_vector_ternary_bitwise_op(Node* n) {
2410 return n->Opcode() == Op_MacroLogicV;
2411 }
2412
2413 static bool is_vector_bitwise_op(Node* n) {
2414 return is_vector_unary_bitwise_op(n) ||
2415 is_vector_binary_bitwise_op(n) ||
2416 is_vector_ternary_bitwise_op(n);
2417 }
2418
2419 static bool is_vector_bitwise_cone_root(Node* n) {
2420 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2421 return false;
2422 }
2423 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2424 if (is_vector_bitwise_op(n->fast_out(i))) {
2425 return false;
2426 }
2427 }
2428 return true;
2429 }
2430
2431 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2432 uint cnt = 0;
2433 if (is_vector_bitwise_op(n)) {
2434 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2435 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2436 for (uint i = 1; i < inp_cnt; i++) {
2437 Node* in = n->in(i);
2438 bool skip = VectorNode::is_all_ones_vector(in);
2439 if (!skip && !inputs.member(in)) {
2440 inputs.push(in);
2441 cnt++;
2442 }
2443 }
2444 assert(cnt <= 1, "not unary");
2445 } else {
2446 uint last_req = inp_cnt;
2447 if (is_vector_ternary_bitwise_op(n)) {
2448 last_req = inp_cnt - 1; // skip last input
2449 }
2450 for (uint i = 1; i < last_req; i++) {
2451 Node* def = n->in(i);
2452 if (!inputs.member(def)) {
2453 inputs.push(def);
2454 cnt++;
2455 }
2456 }
2457 }
2458 } else { // not a bitwise operations
2459 if (!inputs.member(n)) {
2460 inputs.push(n);
2461 cnt++;
2462 }
2463 }
2464 return cnt;
2465 }
2466
2467 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2468 Unique_Node_List useful_nodes;
2469 C->identify_useful_nodes(useful_nodes);
2470
2471 for (uint i = 0; i < useful_nodes.size(); i++) {
2472 Node* n = useful_nodes.at(i);
2473 if (is_vector_bitwise_cone_root(n)) {
2474 list.push(n);
2475 }
2476 }
2477 }
2478
2479 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2480 const TypeVect* vt,
2481 Unique_Node_List& partition,
2482 Unique_Node_List& inputs) {
2483 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2484 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2485 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2486
2487 Node* in1 = inputs.at(0);
2488 Node* in2 = inputs.at(1);
2489 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2490
2491 uint func = compute_truth_table(partition, inputs);
2492
2493 Node* pn = partition.at(partition.size() - 1);
2494 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
2495 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2496 }
2497
2498 static uint extract_bit(uint func, uint pos) {
2499 return (func & (1 << pos)) >> pos;
2500 }
2501
2502 //
2503 // A macro logic node represents a truth table. It has 4 inputs,
2504 // First three inputs corresponds to 3 columns of a truth table
2505 // and fourth input captures the logic function.
2506 //
2507 // eg. fn = (in1 AND in2) OR in3;
2508 //
2509 // MacroNode(in1,in2,in3,fn)
2510 //
2511 // -----------------
2512 // in1 in2 in3 fn
2513 // -----------------
2514 // 0 0 0 0
2515 // 0 0 1 1
2516 // 0 1 0 0
2517 // 0 1 1 1
2518 // 1 0 0 0
2519 // 1 0 1 1
2520 // 1 1 0 1
2521 // 1 1 1 1
2522 //
2523
2524 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2525 int res = 0;
2526 for (int i = 0; i < 8; i++) {
2527 int bit1 = extract_bit(in1, i);
2528 int bit2 = extract_bit(in2, i);
2529 int bit3 = extract_bit(in3, i);
2530
2531 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2532 int func_bit = extract_bit(func, func_bit_pos);
2533
2534 res |= func_bit << i;
2535 }
2536 return res;
2537 }
2538
2539 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2540 assert(n != NULL, "");
2541 assert(eval_map.contains(n), "absent");
2542 return *(eval_map.get(n));
2543 }
2544
2545 static void eval_operands(Node* n,
2546 uint& func1, uint& func2, uint& func3,
2547 ResourceHashtable<Node*,uint>& eval_map) {
2548 assert(is_vector_bitwise_op(n), "");
2549
2550 if (is_vector_unary_bitwise_op(n)) {
2551 Node* opnd = n->in(1);
2552 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2553 opnd = n->in(2);
2554 }
2555 func1 = eval_operand(opnd, eval_map);
2556 } else if (is_vector_binary_bitwise_op(n)) {
2557 func1 = eval_operand(n->in(1), eval_map);
2558 func2 = eval_operand(n->in(2), eval_map);
2559 } else {
2560 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2561 func1 = eval_operand(n->in(1), eval_map);
2562 func2 = eval_operand(n->in(2), eval_map);
2563 func3 = eval_operand(n->in(3), eval_map);
2564 }
2565 }
2566
2567 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2568 assert(inputs.size() <= 3, "sanity");
2569 ResourceMark rm;
2570 uint res = 0;
2571 ResourceHashtable<Node*,uint> eval_map;
2572
2573 // Populate precomputed functions for inputs.
2574 // Each input corresponds to one column of 3 input truth-table.
2575 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
2576 0xCC, // (_, b, _) -> b
2577 0xF0 }; // (a, _, _) -> a
2578 for (uint i = 0; i < inputs.size(); i++) {
2579 eval_map.put(inputs.at(i), input_funcs[2-i]);
2580 }
2581
2582 for (uint i = 0; i < partition.size(); i++) {
2583 Node* n = partition.at(i);
2584
2585 uint func1 = 0, func2 = 0, func3 = 0;
2586 eval_operands(n, func1, func2, func3, eval_map);
2587
2588 switch (n->Opcode()) {
2589 case Op_OrV:
2590 assert(func3 == 0, "not binary");
2591 res = func1 | func2;
2592 break;
2593 case Op_AndV:
2594 assert(func3 == 0, "not binary");
2595 res = func1 & func2;
2596 break;
2597 case Op_XorV:
2598 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2599 assert(func2 == 0 && func3 == 0, "not unary");
2600 res = (~func1) & 0xFF;
2601 } else {
2602 assert(func3 == 0, "not binary");
2603 res = func1 ^ func2;
2604 }
2605 break;
2606 case Op_MacroLogicV:
2607 // Ordering of inputs may change during evaluation of sub-tree
2608 // containing MacroLogic node as a child node, thus a re-evaluation
2609 // makes sure that function is evaluated in context of current
2610 // inputs.
2611 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2612 break;
2613
2614 default: assert(false, "not supported: %s", n->Name());
2615 }
2616 assert(res <= 0xFF, "invalid");
2617 eval_map.put(n, res);
2618 }
2619 return res;
2620 }
2621
2622 // Criteria under which nodes gets packed into a macro logic node:-
2623 // 1) Parent and both child nodes are all unmasked or masked with
2624 // same predicates.
2625 // 2) Masked parent can be packed with left child if it is predicated
2626 // and both have same predicates.
2627 // 3) Masked parent can be packed with right child if its un-predicated
2628 // or has matching predication condition.
2629 // 4) An unmasked parent can be packed with an unmasked child.
2630 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2631 assert(partition.size() == 0, "not empty");
2632 assert(inputs.size() == 0, "not empty");
2633 if (is_vector_ternary_bitwise_op(n)) {
2634 return false;
2635 }
2636
2637 bool is_unary_op = is_vector_unary_bitwise_op(n);
2638 if (is_unary_op) {
2639 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2640 return false; // too few inputs
2641 }
2642
2643 bool pack_left_child = true;
2644 bool pack_right_child = true;
2645
2646 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2647 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2648
2649 int left_child_input_cnt = 0;
2650 int right_child_input_cnt = 0;
2651
2652 bool parent_is_predicated = n->is_predicated_vector();
2653 bool left_child_predicated = n->in(1)->is_predicated_vector();
2654 bool right_child_predicated = n->in(2)->is_predicated_vector();
2655
2656 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : NULL;
2657 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
2658 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
2659
2660 do {
2661 if (pack_left_child && left_child_LOP &&
2662 ((!parent_is_predicated && !left_child_predicated) ||
2663 ((parent_is_predicated && left_child_predicated &&
2664 parent_pred == left_child_pred)))) {
2665 partition.push(n->in(1));
2666 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2667 } else {
2668 inputs.push(n->in(1));
2669 left_child_input_cnt = 1;
2670 }
2671
2672 if (pack_right_child && right_child_LOP &&
2673 (!right_child_predicated ||
2674 (right_child_predicated && parent_is_predicated &&
2675 parent_pred == right_child_pred))) {
2676 partition.push(n->in(2));
2677 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2678 } else {
2679 inputs.push(n->in(2));
2680 right_child_input_cnt = 1;
2681 }
2682
2683 if (inputs.size() > 3) {
2684 assert(partition.size() > 0, "");
2685 inputs.clear();
2686 partition.clear();
2687 if (left_child_input_cnt > right_child_input_cnt) {
2688 pack_left_child = false;
2689 } else {
2690 pack_right_child = false;
2691 }
2692 } else {
2693 break;
2694 }
2695 } while(true);
2696
2697 if(partition.size()) {
2698 partition.push(n);
2699 }
2700
2701 return (partition.size() == 2 || partition.size() == 3) &&
2702 (inputs.size() == 2 || inputs.size() == 3);
2703 }
2704
2705 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2706 assert(is_vector_bitwise_op(n), "not a root");
2707
2708 visited.set(n->_idx);
2709
2710 // 1) Do a DFS walk over the logic cone.
2711 for (uint i = 1; i < n->req(); i++) {
2712 Node* in = n->in(i);
2713 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2714 process_logic_cone_root(igvn, in, visited);
2715 }
2716 }
2717
2718 // 2) Bottom up traversal: Merge node[s] with
2719 // the parent to form macro logic node.
2720 Unique_Node_List partition;
2721 Unique_Node_List inputs;
2722 if (compute_logic_cone(n, partition, inputs)) {
2723 const TypeVect* vt = n->bottom_type()->is_vect();
2724 Node* pn = partition.at(partition.size() - 1);
2725 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
2726 if (mask == NULL ||
2727 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2728 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2729 #ifdef ASSERT
2730 if (TraceNewVectors) {
2731 tty->print("new Vector node: ");
2732 macro_logic->dump();
2733 }
2734 #endif
2735 igvn.replace_node(n, macro_logic);
2736 }
2737 }
2738 }
2739
2740 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2741 ResourceMark rm;
2742 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2743 Unique_Node_List list;
2744 collect_logic_cone_roots(list);
2745
2746 while (list.size() > 0) {
2747 Node* n = list.pop();
2748 const TypeVect* vt = n->bottom_type()->is_vect();
2749 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2750 if (supported) {
2751 VectorSet visited(comp_arena());
2752 process_logic_cone_root(igvn, n, visited);
2753 }
2754 }
2755 }
2756 }
2757
2758 //------------------------------Code_Gen---------------------------------------
2759 // Given a graph, generate code for it
2760 void Compile::Code_Gen() {
2761 if (failing()) {
2762 return;
2763 }
2764
2765 // Perform instruction selection. You might think we could reclaim Matcher
2766 // memory PDQ, but actually the Matcher is used in generating spill code.
2767 // Internals of the Matcher (including some VectorSets) must remain live
2768 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2769 // set a bit in reclaimed memory.
2770
2771 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2772 // nodes. Mapping is only valid at the root of each matched subtree.
2773 NOT_PRODUCT( verify_graph_edges(); )
2774
2775 Matcher matcher;
2776 _matcher = &matcher;
2777 {
2778 TracePhase tp("matcher", &timers[_t_matcher]);
2779 matcher.match();
2780 if (failing()) {
2781 return;
2782 }
2783 }
2784 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2785 // nodes. Mapping is only valid at the root of each matched subtree.
2786 NOT_PRODUCT( verify_graph_edges(); )
2787
2788 // If you have too many nodes, or if matching has failed, bail out
2789 check_node_count(0, "out of nodes matching instructions");
2790 if (failing()) {
2791 return;
2792 }
2793
2794 print_method(PHASE_MATCHING, 2);
2795
2796 // Build a proper-looking CFG
2797 PhaseCFG cfg(node_arena(), root(), matcher);
2798 _cfg = &cfg;
2799 {
2800 TracePhase tp("scheduler", &timers[_t_scheduler]);
2801 bool success = cfg.do_global_code_motion();
2802 if (!success) {
2803 return;
2804 }
2805
2806 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2807 NOT_PRODUCT( verify_graph_edges(); )
2808 cfg.verify();
2809 }
2810
2811 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2812 _regalloc = ®alloc;
2813 {
2814 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2815 // Perform register allocation. After Chaitin, use-def chains are
2816 // no longer accurate (at spill code) and so must be ignored.
2817 // Node->LRG->reg mappings are still accurate.
2818 _regalloc->Register_Allocate();
2819
2820 // Bail out if the allocator builds too many nodes
2821 if (failing()) {
2822 return;
2823 }
2824 }
2825
2826 // Prior to register allocation we kept empty basic blocks in case the
2827 // the allocator needed a place to spill. After register allocation we
2828 // are not adding any new instructions. If any basic block is empty, we
2829 // can now safely remove it.
2830 {
2831 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2832 cfg.remove_empty_blocks();
2833 if (do_freq_based_layout()) {
2834 PhaseBlockLayout layout(cfg);
2835 } else {
2836 cfg.set_loop_alignment();
2837 }
2838 cfg.fixup_flow();
2839 }
2840
2841 // Apply peephole optimizations
2842 if( OptoPeephole ) {
2843 TracePhase tp("peephole", &timers[_t_peephole]);
2844 PhasePeephole peep( _regalloc, cfg);
2845 peep.do_transform();
2846 }
2847
2848 // Do late expand if CPU requires this.
2849 if (Matcher::require_postalloc_expand) {
2850 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2851 cfg.postalloc_expand(_regalloc);
2852 }
2853
2854 // Convert Nodes to instruction bits in a buffer
2855 {
2856 TracePhase tp("output", &timers[_t_output]);
2857 PhaseOutput output;
2858 output.Output();
2859 if (failing()) return;
2860 output.install();
2861 }
2862
2863 print_method(PHASE_FINAL_CODE, 1);
2864
2865 // He's dead, Jim.
2866 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2867 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2868 }
2869
2870 //------------------------------Final_Reshape_Counts---------------------------
2871 // This class defines counters to help identify when a method
2872 // may/must be executed using hardware with only 24-bit precision.
2873 struct Final_Reshape_Counts : public StackObj {
2874 int _call_count; // count non-inlined 'common' calls
2875 int _float_count; // count float ops requiring 24-bit precision
2876 int _double_count; // count double ops requiring more precision
2877 int _java_call_count; // count non-inlined 'java' calls
2878 int _inner_loop_count; // count loops which need alignment
2879 VectorSet _visited; // Visitation flags
2880 Node_List _tests; // Set of IfNodes & PCTableNodes
2881
2882 Final_Reshape_Counts() :
2883 _call_count(0), _float_count(0), _double_count(0),
2884 _java_call_count(0), _inner_loop_count(0) { }
2885
2886 void inc_call_count () { _call_count ++; }
2887 void inc_float_count () { _float_count ++; }
2888 void inc_double_count() { _double_count++; }
2889 void inc_java_call_count() { _java_call_count++; }
2890 void inc_inner_loop_count() { _inner_loop_count++; }
2891
2892 int get_call_count () const { return _call_count ; }
2893 int get_float_count () const { return _float_count ; }
2894 int get_double_count() const { return _double_count; }
2895 int get_java_call_count() const { return _java_call_count; }
2896 int get_inner_loop_count() const { return _inner_loop_count; }
2897 };
2898
2899 // Eliminate trivially redundant StoreCMs and accumulate their
2900 // precedence edges.
2901 void Compile::eliminate_redundant_card_marks(Node* n) {
2902 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2903 if (n->in(MemNode::Address)->outcnt() > 1) {
2904 // There are multiple users of the same address so it might be
2905 // possible to eliminate some of the StoreCMs
2906 Node* mem = n->in(MemNode::Memory);
2907 Node* adr = n->in(MemNode::Address);
2908 Node* val = n->in(MemNode::ValueIn);
2909 Node* prev = n;
2910 bool done = false;
2911 // Walk the chain of StoreCMs eliminating ones that match. As
2912 // long as it's a chain of single users then the optimization is
2913 // safe. Eliminating partially redundant StoreCMs would require
2914 // cloning copies down the other paths.
2915 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2916 if (adr == mem->in(MemNode::Address) &&
2917 val == mem->in(MemNode::ValueIn)) {
2918 // redundant StoreCM
2919 if (mem->req() > MemNode::OopStore) {
2920 // Hasn't been processed by this code yet.
2921 n->add_prec(mem->in(MemNode::OopStore));
2922 } else {
2923 // Already converted to precedence edge
2924 for (uint i = mem->req(); i < mem->len(); i++) {
2925 // Accumulate any precedence edges
2926 if (mem->in(i) != NULL) {
2927 n->add_prec(mem->in(i));
2928 }
2929 }
2930 // Everything above this point has been processed.
2931 done = true;
2932 }
2933 // Eliminate the previous StoreCM
2934 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2935 assert(mem->outcnt() == 0, "should be dead");
2936 mem->disconnect_inputs(this);
2937 } else {
2938 prev = mem;
2939 }
2940 mem = prev->in(MemNode::Memory);
2941 }
2942 }
2943 }
2944
2945 //------------------------------final_graph_reshaping_impl----------------------
2946 // Implement items 1-5 from final_graph_reshaping below.
2947 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2948
2949 if ( n->outcnt() == 0 ) return; // dead node
2950 uint nop = n->Opcode();
2951
2952 // Check for 2-input instruction with "last use" on right input.
2953 // Swap to left input. Implements item (2).
2954 if( n->req() == 3 && // two-input instruction
2955 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2956 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2957 n->in(2)->outcnt() == 1 &&// right use IS a last use
2958 !n->in(2)->is_Con() ) { // right use is not a constant
2959 // Check for commutative opcode
2960 switch( nop ) {
2961 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2962 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
2963 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
2964 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2965 case Op_AndL: case Op_XorL: case Op_OrL:
2966 case Op_AndI: case Op_XorI: case Op_OrI: {
2967 // Move "last use" input to left by swapping inputs
2968 n->swap_edges(1, 2);
2969 break;
2970 }
2971 default:
2972 break;
2973 }
2974 }
2975
2976 #ifdef ASSERT
2977 if( n->is_Mem() ) {
2978 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2979 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2980 // oop will be recorded in oop map if load crosses safepoint
2981 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2982 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2983 "raw memory operations should have control edge");
2984 }
2985 if (n->is_MemBar()) {
2986 MemBarNode* mb = n->as_MemBar();
2987 if (mb->trailing_store() || mb->trailing_load_store()) {
2988 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2989 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2990 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2991 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2992 } else if (mb->leading()) {
2993 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2994 }
2995 }
2996 #endif
2997 // Count FPU ops and common calls, implements item (3)
2998 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2999 if (!gc_handled) {
3000 final_graph_reshaping_main_switch(n, frc, nop);
3001 }
3002
3003 // Collect CFG split points
3004 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3005 frc._tests.push(n);
3006 }
3007 }
3008
3009 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3010 switch( nop ) {
3011 // Count all float operations that may use FPU
3012 case Op_AddF:
3013 case Op_SubF:
3014 case Op_MulF:
3015 case Op_DivF:
3016 case Op_NegF:
3017 case Op_ModF:
3018 case Op_ConvI2F:
3019 case Op_ConF:
3020 case Op_CmpF:
3021 case Op_CmpF3:
3022 case Op_StoreF:
3023 case Op_LoadF:
3024 // case Op_ConvL2F: // longs are split into 32-bit halves
3025 frc.inc_float_count();
3026 break;
3027
3028 case Op_ConvF2D:
3029 case Op_ConvD2F:
3030 frc.inc_float_count();
3031 frc.inc_double_count();
3032 break;
3033
3034 // Count all double operations that may use FPU
3035 case Op_AddD:
3036 case Op_SubD:
3037 case Op_MulD:
3038 case Op_DivD:
3039 case Op_NegD:
3040 case Op_ModD:
3041 case Op_ConvI2D:
3042 case Op_ConvD2I:
3043 // case Op_ConvL2D: // handled by leaf call
3044 // case Op_ConvD2L: // handled by leaf call
3045 case Op_ConD:
3046 case Op_CmpD:
3047 case Op_CmpD3:
3048 case Op_StoreD:
3049 case Op_LoadD:
3050 case Op_LoadD_unaligned:
3051 frc.inc_double_count();
3052 break;
3053 case Op_Opaque1: // Remove Opaque Nodes before matching
3054 case Op_Opaque2: // Remove Opaque Nodes before matching
3055 case Op_Opaque3:
3056 n->subsume_by(n->in(1), this);
3057 break;
3058 case Op_CallStaticJava:
3059 case Op_CallJava:
3060 case Op_CallDynamicJava:
3061 frc.inc_java_call_count(); // Count java call site;
3062 case Op_CallRuntime:
3063 case Op_CallLeaf:
3064 case Op_CallLeafVector:
3065 case Op_CallNative:
3066 case Op_CallLeafNoFP: {
3067 assert (n->is_Call(), "");
3068 CallNode *call = n->as_Call();
3069 // Count call sites where the FP mode bit would have to be flipped.
3070 // Do not count uncommon runtime calls:
3071 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3072 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3073 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3074 frc.inc_call_count(); // Count the call site
3075 } else { // See if uncommon argument is shared
3076 Node *n = call->in(TypeFunc::Parms);
3077 int nop = n->Opcode();
3078 // Clone shared simple arguments to uncommon calls, item (1).
3079 if (n->outcnt() > 1 &&
3080 !n->is_Proj() &&
3081 nop != Op_CreateEx &&
3082 nop != Op_CheckCastPP &&
3083 nop != Op_DecodeN &&
3084 nop != Op_DecodeNKlass &&
3085 !n->is_Mem() &&
3086 !n->is_Phi()) {
3087 Node *x = n->clone();
3088 call->set_req(TypeFunc::Parms, x);
3089 }
3090 }
3091 break;
3092 }
3093
3094 case Op_StoreCM:
3095 {
3096 // Convert OopStore dependence into precedence edge
3097 Node* prec = n->in(MemNode::OopStore);
3098 n->del_req(MemNode::OopStore);
3099 n->add_prec(prec);
3100 eliminate_redundant_card_marks(n);
3101 }
3102
3103 // fall through
3104
3105 case Op_StoreB:
3106 case Op_StoreC:
3107 case Op_StorePConditional:
3108 case Op_StoreI:
3109 case Op_StoreL:
3110 case Op_StoreIConditional:
3111 case Op_StoreLConditional:
3112 case Op_CompareAndSwapB:
3113 case Op_CompareAndSwapS:
3114 case Op_CompareAndSwapI:
3115 case Op_CompareAndSwapL:
3116 case Op_CompareAndSwapP:
3117 case Op_CompareAndSwapN:
3118 case Op_WeakCompareAndSwapB:
3119 case Op_WeakCompareAndSwapS:
3120 case Op_WeakCompareAndSwapI:
3121 case Op_WeakCompareAndSwapL:
3122 case Op_WeakCompareAndSwapP:
3123 case Op_WeakCompareAndSwapN:
3124 case Op_CompareAndExchangeB:
3125 case Op_CompareAndExchangeS:
3126 case Op_CompareAndExchangeI:
3127 case Op_CompareAndExchangeL:
3128 case Op_CompareAndExchangeP:
3129 case Op_CompareAndExchangeN:
3130 case Op_GetAndAddS:
3131 case Op_GetAndAddB:
3132 case Op_GetAndAddI:
3133 case Op_GetAndAddL:
3134 case Op_GetAndSetS:
3135 case Op_GetAndSetB:
3136 case Op_GetAndSetI:
3137 case Op_GetAndSetL:
3138 case Op_GetAndSetP:
3139 case Op_GetAndSetN:
3140 case Op_StoreP:
3141 case Op_StoreN:
3142 case Op_StoreNKlass:
3143 case Op_LoadB:
3144 case Op_LoadUB:
3145 case Op_LoadUS:
3146 case Op_LoadI:
3147 case Op_LoadKlass:
3148 case Op_LoadNKlass:
3149 case Op_LoadL:
3150 case Op_LoadL_unaligned:
3151 case Op_LoadPLocked:
3152 case Op_LoadP:
3153 case Op_LoadN:
3154 case Op_LoadRange:
3155 case Op_LoadS:
3156 break;
3157
3158 case Op_AddP: { // Assert sane base pointers
3159 Node *addp = n->in(AddPNode::Address);
3160 assert( !addp->is_AddP() ||
3161 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3162 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3163 "Base pointers must match (addp %u)", addp->_idx );
3164 #ifdef _LP64
3165 if ((UseCompressedOops || UseCompressedClassPointers) &&
3166 addp->Opcode() == Op_ConP &&
3167 addp == n->in(AddPNode::Base) &&
3168 n->in(AddPNode::Offset)->is_Con()) {
3169 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3170 // on the platform and on the compressed oops mode.
3171 // Use addressing with narrow klass to load with offset on x86.
3172 // Some platforms can use the constant pool to load ConP.
3173 // Do this transformation here since IGVN will convert ConN back to ConP.
3174 const Type* t = addp->bottom_type();
3175 bool is_oop = t->isa_oopptr() != NULL;
3176 bool is_klass = t->isa_klassptr() != NULL;
3177
3178 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3179 (is_klass && Matcher::const_klass_prefer_decode())) {
3180 Node* nn = NULL;
3181
3182 int op = is_oop ? Op_ConN : Op_ConNKlass;
3183
3184 // Look for existing ConN node of the same exact type.
3185 Node* r = root();
3186 uint cnt = r->outcnt();
3187 for (uint i = 0; i < cnt; i++) {
3188 Node* m = r->raw_out(i);
3189 if (m!= NULL && m->Opcode() == op &&
3190 m->bottom_type()->make_ptr() == t) {
3191 nn = m;
3192 break;
3193 }
3194 }
3195 if (nn != NULL) {
3196 // Decode a narrow oop to match address
3197 // [R12 + narrow_oop_reg<<3 + offset]
3198 if (is_oop) {
3199 nn = new DecodeNNode(nn, t);
3200 } else {
3201 nn = new DecodeNKlassNode(nn, t);
3202 }
3203 // Check for succeeding AddP which uses the same Base.
3204 // Otherwise we will run into the assertion above when visiting that guy.
3205 for (uint i = 0; i < n->outcnt(); ++i) {
3206 Node *out_i = n->raw_out(i);
3207 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3208 out_i->set_req(AddPNode::Base, nn);
3209 #ifdef ASSERT
3210 for (uint j = 0; j < out_i->outcnt(); ++j) {
3211 Node *out_j = out_i->raw_out(j);
3212 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3213 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3214 }
3215 #endif
3216 }
3217 }
3218 n->set_req(AddPNode::Base, nn);
3219 n->set_req(AddPNode::Address, nn);
3220 if (addp->outcnt() == 0) {
3221 addp->disconnect_inputs(this);
3222 }
3223 }
3224 }
3225 }
3226 #endif
3227 break;
3228 }
3229
3230 case Op_CastPP: {
3231 // Remove CastPP nodes to gain more freedom during scheduling but
3232 // keep the dependency they encode as control or precedence edges
3233 // (if control is set already) on memory operations. Some CastPP
3234 // nodes don't have a control (don't carry a dependency): skip
3235 // those.
3236 if (n->in(0) != NULL) {
3237 ResourceMark rm;
3238 Unique_Node_List wq;
3239 wq.push(n);
3240 for (uint next = 0; next < wq.size(); ++next) {
3241 Node *m = wq.at(next);
3242 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3243 Node* use = m->fast_out(i);
3244 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3245 use->ensure_control_or_add_prec(n->in(0));
3246 } else {
3247 switch(use->Opcode()) {
3248 case Op_AddP:
3249 case Op_DecodeN:
3250 case Op_DecodeNKlass:
3251 case Op_CheckCastPP:
3252 case Op_CastPP:
3253 wq.push(use);
3254 break;
3255 }
3256 }
3257 }
3258 }
3259 }
3260 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3261 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3262 Node* in1 = n->in(1);
3263 const Type* t = n->bottom_type();
3264 Node* new_in1 = in1->clone();
3265 new_in1->as_DecodeN()->set_type(t);
3266
3267 if (!Matcher::narrow_oop_use_complex_address()) {
3268 //
3269 // x86, ARM and friends can handle 2 adds in addressing mode
3270 // and Matcher can fold a DecodeN node into address by using
3271 // a narrow oop directly and do implicit NULL check in address:
3272 //
3273 // [R12 + narrow_oop_reg<<3 + offset]
3274 // NullCheck narrow_oop_reg
3275 //
3276 // On other platforms (Sparc) we have to keep new DecodeN node and
3277 // use it to do implicit NULL check in address:
3278 //
3279 // decode_not_null narrow_oop_reg, base_reg
3280 // [base_reg + offset]
3281 // NullCheck base_reg
3282 //
3283 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3284 // to keep the information to which NULL check the new DecodeN node
3285 // corresponds to use it as value in implicit_null_check().
3286 //
3287 new_in1->set_req(0, n->in(0));
3288 }
3289
3290 n->subsume_by(new_in1, this);
3291 if (in1->outcnt() == 0) {
3292 in1->disconnect_inputs(this);
3293 }
3294 } else {
3295 n->subsume_by(n->in(1), this);
3296 if (n->outcnt() == 0) {
3297 n->disconnect_inputs(this);
3298 }
3299 }
3300 break;
3301 }
3302 #ifdef _LP64
3303 case Op_CmpP:
3304 // Do this transformation here to preserve CmpPNode::sub() and
3305 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3306 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3307 Node* in1 = n->in(1);
3308 Node* in2 = n->in(2);
3309 if (!in1->is_DecodeNarrowPtr()) {
3310 in2 = in1;
3311 in1 = n->in(2);
3312 }
3313 assert(in1->is_DecodeNarrowPtr(), "sanity");
3314
3315 Node* new_in2 = NULL;
3316 if (in2->is_DecodeNarrowPtr()) {
3317 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3318 new_in2 = in2->in(1);
3319 } else if (in2->Opcode() == Op_ConP) {
3320 const Type* t = in2->bottom_type();
3321 if (t == TypePtr::NULL_PTR) {
3322 assert(in1->is_DecodeN(), "compare klass to null?");
3323 // Don't convert CmpP null check into CmpN if compressed
3324 // oops implicit null check is not generated.
3325 // This will allow to generate normal oop implicit null check.
3326 if (Matcher::gen_narrow_oop_implicit_null_checks())
3327 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3328 //
3329 // This transformation together with CastPP transformation above
3330 // will generated code for implicit NULL checks for compressed oops.
3331 //
3332 // The original code after Optimize()
3333 //
3334 // LoadN memory, narrow_oop_reg
3335 // decode narrow_oop_reg, base_reg
3336 // CmpP base_reg, NULL
3337 // CastPP base_reg // NotNull
3338 // Load [base_reg + offset], val_reg
3339 //
3340 // after these transformations will be
3341 //
3342 // LoadN memory, narrow_oop_reg
3343 // CmpN narrow_oop_reg, NULL
3344 // decode_not_null narrow_oop_reg, base_reg
3345 // Load [base_reg + offset], val_reg
3346 //
3347 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3348 // since narrow oops can be used in debug info now (see the code in
3349 // final_graph_reshaping_walk()).
3350 //
3351 // At the end the code will be matched to
3352 // on x86:
3353 //
3354 // Load_narrow_oop memory, narrow_oop_reg
3355 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3356 // NullCheck narrow_oop_reg
3357 //
3358 // and on sparc:
3359 //
3360 // Load_narrow_oop memory, narrow_oop_reg
3361 // decode_not_null narrow_oop_reg, base_reg
3362 // Load [base_reg + offset], val_reg
3363 // NullCheck base_reg
3364 //
3365 } else if (t->isa_oopptr()) {
3366 new_in2 = ConNode::make(t->make_narrowoop());
3367 } else if (t->isa_klassptr()) {
3368 new_in2 = ConNode::make(t->make_narrowklass());
3369 }
3370 }
3371 if (new_in2 != NULL) {
3372 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3373 n->subsume_by(cmpN, this);
3374 if (in1->outcnt() == 0) {
3375 in1->disconnect_inputs(this);
3376 }
3377 if (in2->outcnt() == 0) {
3378 in2->disconnect_inputs(this);
3379 }
3380 }
3381 }
3382 break;
3383
3384 case Op_DecodeN:
3385 case Op_DecodeNKlass:
3386 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3387 // DecodeN could be pinned when it can't be fold into
3388 // an address expression, see the code for Op_CastPP above.
3389 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3390 break;
3391
3392 case Op_EncodeP:
3393 case Op_EncodePKlass: {
3394 Node* in1 = n->in(1);
3395 if (in1->is_DecodeNarrowPtr()) {
3396 n->subsume_by(in1->in(1), this);
3397 } else if (in1->Opcode() == Op_ConP) {
3398 const Type* t = in1->bottom_type();
3399 if (t == TypePtr::NULL_PTR) {
3400 assert(t->isa_oopptr(), "null klass?");
3401 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3402 } else if (t->isa_oopptr()) {
3403 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3404 } else if (t->isa_klassptr()) {
3405 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3406 }
3407 }
3408 if (in1->outcnt() == 0) {
3409 in1->disconnect_inputs(this);
3410 }
3411 break;
3412 }
3413
3414 case Op_ReverseBytesV:
3415 case Op_ReverseV: {
3416 if ((uint)n->in(1)->Opcode() == nop) {
3417 if (n->is_predicated_vector() && n->in(1)->is_predicated_vector() &&
3418 n->in(2) == n->in(1)->in(2)) {
3419 // Node (Node X , Mask) Mask => X
3420 n->subsume_by(n->in(1)->in(1), this);
3421 } else if (!n->is_predicated_using_blend() && !n->in(1)->is_predicated_using_blend()) {
3422 // Node (Node X) => X
3423 n->subsume_by(n->in(1)->in(1), this);
3424 }
3425 }
3426 break;
3427 }
3428
3429 case Op_Proj: {
3430 if (OptimizeStringConcat || IncrementalInline) {
3431 ProjNode* proj = n->as_Proj();
3432 if (proj->_is_io_use) {
3433 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3434 // Separate projections were used for the exception path which
3435 // are normally removed by a late inline. If it wasn't inlined
3436 // then they will hang around and should just be replaced with
3437 // the original one. Merge them.
3438 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3439 if (non_io_proj != NULL) {
3440 proj->subsume_by(non_io_proj , this);
3441 }
3442 }
3443 }
3444 break;
3445 }
3446
3447 case Op_Phi:
3448 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3449 // The EncodeP optimization may create Phi with the same edges
3450 // for all paths. It is not handled well by Register Allocator.
3451 Node* unique_in = n->in(1);
3452 assert(unique_in != NULL, "");
3453 uint cnt = n->req();
3454 for (uint i = 2; i < cnt; i++) {
3455 Node* m = n->in(i);
3456 assert(m != NULL, "");
3457 if (unique_in != m)
3458 unique_in = NULL;
3459 }
3460 if (unique_in != NULL) {
3461 n->subsume_by(unique_in, this);
3462 }
3463 }
3464 break;
3465
3466 #endif
3467
3468 #ifdef ASSERT
3469 case Op_CastII:
3470 // Verify that all range check dependent CastII nodes were removed.
3471 if (n->isa_CastII()->has_range_check()) {
3472 n->dump(3);
3473 assert(false, "Range check dependent CastII node was not removed");
3474 }
3475 break;
3476 #endif
3477
3478 case Op_ModI:
3479 if (UseDivMod) {
3480 // Check if a%b and a/b both exist
3481 Node* d = n->find_similar(Op_DivI);
3482 if (d) {
3483 // Replace them with a fused divmod if supported
3484 if (Matcher::has_match_rule(Op_DivModI)) {
3485 DivModINode* divmod = DivModINode::make(n);
3486 d->subsume_by(divmod->div_proj(), this);
3487 n->subsume_by(divmod->mod_proj(), this);
3488 } else {
3489 // replace a%b with a-((a/b)*b)
3490 Node* mult = new MulINode(d, d->in(2));
3491 Node* sub = new SubINode(d->in(1), mult);
3492 n->subsume_by(sub, this);
3493 }
3494 }
3495 }
3496 break;
3497
3498 case Op_ModL:
3499 if (UseDivMod) {
3500 // Check if a%b and a/b both exist
3501 Node* d = n->find_similar(Op_DivL);
3502 if (d) {
3503 // Replace them with a fused divmod if supported
3504 if (Matcher::has_match_rule(Op_DivModL)) {
3505 DivModLNode* divmod = DivModLNode::make(n);
3506 d->subsume_by(divmod->div_proj(), this);
3507 n->subsume_by(divmod->mod_proj(), this);
3508 } else {
3509 // replace a%b with a-((a/b)*b)
3510 Node* mult = new MulLNode(d, d->in(2));
3511 Node* sub = new SubLNode(d->in(1), mult);
3512 n->subsume_by(sub, this);
3513 }
3514 }
3515 }
3516 break;
3517
3518 case Op_UModI:
3519 if (UseDivMod) {
3520 // Check if a%b and a/b both exist
3521 Node* d = n->find_similar(Op_UDivI);
3522 if (d) {
3523 // Replace them with a fused unsigned divmod if supported
3524 if (Matcher::has_match_rule(Op_UDivModI)) {
3525 UDivModINode* divmod = UDivModINode::make(n);
3526 d->subsume_by(divmod->div_proj(), this);
3527 n->subsume_by(divmod->mod_proj(), this);
3528 } else {
3529 // replace a%b with a-((a/b)*b)
3530 Node* mult = new MulINode(d, d->in(2));
3531 Node* sub = new SubINode(d->in(1), mult);
3532 n->subsume_by(sub, this);
3533 }
3534 }
3535 }
3536 break;
3537
3538 case Op_UModL:
3539 if (UseDivMod) {
3540 // Check if a%b and a/b both exist
3541 Node* d = n->find_similar(Op_UDivL);
3542 if (d) {
3543 // Replace them with a fused unsigned divmod if supported
3544 if (Matcher::has_match_rule(Op_UDivModL)) {
3545 UDivModLNode* divmod = UDivModLNode::make(n);
3546 d->subsume_by(divmod->div_proj(), this);
3547 n->subsume_by(divmod->mod_proj(), this);
3548 } else {
3549 // replace a%b with a-((a/b)*b)
3550 Node* mult = new MulLNode(d, d->in(2));
3551 Node* sub = new SubLNode(d->in(1), mult);
3552 n->subsume_by(sub, this);
3553 }
3554 }
3555 }
3556 break;
3557
3558 case Op_NoOvfModI:
3559 if (UseDivMod) {
3560 // Check if a%b and a/b both exist
3561 Node* d = n->find_similar(Op_NoOvfDivI);
3562 if (d) {
3563 // Replace them with a fused divmod if supported
3564 if (Matcher::has_match_rule(Op_NoOvfDivModI)) {
3565 NoOvfDivModINode* divmod = NoOvfDivModINode::make(n);
3566 d->subsume_by(divmod->div_proj(), this);
3567 n->subsume_by(divmod->mod_proj(), this);
3568 } else {
3569 // replace a%b with a-((a/b)*b)
3570 Node* mult = new MulINode(d, d->in(2));
3571 Node* sub = new SubINode(d->in(1), mult);
3572 n->subsume_by(sub, this);
3573 }
3574 }
3575 }
3576 break;
3577
3578 case Op_NoOvfModL:
3579 if (UseDivMod) {
3580 // Check if a%b and a/b both exist
3581 Node* d = n->find_similar(Op_NoOvfDivL);
3582 if (d) {
3583 // Replace them with a fused divmod if supported
3584 if (Matcher::has_match_rule(Op_NoOvfDivModL)) {
3585 NoOvfDivModLNode* divmod = NoOvfDivModLNode::make(n);
3586 d->subsume_by(divmod->div_proj(), this);
3587 n->subsume_by(divmod->mod_proj(), this);
3588 } else {
3589 // replace a%b with a-((a/b)*b)
3590 Node* mult = new MulLNode(d, d->in(2));
3591 Node* sub = new SubLNode(d->in(1), mult);
3592 n->subsume_by(sub, this);
3593 }
3594 }
3595 }
3596 break;
3597
3598 case Op_LoadVector:
3599 case Op_StoreVector:
3600 case Op_LoadVectorGather:
3601 case Op_StoreVectorScatter:
3602 case Op_LoadVectorGatherMasked:
3603 case Op_StoreVectorScatterMasked:
3604 case Op_VectorCmpMasked:
3605 case Op_VectorMaskGen:
3606 case Op_LoadVectorMasked:
3607 case Op_StoreVectorMasked:
3608 break;
3609
3610 case Op_AddReductionVI:
3611 case Op_AddReductionVL:
3612 case Op_AddReductionVF:
3613 case Op_AddReductionVD:
3614 case Op_MulReductionVI:
3615 case Op_MulReductionVL:
3616 case Op_MulReductionVF:
3617 case Op_MulReductionVD:
3618 case Op_MinReductionV:
3619 case Op_MaxReductionV:
3620 case Op_AndReductionV:
3621 case Op_OrReductionV:
3622 case Op_XorReductionV:
3623 break;
3624
3625 case Op_PackB:
3626 case Op_PackS:
3627 case Op_PackI:
3628 case Op_PackF:
3629 case Op_PackL:
3630 case Op_PackD:
3631 if (n->req()-1 > 2) {
3632 // Replace many operand PackNodes with a binary tree for matching
3633 PackNode* p = (PackNode*) n;
3634 Node* btp = p->binary_tree_pack(1, n->req());
3635 n->subsume_by(btp, this);
3636 }
3637 break;
3638 case Op_Loop:
3639 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3640 case Op_CountedLoop:
3641 case Op_LongCountedLoop:
3642 case Op_OuterStripMinedLoop:
3643 if (n->as_Loop()->is_inner_loop()) {
3644 frc.inc_inner_loop_count();
3645 }
3646 n->as_Loop()->verify_strip_mined(0);
3647 break;
3648 case Op_LShiftI:
3649 case Op_RShiftI:
3650 case Op_URShiftI:
3651 case Op_LShiftL:
3652 case Op_RShiftL:
3653 case Op_URShiftL:
3654 if (Matcher::need_masked_shift_count) {
3655 // The cpu's shift instructions don't restrict the count to the
3656 // lower 5/6 bits. We need to do the masking ourselves.
3657 Node* in2 = n->in(2);
3658 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3659 const TypeInt* t = in2->find_int_type();
3660 if (t != NULL && t->is_con()) {
3661 juint shift = t->get_con();
3662 if (shift > mask) { // Unsigned cmp
3663 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3664 }
3665 } else {
3666 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3667 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3668 n->set_req(2, shift);
3669 }
3670 }
3671 if (in2->outcnt() == 0) { // Remove dead node
3672 in2->disconnect_inputs(this);
3673 }
3674 }
3675 break;
3676 case Op_MemBarStoreStore:
3677 case Op_MemBarRelease:
3678 // Break the link with AllocateNode: it is no longer useful and
3679 // confuses register allocation.
3680 if (n->req() > MemBarNode::Precedent) {
3681 n->set_req(MemBarNode::Precedent, top());
3682 }
3683 break;
3684 case Op_MemBarAcquire: {
3685 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3686 // At parse time, the trailing MemBarAcquire for a volatile load
3687 // is created with an edge to the load. After optimizations,
3688 // that input may be a chain of Phis. If those phis have no
3689 // other use, then the MemBarAcquire keeps them alive and
3690 // register allocation can be confused.
3691 ResourceMark rm;
3692 Unique_Node_List wq;
3693 wq.push(n->in(MemBarNode::Precedent));
3694 n->set_req(MemBarNode::Precedent, top());
3695 while (wq.size() > 0) {
3696 Node* m = wq.pop();
3697 if (m->outcnt() == 0) {
3698 for (uint j = 0; j < m->req(); j++) {
3699 Node* in = m->in(j);
3700 if (in != NULL) {
3701 wq.push(in);
3702 }
3703 }
3704 m->disconnect_inputs(this);
3705 }
3706 }
3707 }
3708 break;
3709 }
3710 case Op_Blackhole:
3711 break;
3712 case Op_RangeCheck: {
3713 RangeCheckNode* rc = n->as_RangeCheck();
3714 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3715 n->subsume_by(iff, this);
3716 frc._tests.push(iff);
3717 break;
3718 }
3719 case Op_ConvI2L: {
3720 if (!Matcher::convi2l_type_required) {
3721 // Code generation on some platforms doesn't need accurate
3722 // ConvI2L types. Widening the type can help remove redundant
3723 // address computations.
3724 n->as_Type()->set_type(TypeLong::INT);
3725 ResourceMark rm;
3726 Unique_Node_List wq;
3727 wq.push(n);
3728 for (uint next = 0; next < wq.size(); next++) {
3729 Node *m = wq.at(next);
3730
3731 for(;;) {
3732 // Loop over all nodes with identical inputs edges as m
3733 Node* k = m->find_similar(m->Opcode());
3734 if (k == NULL) {
3735 break;
3736 }
3737 // Push their uses so we get a chance to remove node made
3738 // redundant
3739 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3740 Node* u = k->fast_out(i);
3741 if (u->Opcode() == Op_LShiftL ||
3742 u->Opcode() == Op_AddL ||
3743 u->Opcode() == Op_SubL ||
3744 u->Opcode() == Op_AddP) {
3745 wq.push(u);
3746 }
3747 }
3748 // Replace all nodes with identical edges as m with m
3749 k->subsume_by(m, this);
3750 }
3751 }
3752 }
3753 break;
3754 }
3755 case Op_CmpUL: {
3756 if (!Matcher::has_match_rule(Op_CmpUL)) {
3757 // No support for unsigned long comparisons
3758 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3759 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3760 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3761 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3762 Node* andl = new AndLNode(orl, remove_sign_mask);
3763 Node* cmp = new CmpLNode(andl, n->in(2));
3764 n->subsume_by(cmp, this);
3765 }
3766 break;
3767 }
3768 default:
3769 assert(!n->is_Call(), "");
3770 assert(!n->is_Mem(), "");
3771 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3772 break;
3773 }
3774 }
3775
3776 //------------------------------final_graph_reshaping_walk---------------------
3777 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3778 // requires that the walk visits a node's inputs before visiting the node.
3779 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3780 Unique_Node_List sfpt;
3781
3782 frc._visited.set(root->_idx); // first, mark node as visited
3783 uint cnt = root->req();
3784 Node *n = root;
3785 uint i = 0;
3786 while (true) {
3787 if (i < cnt) {
3788 // Place all non-visited non-null inputs onto stack
3789 Node* m = n->in(i);
3790 ++i;
3791 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3792 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3793 // compute worst case interpreter size in case of a deoptimization
3794 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3795
3796 sfpt.push(m);
3797 }
3798 cnt = m->req();
3799 nstack.push(n, i); // put on stack parent and next input's index
3800 n = m;
3801 i = 0;
3802 }
3803 } else {
3804 // Now do post-visit work
3805 final_graph_reshaping_impl( n, frc );
3806 if (nstack.is_empty())
3807 break; // finished
3808 n = nstack.node(); // Get node from stack
3809 cnt = n->req();
3810 i = nstack.index();
3811 nstack.pop(); // Shift to the next node on stack
3812 }
3813 }
3814
3815 // Skip next transformation if compressed oops are not used.
3816 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3817 (!UseCompressedOops && !UseCompressedClassPointers))
3818 return;
3819
3820 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3821 // It could be done for an uncommon traps or any safepoints/calls
3822 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3823 while (sfpt.size() > 0) {
3824 n = sfpt.pop();
3825 JVMState *jvms = n->as_SafePoint()->jvms();
3826 assert(jvms != NULL, "sanity");
3827 int start = jvms->debug_start();
3828 int end = n->req();
3829 bool is_uncommon = (n->is_CallStaticJava() &&
3830 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3831 for (int j = start; j < end; j++) {
3832 Node* in = n->in(j);
3833 if (in->is_DecodeNarrowPtr()) {
3834 bool safe_to_skip = true;
3835 if (!is_uncommon ) {
3836 // Is it safe to skip?
3837 for (uint i = 0; i < in->outcnt(); i++) {
3838 Node* u = in->raw_out(i);
3839 if (!u->is_SafePoint() ||
3840 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3841 safe_to_skip = false;
3842 }
3843 }
3844 }
3845 if (safe_to_skip) {
3846 n->set_req(j, in->in(1));
3847 }
3848 if (in->outcnt() == 0) {
3849 in->disconnect_inputs(this);
3850 }
3851 }
3852 }
3853 }
3854 }
3855
3856 //------------------------------final_graph_reshaping--------------------------
3857 // Final Graph Reshaping.
3858 //
3859 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3860 // and not commoned up and forced early. Must come after regular
3861 // optimizations to avoid GVN undoing the cloning. Clone constant
3862 // inputs to Loop Phis; these will be split by the allocator anyways.
3863 // Remove Opaque nodes.
3864 // (2) Move last-uses by commutative operations to the left input to encourage
3865 // Intel update-in-place two-address operations and better register usage
3866 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3867 // calls canonicalizing them back.
3868 // (3) Count the number of double-precision FP ops, single-precision FP ops
3869 // and call sites. On Intel, we can get correct rounding either by
3870 // forcing singles to memory (requires extra stores and loads after each
3871 // FP bytecode) or we can set a rounding mode bit (requires setting and
3872 // clearing the mode bit around call sites). The mode bit is only used
3873 // if the relative frequency of single FP ops to calls is low enough.
3874 // This is a key transform for SPEC mpeg_audio.
3875 // (4) Detect infinite loops; blobs of code reachable from above but not
3876 // below. Several of the Code_Gen algorithms fail on such code shapes,
3877 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3878 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3879 // Detection is by looking for IfNodes where only 1 projection is
3880 // reachable from below or CatchNodes missing some targets.
3881 // (5) Assert for insane oop offsets in debug mode.
3882
3883 bool Compile::final_graph_reshaping() {
3884 // an infinite loop may have been eliminated by the optimizer,
3885 // in which case the graph will be empty.
3886 if (root()->req() == 1) {
3887 record_method_not_compilable("trivial infinite loop");
3888 return true;
3889 }
3890
3891 // Expensive nodes have their control input set to prevent the GVN
3892 // from freely commoning them. There's no GVN beyond this point so
3893 // no need to keep the control input. We want the expensive nodes to
3894 // be freely moved to the least frequent code path by gcm.
3895 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3896 for (int i = 0; i < expensive_count(); i++) {
3897 _expensive_nodes.at(i)->set_req(0, NULL);
3898 }
3899
3900 Final_Reshape_Counts frc;
3901
3902 // Visit everybody reachable!
3903 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3904 Node_Stack nstack(live_nodes() >> 1);
3905 final_graph_reshaping_walk(nstack, root(), frc);
3906
3907 // Check for unreachable (from below) code (i.e., infinite loops).
3908 for( uint i = 0; i < frc._tests.size(); i++ ) {
3909 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3910 // Get number of CFG targets.
3911 // Note that PCTables include exception targets after calls.
3912 uint required_outcnt = n->required_outcnt();
3913 if (n->outcnt() != required_outcnt) {
3914 // Check for a few special cases. Rethrow Nodes never take the
3915 // 'fall-thru' path, so expected kids is 1 less.
3916 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3917 if (n->in(0)->in(0)->is_Call()) {
3918 CallNode* call = n->in(0)->in(0)->as_Call();
3919 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3920 required_outcnt--; // Rethrow always has 1 less kid
3921 } else if (call->req() > TypeFunc::Parms &&
3922 call->is_CallDynamicJava()) {
3923 // Check for null receiver. In such case, the optimizer has
3924 // detected that the virtual call will always result in a null
3925 // pointer exception. The fall-through projection of this CatchNode
3926 // will not be populated.
3927 Node* arg0 = call->in(TypeFunc::Parms);
3928 if (arg0->is_Type() &&
3929 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3930 required_outcnt--;
3931 }
3932 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3933 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3934 // Check for illegal array length. In such case, the optimizer has
3935 // detected that the allocation attempt will always result in an
3936 // exception. There is no fall-through projection of this CatchNode .
3937 assert(call->is_CallStaticJava(), "static call expected");
3938 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3939 Node* valid_length_test = call->in(call->req()-1);
3940 call->del_req(call->req()-1);
3941 if (valid_length_test->find_int_con(1) == 0) {
3942 required_outcnt--;
3943 }
3944 assert(n->outcnt() == required_outcnt, "malformed control flow");
3945 continue;
3946 }
3947 }
3948 }
3949 // Recheck with a better notion of 'required_outcnt'
3950 if (n->outcnt() != required_outcnt) {
3951 record_method_not_compilable("malformed control flow");
3952 return true; // Not all targets reachable!
3953 }
3954 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
3955 CallNode* call = n->in(0)->in(0)->as_Call();
3956 if (call->entry_point() == OptoRuntime::new_array_Java() ||
3957 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3958 assert(call->is_CallStaticJava(), "static call expected");
3959 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3960 call->del_req(call->req()-1); // valid length test useless now
3961 }
3962 }
3963 // Check that I actually visited all kids. Unreached kids
3964 // must be infinite loops.
3965 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3966 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3967 record_method_not_compilable("infinite loop");
3968 return true; // Found unvisited kid; must be unreach
3969 }
3970
3971 // Here so verification code in final_graph_reshaping_walk()
3972 // always see an OuterStripMinedLoopEnd
3973 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
3974 IfNode* init_iff = n->as_If();
3975 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3976 n->subsume_by(iff, this);
3977 }
3978 }
3979
3980 #ifdef IA32
3981 // If original bytecodes contained a mixture of floats and doubles
3982 // check if the optimizer has made it homogeneous, item (3).
3983 if (UseSSE == 0 &&
3984 frc.get_float_count() > 32 &&
3985 frc.get_double_count() == 0 &&
3986 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3987 set_24_bit_selection_and_mode(false, true);
3988 }
3989 #endif // IA32
3990
3991 set_java_calls(frc.get_java_call_count());
3992 set_inner_loops(frc.get_inner_loop_count());
3993
3994 // No infinite loops, no reason to bail out.
3995 return false;
3996 }
3997
3998 //-----------------------------too_many_traps----------------------------------
3999 // Report if there are too many traps at the current method and bci.
4000 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4001 bool Compile::too_many_traps(ciMethod* method,
4002 int bci,
4003 Deoptimization::DeoptReason reason) {
4004 ciMethodData* md = method->method_data();
4005 if (md->is_empty()) {
4006 // Assume the trap has not occurred, or that it occurred only
4007 // because of a transient condition during start-up in the interpreter.
4008 return false;
4009 }
4010 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4011 if (md->has_trap_at(bci, m, reason) != 0) {
4012 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4013 // Also, if there are multiple reasons, or if there is no per-BCI record,
4014 // assume the worst.
4015 if (log())
4016 log()->elem("observe trap='%s' count='%d'",
4017 Deoptimization::trap_reason_name(reason),
4018 md->trap_count(reason));
4019 return true;
4020 } else {
4021 // Ignore method/bci and see if there have been too many globally.
4022 return too_many_traps(reason, md);
4023 }
4024 }
4025
4026 // Less-accurate variant which does not require a method and bci.
4027 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4028 ciMethodData* logmd) {
4029 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4030 // Too many traps globally.
4031 // Note that we use cumulative trap_count, not just md->trap_count.
4032 if (log()) {
4033 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4034 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4035 Deoptimization::trap_reason_name(reason),
4036 mcount, trap_count(reason));
4037 }
4038 return true;
4039 } else {
4040 // The coast is clear.
4041 return false;
4042 }
4043 }
4044
4045 //--------------------------too_many_recompiles--------------------------------
4046 // Report if there are too many recompiles at the current method and bci.
4047 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4048 // Is not eager to return true, since this will cause the compiler to use
4049 // Action_none for a trap point, to avoid too many recompilations.
4050 bool Compile::too_many_recompiles(ciMethod* method,
4051 int bci,
4052 Deoptimization::DeoptReason reason) {
4053 ciMethodData* md = method->method_data();
4054 if (md->is_empty()) {
4055 // Assume the trap has not occurred, or that it occurred only
4056 // because of a transient condition during start-up in the interpreter.
4057 return false;
4058 }
4059 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4060 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4061 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4062 Deoptimization::DeoptReason per_bc_reason
4063 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4064 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4065 if ((per_bc_reason == Deoptimization::Reason_none
4066 || md->has_trap_at(bci, m, reason) != 0)
4067 // The trap frequency measure we care about is the recompile count:
4068 && md->trap_recompiled_at(bci, m)
4069 && md->overflow_recompile_count() >= bc_cutoff) {
4070 // Do not emit a trap here if it has already caused recompilations.
4071 // Also, if there are multiple reasons, or if there is no per-BCI record,
4072 // assume the worst.
4073 if (log())
4074 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4075 Deoptimization::trap_reason_name(reason),
4076 md->trap_count(reason),
4077 md->overflow_recompile_count());
4078 return true;
4079 } else if (trap_count(reason) != 0
4080 && decompile_count() >= m_cutoff) {
4081 // Too many recompiles globally, and we have seen this sort of trap.
4082 // Use cumulative decompile_count, not just md->decompile_count.
4083 if (log())
4084 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4085 Deoptimization::trap_reason_name(reason),
4086 md->trap_count(reason), trap_count(reason),
4087 md->decompile_count(), decompile_count());
4088 return true;
4089 } else {
4090 // The coast is clear.
4091 return false;
4092 }
4093 }
4094
4095 // Compute when not to trap. Used by matching trap based nodes and
4096 // NullCheck optimization.
4097 void Compile::set_allowed_deopt_reasons() {
4098 _allowed_reasons = 0;
4099 if (is_method_compilation()) {
4100 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4101 assert(rs < BitsPerInt, "recode bit map");
4102 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4103 _allowed_reasons |= nth_bit(rs);
4104 }
4105 }
4106 }
4107 }
4108
4109 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4110 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4111 }
4112
4113 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4114 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4115 }
4116
4117 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4118 if (holder->is_initialized()) {
4119 return false;
4120 }
4121 if (holder->is_being_initialized()) {
4122 if (accessing_method->holder() == holder) {
4123 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4124 // <init>, or a static method. In all those cases, there was an initialization
4125 // barrier on the holder klass passed.
4126 if (accessing_method->is_static_initializer() ||
4127 accessing_method->is_object_initializer() ||
4128 accessing_method->is_static()) {
4129 return false;
4130 }
4131 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4132 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4133 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4134 // child class can become fully initialized while its parent class is still being initialized.
4135 if (accessing_method->is_static_initializer()) {
4136 return false;
4137 }
4138 }
4139 ciMethod* root = method(); // the root method of compilation
4140 if (root != accessing_method) {
4141 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4142 }
4143 }
4144 return true;
4145 }
4146
4147 #ifndef PRODUCT
4148 //------------------------------verify_graph_edges---------------------------
4149 // Walk the Graph and verify that there is a one-to-one correspondence
4150 // between Use-Def edges and Def-Use edges in the graph.
4151 void Compile::verify_graph_edges(bool no_dead_code) {
4152 if (VerifyGraphEdges) {
4153 Unique_Node_List visited;
4154 // Call recursive graph walk to check edges
4155 _root->verify_edges(visited);
4156 if (no_dead_code) {
4157 // Now make sure that no visited node is used by an unvisited node.
4158 bool dead_nodes = false;
4159 Unique_Node_List checked;
4160 while (visited.size() > 0) {
4161 Node* n = visited.pop();
4162 checked.push(n);
4163 for (uint i = 0; i < n->outcnt(); i++) {
4164 Node* use = n->raw_out(i);
4165 if (checked.member(use)) continue; // already checked
4166 if (visited.member(use)) continue; // already in the graph
4167 if (use->is_Con()) continue; // a dead ConNode is OK
4168 // At this point, we have found a dead node which is DU-reachable.
4169 if (!dead_nodes) {
4170 tty->print_cr("*** Dead nodes reachable via DU edges:");
4171 dead_nodes = true;
4172 }
4173 use->dump(2);
4174 tty->print_cr("---");
4175 checked.push(use); // No repeats; pretend it is now checked.
4176 }
4177 }
4178 assert(!dead_nodes, "using nodes must be reachable from root");
4179 }
4180 }
4181 }
4182 #endif
4183
4184 // The Compile object keeps track of failure reasons separately from the ciEnv.
4185 // This is required because there is not quite a 1-1 relation between the
4186 // ciEnv and its compilation task and the Compile object. Note that one
4187 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4188 // to backtrack and retry without subsuming loads. Other than this backtracking
4189 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4190 // by the logic in C2Compiler.
4191 void Compile::record_failure(const char* reason) {
4192 if (log() != NULL) {
4193 log()->elem("failure reason='%s' phase='compile'", reason);
4194 }
4195 if (_failure_reason == NULL) {
4196 // Record the first failure reason.
4197 _failure_reason = reason;
4198 }
4199
4200 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4201 C->print_method(PHASE_FAILURE, 1);
4202 }
4203 _root = NULL; // flush the graph, too
4204 }
4205
4206 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4207 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4208 _phase_name(name), _dolog(CITimeVerbose)
4209 {
4210 if (_dolog) {
4211 C = Compile::current();
4212 _log = C->log();
4213 } else {
4214 C = NULL;
4215 _log = NULL;
4216 }
4217 if (_log != NULL) {
4218 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4219 _log->stamp();
4220 _log->end_head();
4221 }
4222 }
4223
4224 Compile::TracePhase::~TracePhase() {
4225
4226 C = Compile::current();
4227 if (_dolog) {
4228 _log = C->log();
4229 } else {
4230 _log = NULL;
4231 }
4232
4233 #ifdef ASSERT
4234 if (PrintIdealNodeCount) {
4235 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4236 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4237 }
4238
4239 if (VerifyIdealNodeCount) {
4240 Compile::current()->print_missing_nodes();
4241 }
4242 #endif
4243
4244 if (_log != NULL) {
4245 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4246 }
4247 }
4248
4249 //----------------------------static_subtype_check-----------------------------
4250 // Shortcut important common cases when superklass is exact:
4251 // (0) superklass is java.lang.Object (can occur in reflective code)
4252 // (1) subklass is already limited to a subtype of superklass => always ok
4253 // (2) subklass does not overlap with superklass => always fail
4254 // (3) superklass has NO subtypes and we can check with a simple compare.
4255 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4256 if (StressReflectiveCode) {
4257 return SSC_full_test; // Let caller generate the general case.
4258 }
4259
4260 if (superk == env()->Object_klass()) {
4261 return SSC_always_true; // (0) this test cannot fail
4262 }
4263
4264 ciType* superelem = superk;
4265 ciType* subelem = subk;
4266 if (superelem->is_array_klass()) {
4267 superelem = superelem->as_array_klass()->base_element_type();
4268 }
4269 if (subelem->is_array_klass()) {
4270 subelem = subelem->as_array_klass()->base_element_type();
4271 }
4272
4273 if (!subk->is_interface()) { // cannot trust static interface types yet
4274 if (subk->is_subtype_of(superk)) {
4275 return SSC_always_true; // (1) false path dead; no dynamic test needed
4276 }
4277 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4278 !(subelem->is_klass() && subelem->as_klass()->is_interface()) &&
4279 !superk->is_subtype_of(subk)) {
4280 return SSC_always_false; // (2) true path dead; no dynamic test needed
4281 }
4282 }
4283
4284 // If casting to an instance klass, it must have no subtypes
4285 if (superk->is_interface()) {
4286 // Cannot trust interfaces yet.
4287 // %%% S.B. superk->nof_implementors() == 1
4288 } else if (superelem->is_instance_klass()) {
4289 ciInstanceKlass* ik = superelem->as_instance_klass();
4290 if (!ik->has_subklass() && !ik->is_interface()) {
4291 if (!ik->is_final()) {
4292 // Add a dependency if there is a chance of a later subclass.
4293 dependencies()->assert_leaf_type(ik);
4294 }
4295 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4296 }
4297 } else {
4298 // A primitive array type has no subtypes.
4299 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4300 }
4301
4302 return SSC_full_test;
4303 }
4304
4305 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4306 #ifdef _LP64
4307 // The scaled index operand to AddP must be a clean 64-bit value.
4308 // Java allows a 32-bit int to be incremented to a negative
4309 // value, which appears in a 64-bit register as a large
4310 // positive number. Using that large positive number as an
4311 // operand in pointer arithmetic has bad consequences.
4312 // On the other hand, 32-bit overflow is rare, and the possibility
4313 // can often be excluded, if we annotate the ConvI2L node with
4314 // a type assertion that its value is known to be a small positive
4315 // number. (The prior range check has ensured this.)
4316 // This assertion is used by ConvI2LNode::Ideal.
4317 int index_max = max_jint - 1; // array size is max_jint, index is one less
4318 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4319 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4320 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4321 #endif
4322 return idx;
4323 }
4324
4325 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4326 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4327 if (ctrl != NULL) {
4328 // Express control dependency by a CastII node with a narrow type.
4329 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4330 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4331 // node from floating above the range check during loop optimizations. Otherwise, the
4332 // ConvI2L node may be eliminated independently of the range check, causing the data path
4333 // to become TOP while the control path is still there (although it's unreachable).
4334 value->set_req(0, ctrl);
4335 value = phase->transform(value);
4336 }
4337 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4338 return phase->transform(new ConvI2LNode(value, ltype));
4339 }
4340
4341 void Compile::print_inlining_stream_free() {
4342 if (_print_inlining_stream != NULL) {
4343 _print_inlining_stream->~stringStream();
4344 _print_inlining_stream = NULL;
4345 }
4346 }
4347
4348 // The message about the current inlining is accumulated in
4349 // _print_inlining_stream and transferred into the _print_inlining_list
4350 // once we know whether inlining succeeds or not. For regular
4351 // inlining, messages are appended to the buffer pointed by
4352 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4353 // a new buffer is added after _print_inlining_idx in the list. This
4354 // way we can update the inlining message for late inlining call site
4355 // when the inlining is attempted again.
4356 void Compile::print_inlining_init() {
4357 if (print_inlining() || print_intrinsics()) {
4358 // print_inlining_init is actually called several times.
4359 print_inlining_stream_free();
4360 _print_inlining_stream = new stringStream();
4361 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4362 }
4363 }
4364
4365 void Compile::print_inlining_reinit() {
4366 if (print_inlining() || print_intrinsics()) {
4367 print_inlining_stream_free();
4368 // Re allocate buffer when we change ResourceMark
4369 _print_inlining_stream = new stringStream();
4370 }
4371 }
4372
4373 void Compile::print_inlining_reset() {
4374 _print_inlining_stream->reset();
4375 }
4376
4377 void Compile::print_inlining_commit() {
4378 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4379 // Transfer the message from _print_inlining_stream to the current
4380 // _print_inlining_list buffer and clear _print_inlining_stream.
4381 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4382 print_inlining_reset();
4383 }
4384
4385 void Compile::print_inlining_push() {
4386 // Add new buffer to the _print_inlining_list at current position
4387 _print_inlining_idx++;
4388 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4389 }
4390
4391 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4392 return _print_inlining_list->at(_print_inlining_idx);
4393 }
4394
4395 void Compile::print_inlining_update(CallGenerator* cg) {
4396 if (print_inlining() || print_intrinsics()) {
4397 if (cg->is_late_inline()) {
4398 if (print_inlining_current()->cg() != cg &&
4399 (print_inlining_current()->cg() != NULL ||
4400 print_inlining_current()->ss()->size() != 0)) {
4401 print_inlining_push();
4402 }
4403 print_inlining_commit();
4404 print_inlining_current()->set_cg(cg);
4405 } else {
4406 if (print_inlining_current()->cg() != NULL) {
4407 print_inlining_push();
4408 }
4409 print_inlining_commit();
4410 }
4411 }
4412 }
4413
4414 void Compile::print_inlining_move_to(CallGenerator* cg) {
4415 // We resume inlining at a late inlining call site. Locate the
4416 // corresponding inlining buffer so that we can update it.
4417 if (print_inlining() || print_intrinsics()) {
4418 for (int i = 0; i < _print_inlining_list->length(); i++) {
4419 if (_print_inlining_list->at(i)->cg() == cg) {
4420 _print_inlining_idx = i;
4421 return;
4422 }
4423 }
4424 ShouldNotReachHere();
4425 }
4426 }
4427
4428 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4429 if (print_inlining() || print_intrinsics()) {
4430 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4431 assert(print_inlining_current()->cg() == cg, "wrong entry");
4432 // replace message with new message
4433 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4434 print_inlining_commit();
4435 print_inlining_current()->set_cg(cg);
4436 }
4437 }
4438
4439 void Compile::print_inlining_assert_ready() {
4440 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4441 }
4442
4443 void Compile::process_print_inlining() {
4444 assert(_late_inlines.length() == 0, "not drained yet");
4445 if (print_inlining() || print_intrinsics()) {
4446 ResourceMark rm;
4447 stringStream ss;
4448 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4449 for (int i = 0; i < _print_inlining_list->length(); i++) {
4450 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4451 ss.print("%s", pib->ss()->as_string());
4452 delete pib;
4453 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
4454 }
4455 // Reset _print_inlining_list, it only contains destructed objects.
4456 // It is on the arena, so it will be freed when the arena is reset.
4457 _print_inlining_list = NULL;
4458 // _print_inlining_stream won't be used anymore, either.
4459 print_inlining_stream_free();
4460 size_t end = ss.size();
4461 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4462 strncpy(_print_inlining_output, ss.base(), end+1);
4463 _print_inlining_output[end] = 0;
4464 }
4465 }
4466
4467 void Compile::dump_print_inlining() {
4468 if (_print_inlining_output != NULL) {
4469 tty->print_raw(_print_inlining_output);
4470 }
4471 }
4472
4473 void Compile::log_late_inline(CallGenerator* cg) {
4474 if (log() != NULL) {
4475 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4476 cg->unique_id());
4477 JVMState* p = cg->call_node()->jvms();
4478 while (p != NULL) {
4479 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4480 p = p->caller();
4481 }
4482 log()->tail("late_inline");
4483 }
4484 }
4485
4486 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4487 log_late_inline(cg);
4488 if (log() != NULL) {
4489 log()->inline_fail(msg);
4490 }
4491 }
4492
4493 void Compile::log_inline_id(CallGenerator* cg) {
4494 if (log() != NULL) {
4495 // The LogCompilation tool needs a unique way to identify late
4496 // inline call sites. This id must be unique for this call site in
4497 // this compilation. Try to have it unique across compilations as
4498 // well because it can be convenient when grepping through the log
4499 // file.
4500 // Distinguish OSR compilations from others in case CICountOSR is
4501 // on.
4502 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4503 cg->set_unique_id(id);
4504 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4505 }
4506 }
4507
4508 void Compile::log_inline_failure(const char* msg) {
4509 if (C->log() != NULL) {
4510 C->log()->inline_fail(msg);
4511 }
4512 }
4513
4514
4515 // Dump inlining replay data to the stream.
4516 // Don't change thread state and acquire any locks.
4517 void Compile::dump_inline_data(outputStream* out) {
4518 InlineTree* inl_tree = ilt();
4519 if (inl_tree != NULL) {
4520 out->print(" inline %d", inl_tree->count());
4521 inl_tree->dump_replay_data(out);
4522 }
4523 }
4524
4525 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4526 if (n1->Opcode() < n2->Opcode()) return -1;
4527 else if (n1->Opcode() > n2->Opcode()) return 1;
4528
4529 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4530 for (uint i = 1; i < n1->req(); i++) {
4531 if (n1->in(i) < n2->in(i)) return -1;
4532 else if (n1->in(i) > n2->in(i)) return 1;
4533 }
4534
4535 return 0;
4536 }
4537
4538 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4539 Node* n1 = *n1p;
4540 Node* n2 = *n2p;
4541
4542 return cmp_expensive_nodes(n1, n2);
4543 }
4544
4545 void Compile::sort_expensive_nodes() {
4546 if (!expensive_nodes_sorted()) {
4547 _expensive_nodes.sort(cmp_expensive_nodes);
4548 }
4549 }
4550
4551 bool Compile::expensive_nodes_sorted() const {
4552 for (int i = 1; i < _expensive_nodes.length(); i++) {
4553 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4554 return false;
4555 }
4556 }
4557 return true;
4558 }
4559
4560 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4561 if (_expensive_nodes.length() == 0) {
4562 return false;
4563 }
4564
4565 assert(OptimizeExpensiveOps, "optimization off?");
4566
4567 // Take this opportunity to remove dead nodes from the list
4568 int j = 0;
4569 for (int i = 0; i < _expensive_nodes.length(); i++) {
4570 Node* n = _expensive_nodes.at(i);
4571 if (!n->is_unreachable(igvn)) {
4572 assert(n->is_expensive(), "should be expensive");
4573 _expensive_nodes.at_put(j, n);
4574 j++;
4575 }
4576 }
4577 _expensive_nodes.trunc_to(j);
4578
4579 // Then sort the list so that similar nodes are next to each other
4580 // and check for at least two nodes of identical kind with same data
4581 // inputs.
4582 sort_expensive_nodes();
4583
4584 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4585 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4586 return true;
4587 }
4588 }
4589
4590 return false;
4591 }
4592
4593 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4594 if (_expensive_nodes.length() == 0) {
4595 return;
4596 }
4597
4598 assert(OptimizeExpensiveOps, "optimization off?");
4599
4600 // Sort to bring similar nodes next to each other and clear the
4601 // control input of nodes for which there's only a single copy.
4602 sort_expensive_nodes();
4603
4604 int j = 0;
4605 int identical = 0;
4606 int i = 0;
4607 bool modified = false;
4608 for (; i < _expensive_nodes.length()-1; i++) {
4609 assert(j <= i, "can't write beyond current index");
4610 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4611 identical++;
4612 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4613 continue;
4614 }
4615 if (identical > 0) {
4616 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4617 identical = 0;
4618 } else {
4619 Node* n = _expensive_nodes.at(i);
4620 igvn.replace_input_of(n, 0, NULL);
4621 igvn.hash_insert(n);
4622 modified = true;
4623 }
4624 }
4625 if (identical > 0) {
4626 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4627 } else if (_expensive_nodes.length() >= 1) {
4628 Node* n = _expensive_nodes.at(i);
4629 igvn.replace_input_of(n, 0, NULL);
4630 igvn.hash_insert(n);
4631 modified = true;
4632 }
4633 _expensive_nodes.trunc_to(j);
4634 if (modified) {
4635 igvn.optimize();
4636 }
4637 }
4638
4639 void Compile::add_expensive_node(Node * n) {
4640 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4641 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4642 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4643 if (OptimizeExpensiveOps) {
4644 _expensive_nodes.append(n);
4645 } else {
4646 // Clear control input and let IGVN optimize expensive nodes if
4647 // OptimizeExpensiveOps is off.
4648 n->set_req(0, NULL);
4649 }
4650 }
4651
4652 /**
4653 * Track coarsened Lock and Unlock nodes.
4654 */
4655
4656 class Lock_List : public Node_List {
4657 uint _origin_cnt;
4658 public:
4659 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4660 uint origin_cnt() const { return _origin_cnt; }
4661 };
4662
4663 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4664 int length = locks.length();
4665 if (length > 0) {
4666 // Have to keep this list until locks elimination during Macro nodes elimination.
4667 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4668 for (int i = 0; i < length; i++) {
4669 AbstractLockNode* lock = locks.at(i);
4670 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4671 locks_list->push(lock);
4672 }
4673 _coarsened_locks.append(locks_list);
4674 }
4675 }
4676
4677 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4678 int count = coarsened_count();
4679 for (int i = 0; i < count; i++) {
4680 Node_List* locks_list = _coarsened_locks.at(i);
4681 for (uint j = 0; j < locks_list->size(); j++) {
4682 Node* lock = locks_list->at(j);
4683 assert(lock->is_AbstractLock(), "sanity");
4684 if (!useful.member(lock)) {
4685 locks_list->yank(lock);
4686 }
4687 }
4688 }
4689 }
4690
4691 void Compile::remove_coarsened_lock(Node* n) {
4692 if (n->is_AbstractLock()) {
4693 int count = coarsened_count();
4694 for (int i = 0; i < count; i++) {
4695 Node_List* locks_list = _coarsened_locks.at(i);
4696 locks_list->yank(n);
4697 }
4698 }
4699 }
4700
4701 bool Compile::coarsened_locks_consistent() {
4702 int count = coarsened_count();
4703 for (int i = 0; i < count; i++) {
4704 bool unbalanced = false;
4705 bool modified = false; // track locks kind modifications
4706 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4707 uint size = locks_list->size();
4708 if (size == 0) {
4709 unbalanced = false; // All locks were eliminated - good
4710 } else if (size != locks_list->origin_cnt()) {
4711 unbalanced = true; // Some locks were removed from list
4712 } else {
4713 for (uint j = 0; j < size; j++) {
4714 Node* lock = locks_list->at(j);
4715 // All nodes in group should have the same state (modified or not)
4716 if (!lock->as_AbstractLock()->is_coarsened()) {
4717 if (j == 0) {
4718 // first on list was modified, the rest should be too for consistency
4719 modified = true;
4720 } else if (!modified) {
4721 // this lock was modified but previous locks on the list were not
4722 unbalanced = true;
4723 break;
4724 }
4725 } else if (modified) {
4726 // previous locks on list were modified but not this lock
4727 unbalanced = true;
4728 break;
4729 }
4730 }
4731 }
4732 if (unbalanced) {
4733 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4734 #ifdef ASSERT
4735 if (PrintEliminateLocks) {
4736 tty->print_cr("=== unbalanced coarsened locks ===");
4737 for (uint l = 0; l < size; l++) {
4738 locks_list->at(l)->dump();
4739 }
4740 }
4741 #endif
4742 record_failure(C2Compiler::retry_no_locks_coarsening());
4743 return false;
4744 }
4745 }
4746 return true;
4747 }
4748
4749 /**
4750 * Remove the speculative part of types and clean up the graph
4751 */
4752 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4753 if (UseTypeSpeculation) {
4754 Unique_Node_List worklist;
4755 worklist.push(root());
4756 int modified = 0;
4757 // Go over all type nodes that carry a speculative type, drop the
4758 // speculative part of the type and enqueue the node for an igvn
4759 // which may optimize it out.
4760 for (uint next = 0; next < worklist.size(); ++next) {
4761 Node *n = worklist.at(next);
4762 if (n->is_Type()) {
4763 TypeNode* tn = n->as_Type();
4764 const Type* t = tn->type();
4765 const Type* t_no_spec = t->remove_speculative();
4766 if (t_no_spec != t) {
4767 bool in_hash = igvn.hash_delete(n);
4768 assert(in_hash, "node should be in igvn hash table");
4769 tn->set_type(t_no_spec);
4770 igvn.hash_insert(n);
4771 igvn._worklist.push(n); // give it a chance to go away
4772 modified++;
4773 }
4774 }
4775 // Iterate over outs - endless loops is unreachable from below
4776 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4777 Node *m = n->fast_out(i);
4778 if (not_a_node(m)) {
4779 continue;
4780 }
4781 worklist.push(m);
4782 }
4783 }
4784 // Drop the speculative part of all types in the igvn's type table
4785 igvn.remove_speculative_types();
4786 if (modified > 0) {
4787 igvn.optimize();
4788 }
4789 #ifdef ASSERT
4790 // Verify that after the IGVN is over no speculative type has resurfaced
4791 worklist.clear();
4792 worklist.push(root());
4793 for (uint next = 0; next < worklist.size(); ++next) {
4794 Node *n = worklist.at(next);
4795 const Type* t = igvn.type_or_null(n);
4796 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4797 if (n->is_Type()) {
4798 t = n->as_Type()->type();
4799 assert(t == t->remove_speculative(), "no more speculative types");
4800 }
4801 // Iterate over outs - endless loops is unreachable from below
4802 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4803 Node *m = n->fast_out(i);
4804 if (not_a_node(m)) {
4805 continue;
4806 }
4807 worklist.push(m);
4808 }
4809 }
4810 igvn.check_no_speculative_types();
4811 #endif
4812 }
4813 }
4814
4815 // Auxiliary methods to support randomized stressing/fuzzing.
4816
4817 int Compile::random() {
4818 _stress_seed = os::next_random(_stress_seed);
4819 return static_cast<int>(_stress_seed);
4820 }
4821
4822 // This method can be called the arbitrary number of times, with current count
4823 // as the argument. The logic allows selecting a single candidate from the
4824 // running list of candidates as follows:
4825 // int count = 0;
4826 // Cand* selected = null;
4827 // while(cand = cand->next()) {
4828 // if (randomized_select(++count)) {
4829 // selected = cand;
4830 // }
4831 // }
4832 //
4833 // Including count equalizes the chances any candidate is "selected".
4834 // This is useful when we don't have the complete list of candidates to choose
4835 // from uniformly. In this case, we need to adjust the randomicity of the
4836 // selection, or else we will end up biasing the selection towards the latter
4837 // candidates.
4838 //
4839 // Quick back-envelope calculation shows that for the list of n candidates
4840 // the equal probability for the candidate to persist as "best" can be
4841 // achieved by replacing it with "next" k-th candidate with the probability
4842 // of 1/k. It can be easily shown that by the end of the run, the
4843 // probability for any candidate is converged to 1/n, thus giving the
4844 // uniform distribution among all the candidates.
4845 //
4846 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4847 #define RANDOMIZED_DOMAIN_POW 29
4848 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4849 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4850 bool Compile::randomized_select(int count) {
4851 assert(count > 0, "only positive");
4852 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4853 }
4854
4855 CloneMap& Compile::clone_map() { return _clone_map; }
4856 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4857
4858 void NodeCloneInfo::dump() const {
4859 tty->print(" {%d:%d} ", idx(), gen());
4860 }
4861
4862 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4863 uint64_t val = value(old->_idx);
4864 NodeCloneInfo cio(val);
4865 assert(val != 0, "old node should be in the map");
4866 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4867 insert(nnn->_idx, cin.get());
4868 #ifndef PRODUCT
4869 if (is_debug()) {
4870 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4871 }
4872 #endif
4873 }
4874
4875 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4876 NodeCloneInfo cio(value(old->_idx));
4877 if (cio.get() == 0) {
4878 cio.set(old->_idx, 0);
4879 insert(old->_idx, cio.get());
4880 #ifndef PRODUCT
4881 if (is_debug()) {
4882 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4883 }
4884 #endif
4885 }
4886 clone(old, nnn, gen);
4887 }
4888
4889 int CloneMap::max_gen() const {
4890 int g = 0;
4891 DictI di(_dict);
4892 for(; di.test(); ++di) {
4893 int t = gen(di._key);
4894 if (g < t) {
4895 g = t;
4896 #ifndef PRODUCT
4897 if (is_debug()) {
4898 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4899 }
4900 #endif
4901 }
4902 }
4903 return g;
4904 }
4905
4906 void CloneMap::dump(node_idx_t key) const {
4907 uint64_t val = value(key);
4908 if (val != 0) {
4909 NodeCloneInfo ni(val);
4910 ni.dump();
4911 }
4912 }
4913
4914 // Move Allocate nodes to the start of the list
4915 void Compile::sort_macro_nodes() {
4916 int count = macro_count();
4917 int allocates = 0;
4918 for (int i = 0; i < count; i++) {
4919 Node* n = macro_node(i);
4920 if (n->is_Allocate()) {
4921 if (i != allocates) {
4922 Node* tmp = macro_node(allocates);
4923 _macro_nodes.at_put(allocates, n);
4924 _macro_nodes.at_put(i, tmp);
4925 }
4926 allocates++;
4927 }
4928 }
4929 }
4930
4931 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
4932 EventCompilerPhase event;
4933 if (event.should_commit()) {
4934 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
4935 }
4936 #ifndef PRODUCT
4937 ResourceMark rm;
4938 stringStream ss;
4939 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
4940 if (n != nullptr) {
4941 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
4942 }
4943
4944 const char* name = ss.as_string();
4945 if (should_print_igv(level)) {
4946 _igv_printer->print_method(name, level);
4947 }
4948 if (should_print_phase(cpt)) {
4949 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
4950 }
4951 #endif
4952 C->_latest_stage_start_counter.stamp();
4953 }
4954
4955 // Only used from CompileWrapper
4956 void Compile::begin_method() {
4957 #ifndef PRODUCT
4958 if (_method != NULL && should_print_igv(1)) {
4959 _igv_printer->begin_method();
4960 }
4961 #endif
4962 C->_latest_stage_start_counter.stamp();
4963 }
4964
4965 // Only used from CompileWrapper
4966 void Compile::end_method() {
4967 EventCompilerPhase event;
4968 if (event.should_commit()) {
4969 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
4970 }
4971
4972 #ifndef PRODUCT
4973 if (_method != NULL && should_print_igv(1)) {
4974 _igv_printer->end_method();
4975 }
4976 #endif
4977 }
4978
4979 bool Compile::should_print_phase(CompilerPhaseType cpt) {
4980 #ifndef PRODUCT
4981 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
4982 return true;
4983 }
4984 #endif
4985 return false;
4986 }
4987
4988 bool Compile::should_print_igv(int level) {
4989 #ifndef PRODUCT
4990 if (PrintIdealGraphLevel < 0) { // disabled by the user
4991 return false;
4992 }
4993
4994 bool need = directive()->IGVPrintLevelOption >= level;
4995 if (need && !_igv_printer) {
4996 _igv_printer = IdealGraphPrinter::printer();
4997 _igv_printer->set_compile(this);
4998 }
4999 return need;
5000 #else
5001 return false;
5002 #endif
5003 }
5004
5005 #ifndef PRODUCT
5006 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
5007 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
5008
5009 // Called from debugger. Prints method to the default file with the default phase name.
5010 // This works regardless of any Ideal Graph Visualizer flags set or not.
5011 void igv_print() {
5012 Compile::current()->igv_print_method_to_file();
5013 }
5014
5015 // Same as igv_print() above but with a specified phase name.
5016 void igv_print(const char* phase_name) {
5017 Compile::current()->igv_print_method_to_file(phase_name);
5018 }
5019
5020 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5021 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5022 // This works regardless of any Ideal Graph Visualizer flags set or not.
5023 void igv_print(bool network) {
5024 if (network) {
5025 Compile::current()->igv_print_method_to_network();
5026 } else {
5027 Compile::current()->igv_print_method_to_file();
5028 }
5029 }
5030
5031 // Same as igv_print(bool network) above but with a specified phase name.
5032 void igv_print(bool network, const char* phase_name) {
5033 if (network) {
5034 Compile::current()->igv_print_method_to_network(phase_name);
5035 } else {
5036 Compile::current()->igv_print_method_to_file(phase_name);
5037 }
5038 }
5039
5040 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5041 void igv_print_default() {
5042 Compile::current()->print_method(PHASE_DEBUG, 0);
5043 }
5044
5045 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5046 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5047 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5048 void igv_append() {
5049 Compile::current()->igv_print_method_to_file("Debug", true);
5050 }
5051
5052 // Same as igv_append() above but with a specified phase name.
5053 void igv_append(const char* phase_name) {
5054 Compile::current()->igv_print_method_to_file(phase_name, true);
5055 }
5056
5057 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5058 const char* file_name = "custom_debug.xml";
5059 if (_debug_file_printer == NULL) {
5060 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5061 } else {
5062 _debug_file_printer->update_compiled_method(C->method());
5063 }
5064 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5065 _debug_file_printer->print(phase_name, (Node*)C->root());
5066 }
5067
5068 void Compile::igv_print_method_to_network(const char* phase_name) {
5069 if (_debug_network_printer == NULL) {
5070 _debug_network_printer = new IdealGraphPrinter(C);
5071 } else {
5072 _debug_network_printer->update_compiled_method(C->method());
5073 }
5074 tty->print_cr("Method printed over network stream to IGV");
5075 _debug_network_printer->print(phase_name, (Node*)C->root());
5076 }
5077 #endif
5078
5079 void Compile::add_native_invoker(RuntimeStub* stub) {
5080 _native_invokers.append(stub);
5081 }
5082
5083 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5084 if (type != NULL && phase->type(value)->higher_equal(type)) {
5085 return value;
5086 }
5087 Node* result = NULL;
5088 if (bt == T_BYTE) {
5089 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5090 result = new RShiftINode(result, phase->intcon(24));
5091 } else if (bt == T_BOOLEAN) {
5092 result = new AndINode(value, phase->intcon(0xFF));
5093 } else if (bt == T_CHAR) {
5094 result = new AndINode(value,phase->intcon(0xFFFF));
5095 } else {
5096 assert(bt == T_SHORT, "unexpected narrow type");
5097 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5098 result = new RShiftINode(result, phase->intcon(16));
5099 }
5100 if (transform_res) {
5101 result = phase->transform(result);
5102 }
5103 return result;
5104 }
5105