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
25 #include "precompiled.hpp"
26 #include "jvm_io.h"
27 #include "asm/macroAssembler.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "ci/ciReplay.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "code/exceptionHandlerTable.hpp"
32 #include "code/nmethod.hpp"
33 #include "compiler/compileBroker.hpp"
34 #include "compiler/compileLog.hpp"
35 #include "compiler/disassembler.hpp"
36 #include "compiler/oopMap.hpp"
37 #include "gc/shared/barrierSet.hpp"
38 #include "gc/shared/c2/barrierSetC2.hpp"
39 #include "jfr/jfrEvents.hpp"
40 #include "memory/resourceArea.hpp"
41 #include "opto/addnode.hpp"
42 #include "opto/block.hpp"
43 #include "opto/c2compiler.hpp"
44 #include "opto/callGenerator.hpp"
45 #include "opto/callnode.hpp"
46 #include "opto/castnode.hpp"
47 #include "opto/cfgnode.hpp"
48 #include "opto/chaitin.hpp"
49 #include "opto/compile.hpp"
50 #include "opto/connode.hpp"
51 #include "opto/convertnode.hpp"
52 #include "opto/divnode.hpp"
53 #include "opto/escape.hpp"
54 #include "opto/idealGraphPrinter.hpp"
55 #include "opto/inlinetypenode.hpp"
56 #include "opto/loopnode.hpp"
57 #include "opto/machnode.hpp"
58 #include "opto/macro.hpp"
59 #include "opto/matcher.hpp"
60 #include "opto/mathexactnode.hpp"
61 #include "opto/memnode.hpp"
62 #include "opto/mulnode.hpp"
63 #include "opto/narrowptrnode.hpp"
64 #include "opto/node.hpp"
65 #include "opto/opcodes.hpp"
66 #include "opto/output.hpp"
67 #include "opto/parse.hpp"
68 #include "opto/phaseX.hpp"
69 #include "opto/rootnode.hpp"
70 #include "opto/runtime.hpp"
71 #include "opto/stringopts.hpp"
72 #include "opto/type.hpp"
73 #include "opto/vector.hpp"
74 #include "opto/vectornode.hpp"
75 #include "runtime/globals_extension.hpp"
76 #include "runtime/sharedRuntime.hpp"
77 #include "runtime/signature.hpp"
78 #include "runtime/stubRoutines.hpp"
79 #include "runtime/timer.hpp"
80 #include "utilities/align.hpp"
81 #include "utilities/copy.hpp"
82 #include "utilities/macros.hpp"
83 #include "utilities/resourceHash.hpp"
84
85
86 // -------------------- Compile::mach_constant_base_node -----------------------
87 // Constant table base node singleton.
88 MachConstantBaseNode* Compile::mach_constant_base_node() {
89 if (_mach_constant_base_node == NULL) {
90 _mach_constant_base_node = new MachConstantBaseNode();
91 _mach_constant_base_node->add_req(C->root());
92 }
93 return _mach_constant_base_node;
94 }
95
96
97 /// Support for intrinsics.
98
99 // Return the index at which m must be inserted (or already exists).
100 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
101 class IntrinsicDescPair {
102 private:
103 ciMethod* _m;
104 bool _is_virtual;
105 public:
106 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
107 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
108 ciMethod* m= elt->method();
109 ciMethod* key_m = key->_m;
110 if (key_m < m) return -1;
111 else if (key_m > m) return 1;
112 else {
113 bool is_virtual = elt->is_virtual();
114 bool key_virtual = key->_is_virtual;
115 if (key_virtual < is_virtual) return -1;
116 else if (key_virtual > is_virtual) return 1;
117 else return 0;
118 }
119 }
120 };
121 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
122 #ifdef ASSERT
123 for (int i = 1; i < _intrinsics.length(); i++) {
124 CallGenerator* cg1 = _intrinsics.at(i-1);
125 CallGenerator* cg2 = _intrinsics.at(i);
126 assert(cg1->method() != cg2->method()
127 ? cg1->method() < cg2->method()
128 : cg1->is_virtual() < cg2->is_virtual(),
129 "compiler intrinsics list must stay sorted");
130 }
131 #endif
132 IntrinsicDescPair pair(m, is_virtual);
133 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
134 }
135
136 void Compile::register_intrinsic(CallGenerator* cg) {
137 bool found = false;
138 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
139 assert(!found, "registering twice");
140 _intrinsics.insert_before(index, cg);
141 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
142 }
143
144 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
145 assert(m->is_loaded(), "don't try this on unloaded methods");
146 if (_intrinsics.length() > 0) {
147 bool found = false;
148 int index = intrinsic_insertion_index(m, is_virtual, found);
149 if (found) {
150 return _intrinsics.at(index);
151 }
152 }
153 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
154 if (m->intrinsic_id() != vmIntrinsics::_none &&
155 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
156 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
157 if (cg != NULL) {
158 // Save it for next time:
159 register_intrinsic(cg);
160 return cg;
161 } else {
162 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
163 }
164 }
165 return NULL;
166 }
167
168 // Compile::make_vm_intrinsic is defined in library_call.cpp.
169
170 #ifndef PRODUCT
171 // statistics gathering...
172
173 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
174 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
175
176 inline int as_int(vmIntrinsics::ID id) {
177 return vmIntrinsics::as_int(id);
178 }
179
180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
182 int oflags = _intrinsic_hist_flags[as_int(id)];
183 assert(flags != 0, "what happened?");
184 if (is_virtual) {
185 flags |= _intrinsic_virtual;
186 }
187 bool changed = (flags != oflags);
188 if ((flags & _intrinsic_worked) != 0) {
189 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
190 if (count == 1) {
191 changed = true; // first time
192 }
193 // increment the overall count also:
194 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
195 }
196 if (changed) {
197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
198 // Something changed about the intrinsic's virtuality.
199 if ((flags & _intrinsic_virtual) != 0) {
200 // This is the first use of this intrinsic as a virtual call.
201 if (oflags != 0) {
202 // We already saw it as a non-virtual, so note both cases.
203 flags |= _intrinsic_both;
204 }
205 } else if ((oflags & _intrinsic_both) == 0) {
206 // This is the first use of this intrinsic as a non-virtual
207 flags |= _intrinsic_both;
208 }
209 }
210 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
211 }
212 // update the overall flags also:
213 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
214 return changed;
215 }
216
217 static char* format_flags(int flags, char* buf) {
218 buf[0] = 0;
219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
224 if (buf[0] == 0) strcat(buf, ",");
225 assert(buf[0] == ',', "must be");
226 return &buf[1];
227 }
228
229 void Compile::print_intrinsic_statistics() {
230 char flagsbuf[100];
231 ttyLocker ttyl;
232 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
233 tty->print_cr("Compiler intrinsic usage:");
234 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
235 if (total == 0) total = 1; // avoid div0 in case of no successes
236 #define PRINT_STAT_LINE(name, c, f) \
237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
238 for (auto id : EnumRange<vmIntrinsicID>{}) {
239 int flags = _intrinsic_hist_flags[as_int(id)];
240 juint count = _intrinsic_hist_count[as_int(id)];
241 if ((flags | count) != 0) {
242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
243 }
244 }
245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
246 if (xtty != NULL) xtty->tail("statistics");
247 }
248
249 void Compile::print_statistics() {
250 { ttyLocker ttyl;
251 if (xtty != NULL) xtty->head("statistics type='opto'");
252 Parse::print_statistics();
253 PhaseCCP::print_statistics();
254 PhaseRegAlloc::print_statistics();
255 PhaseOutput::print_statistics();
256 PhasePeephole::print_statistics();
257 PhaseIdealLoop::print_statistics();
258 if (xtty != NULL) xtty->tail("statistics");
259 }
260 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
261 // put this under its own <statistics> element.
262 print_intrinsic_statistics();
263 }
264 }
265 #endif //PRODUCT
266
267 void Compile::gvn_replace_by(Node* n, Node* nn) {
268 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
269 Node* use = n->last_out(i);
270 bool is_in_table = initial_gvn()->hash_delete(use);
271 uint uses_found = 0;
272 for (uint j = 0; j < use->len(); j++) {
273 if (use->in(j) == n) {
274 if (j < use->req())
275 use->set_req(j, nn);
276 else
277 use->set_prec(j, nn);
278 uses_found++;
279 }
280 }
281 if (is_in_table) {
282 // reinsert into table
283 initial_gvn()->hash_find_insert(use);
284 }
285 record_for_igvn(use);
286 i -= uses_found; // we deleted 1 or more copies of this edge
287 }
288 }
289
290
291 // Identify all nodes that are reachable from below, useful.
292 // Use breadth-first pass that records state in a Unique_Node_List,
293 // recursive traversal is slower.
294 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
295 int estimated_worklist_size = live_nodes();
296 useful.map( estimated_worklist_size, NULL ); // preallocate space
297
298 // Initialize worklist
299 if (root() != NULL) { useful.push(root()); }
300 // If 'top' is cached, declare it useful to preserve cached node
301 if( cached_top_node() ) { useful.push(cached_top_node()); }
302
303 // Push all useful nodes onto the list, breadthfirst
304 for( uint next = 0; next < useful.size(); ++next ) {
305 assert( next < unique(), "Unique useful nodes < total nodes");
306 Node *n = useful.at(next);
307 uint max = n->len();
308 for( uint i = 0; i < max; ++i ) {
309 Node *m = n->in(i);
310 if (not_a_node(m)) continue;
311 useful.push(m);
312 }
313 }
314 }
315
316 // Update dead_node_list with any missing dead nodes using useful
317 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
318 void Compile::update_dead_node_list(Unique_Node_List &useful) {
319 uint max_idx = unique();
320 VectorSet& useful_node_set = useful.member_set();
321
322 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
323 // If node with index node_idx is not in useful set,
324 // mark it as dead in dead node list.
325 if (!useful_node_set.test(node_idx)) {
326 record_dead_node(node_idx);
327 }
328 }
329 }
330
331 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
332 int shift = 0;
333 for (int i = 0; i < inlines->length(); i++) {
334 CallGenerator* cg = inlines->at(i);
335 if (useful.member(cg->call_node())) {
336 if (shift > 0) {
337 inlines->at_put(i - shift, cg);
338 }
339 } else {
340 shift++; // skip over the dead element
341 }
342 }
343 if (shift > 0) {
344 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
345 }
346 }
347
348 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
349 assert(dead != NULL && dead->is_Call(), "sanity");
350 int found = 0;
351 for (int i = 0; i < inlines->length(); i++) {
352 if (inlines->at(i)->call_node() == dead) {
353 inlines->remove_at(i);
354 found++;
355 NOT_DEBUG( break; ) // elements are unique, so exit early
356 }
357 }
358 assert(found <= 1, "not unique");
359 }
360
361 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
362 for (int i = node_list.length() - 1; i >= 0; i--) {
363 Node* n = node_list.at(i);
364 if (!useful.member(n)) {
365 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
366 }
367 }
368 }
369
370 void Compile::remove_useless_node(Node* dead) {
371 remove_modified_node(dead);
372
373 // Constant node that has no out-edges and has only one in-edge from
374 // root is usually dead. However, sometimes reshaping walk makes
375 // it reachable by adding use edges. So, we will NOT count Con nodes
376 // as dead to be conservative about the dead node count at any
377 // given time.
378 if (!dead->is_Con()) {
379 record_dead_node(dead->_idx);
380 }
381 if (dead->is_macro()) {
382 remove_macro_node(dead);
383 }
384 if (dead->is_expensive()) {
385 remove_expensive_node(dead);
386 }
387 if (dead->Opcode() == Op_Opaque4) {
388 remove_skeleton_predicate_opaq(dead);
389 }
390 if (dead->for_post_loop_opts_igvn()) {
391 remove_from_post_loop_opts_igvn(dead);
392 }
393 if (dead->is_InlineTypeBase()) {
394 remove_inline_type(dead);
395 }
396 if (dead->is_Call()) {
397 remove_useless_late_inlines( &_late_inlines, dead);
398 remove_useless_late_inlines( &_string_late_inlines, dead);
399 remove_useless_late_inlines( &_boxing_late_inlines, dead);
400 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
401 }
402 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
403 bs->unregister_potential_barrier_node(dead);
404 }
405
406 // Disconnect all useless nodes by disconnecting those at the boundary.
407 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
408 uint next = 0;
409 while (next < useful.size()) {
410 Node *n = useful.at(next++);
411 if (n->is_SafePoint()) {
412 // We're done with a parsing phase. Replaced nodes are not valid
413 // beyond that point.
414 n->as_SafePoint()->delete_replaced_nodes();
415 }
416 // Use raw traversal of out edges since this code removes out edges
417 int max = n->outcnt();
418 for (int j = 0; j < max; ++j) {
419 Node* child = n->raw_out(j);
420 if (!useful.member(child)) {
421 assert(!child->is_top() || child != top(),
422 "If top is cached in Compile object it is in useful list");
423 // Only need to remove this out-edge to the useless node
424 n->raw_del_out(j);
425 --j;
426 --max;
427 }
428 }
429 if (n->outcnt() == 1 && n->has_special_unique_user()) {
430 worklist->push(n->unique_out());
431 }
432 if (n->outcnt() == 0) {
433 worklist->push(n);
434 }
435 }
436
437 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
438 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes
439 remove_useless_nodes(_skeleton_predicate_opaqs, useful);
440 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
441 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
442 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
443 #ifdef ASSERT
444 if (_modified_nodes != NULL) {
445 _modified_nodes->remove_useless_nodes(useful.member_set());
446 }
447 #endif
448 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
449
450 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
451 bs->eliminate_useless_gc_barriers(useful, this);
452 // clean up the late inline lists
453 remove_useless_late_inlines( &_late_inlines, useful);
454 remove_useless_late_inlines( &_string_late_inlines, useful);
455 remove_useless_late_inlines( &_boxing_late_inlines, useful);
456 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
457 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
458 }
459
460 // ============================================================================
461 //------------------------------CompileWrapper---------------------------------
462 class CompileWrapper : public StackObj {
463 Compile *const _compile;
464 public:
465 CompileWrapper(Compile* compile);
466
467 ~CompileWrapper();
468 };
469
470 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
471 // the Compile* pointer is stored in the current ciEnv:
472 ciEnv* env = compile->env();
473 assert(env == ciEnv::current(), "must already be a ciEnv active");
474 assert(env->compiler_data() == NULL, "compile already active?");
475 env->set_compiler_data(compile);
476 assert(compile == Compile::current(), "sanity");
477
478 compile->set_type_dict(NULL);
479 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
480 compile->clone_map().set_clone_idx(0);
481 compile->set_type_last_size(0);
482 compile->set_last_tf(NULL, NULL);
483 compile->set_indexSet_arena(NULL);
484 compile->set_indexSet_free_block_list(NULL);
485 compile->init_type_arena();
486 Type::Initialize(compile);
487 _compile->begin_method();
488 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
489 }
490 CompileWrapper::~CompileWrapper() {
491 // simulate crash during compilation
492 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
493
494 _compile->end_method();
495 _compile->env()->set_compiler_data(NULL);
496 }
497
498
499 //----------------------------print_compile_messages---------------------------
500 void Compile::print_compile_messages() {
501 #ifndef PRODUCT
502 // Check if recompiling
503 if (!subsume_loads() && PrintOpto) {
504 // Recompiling without allowing machine instructions to subsume loads
505 tty->print_cr("*********************************************************");
506 tty->print_cr("** Bailout: Recompile without subsuming loads **");
507 tty->print_cr("*********************************************************");
508 }
509 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
510 // Recompiling without escape analysis
511 tty->print_cr("*********************************************************");
512 tty->print_cr("** Bailout: Recompile without escape analysis **");
513 tty->print_cr("*********************************************************");
514 }
515 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
516 // Recompiling without iterative escape analysis
517 tty->print_cr("*********************************************************");
518 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
519 tty->print_cr("*********************************************************");
520 }
521 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
522 // Recompiling without boxing elimination
523 tty->print_cr("*********************************************************");
524 tty->print_cr("** Bailout: Recompile without boxing elimination **");
525 tty->print_cr("*********************************************************");
526 }
527 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
528 // Recompiling without locks coarsening
529 tty->print_cr("*********************************************************");
530 tty->print_cr("** Bailout: Recompile without locks coarsening **");
531 tty->print_cr("*********************************************************");
532 }
533 if (env()->break_at_compile()) {
534 // Open the debugger when compiling this method.
535 tty->print("### Breaking when compiling: ");
536 method()->print_short_name();
537 tty->cr();
538 BREAKPOINT;
539 }
540
541 if( PrintOpto ) {
542 if (is_osr_compilation()) {
543 tty->print("[OSR]%3d", _compile_id);
544 } else {
545 tty->print("%3d", _compile_id);
546 }
547 }
548 #endif
549 }
550
551 #ifndef PRODUCT
552 void Compile::print_ideal_ir(const char* phase_name) {
553 ttyLocker ttyl;
554 // keep the following output all in one block
555 // This output goes directly to the tty, not the compiler log.
556 // To enable tools to match it up with the compilation activity,
557 // be sure to tag this tty output with the compile ID.
558 if (xtty != NULL) {
559 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
560 compile_id(),
561 is_osr_compilation() ? " compile_kind='osr'" : "",
562 phase_name);
563 }
564 root()->dump(9999);
565 if (xtty != NULL) {
566 xtty->tail("ideal");
567 }
568 }
569 #endif
570
571 // ============================================================================
572 //------------------------------Compile standard-------------------------------
573 debug_only( int Compile::_debug_idx = 100000; )
574
575 // Compile a method. entry_bci is -1 for normal compilations and indicates
576 // the continuation bci for on stack replacement.
577
578
579 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
580 Options options, DirectiveSet* directive)
581 : Phase(Compiler),
582 _compile_id(ci_env->compile_id()),
583 _options(options),
584 _method(target),
585 _entry_bci(osr_bci),
586 _ilt(NULL),
587 _stub_function(NULL),
588 _stub_name(NULL),
589 _stub_entry_point(NULL),
590 _max_node_limit(MaxNodeLimit),
591 _post_loop_opts_phase(false),
592 _inlining_progress(false),
593 _inlining_incrementally(false),
594 _do_cleanup(false),
595 _has_reserved_stack_access(target->has_reserved_stack_access()),
596 #ifndef PRODUCT
597 _igv_idx(0),
598 _trace_opto_output(directive->TraceOptoOutputOption),
599 #endif
600 _has_method_handle_invokes(false),
601 _clinit_barrier_on_entry(false),
602 _stress_seed(0),
603 _comp_arena(mtCompiler),
604 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
605 _env(ci_env),
606 _directive(directive),
607 _log(ci_env->log()),
608 _failure_reason(NULL),
609 _intrinsics (comp_arena(), 0, 0, NULL),
610 _macro_nodes (comp_arena(), 8, 0, NULL),
611 _predicate_opaqs (comp_arena(), 8, 0, NULL),
612 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
613 _expensive_nodes (comp_arena(), 8, 0, NULL),
614 _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
615 _inline_type_nodes (comp_arena(), 8, 0, NULL),
616 _coarsened_locks (comp_arena(), 8, 0, NULL),
617 _congraph(NULL),
618 NOT_PRODUCT(_igv_printer(NULL) COMMA)
619 _dead_node_list(comp_arena()),
620 _dead_node_count(0),
621 _node_arena(mtCompiler),
622 _old_arena(mtCompiler),
623 _mach_constant_base_node(NULL),
624 _Compile_types(mtCompiler),
625 _initial_gvn(NULL),
626 _for_igvn(NULL),
627 _late_inlines(comp_arena(), 2, 0, NULL),
628 _string_late_inlines(comp_arena(), 2, 0, NULL),
629 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
630 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
631 _late_inlines_pos(0),
632 _number_of_mh_late_inlines(0),
633 _native_invokers(comp_arena(), 1, 0, NULL),
634 _print_inlining_stream(NULL),
635 _print_inlining_list(NULL),
636 _print_inlining_idx(0),
637 _print_inlining_output(NULL),
638 _replay_inline_data(NULL),
639 _java_calls(0),
640 _inner_loops(0),
641 _interpreter_frame_size(0)
642 #ifndef PRODUCT
643 , _in_dump_cnt(0)
644 #endif
645 {
646 C = this;
647 CompileWrapper cw(this);
648
649 if (CITimeVerbose) {
650 tty->print(" ");
651 target->holder()->name()->print();
652 tty->print(".");
653 target->print_short_name();
654 tty->print(" ");
655 }
656 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
657 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
658
659 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
660 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
661 // We can always print a disassembly, either abstract (hex dump) or
662 // with the help of a suitable hsdis library. Thus, we should not
663 // couple print_assembly and print_opto_assembly controls.
664 // But: always print opto and regular assembly on compile command 'print'.
665 bool print_assembly = directive->PrintAssemblyOption;
666 set_print_assembly(print_opto_assembly || print_assembly);
667 #else
668 set_print_assembly(false); // must initialize.
669 #endif
670
671 #ifndef PRODUCT
672 set_parsed_irreducible_loop(false);
673
674 if (directive->ReplayInlineOption) {
675 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
676 }
677 #endif
678 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
679 set_print_intrinsics(directive->PrintIntrinsicsOption);
680 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
681
682 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
683 // Make sure the method being compiled gets its own MDO,
684 // so we can at least track the decompile_count().
685 // Need MDO to record RTM code generation state.
686 method()->ensure_method_data();
687 }
688
689 Init(::AliasLevel);
690
691
692 print_compile_messages();
693
694 _ilt = InlineTree::build_inline_tree_root();
695
696 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
697 assert(num_alias_types() >= AliasIdxRaw, "");
698
699 #define MINIMUM_NODE_HASH 1023
700 // Node list that Iterative GVN will start with
701 Unique_Node_List for_igvn(comp_arena());
702 set_for_igvn(&for_igvn);
703
704 // GVN that will be run immediately on new nodes
705 uint estimated_size = method()->code_size()*4+64;
706 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
707 PhaseGVN gvn(node_arena(), estimated_size);
708 set_initial_gvn(&gvn);
709
710 print_inlining_init();
711 { // Scope for timing the parser
712 TracePhase tp("parse", &timers[_t_parser]);
713
714 // Put top into the hash table ASAP.
715 initial_gvn()->transform_no_reclaim(top());
716
717 // Set up tf(), start(), and find a CallGenerator.
718 CallGenerator* cg = NULL;
719 if (is_osr_compilation()) {
720 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
721 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
722 initial_gvn()->set_type_bottom(s);
723 init_start(s);
724 cg = CallGenerator::for_osr(method(), entry_bci());
725 } else {
726 // Normal case.
727 init_tf(TypeFunc::make(method()));
728 StartNode* s = new StartNode(root(), tf()->domain_cc());
729 initial_gvn()->set_type_bottom(s);
730 init_start(s);
731 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
732 // With java.lang.ref.reference.get() we must go through the
733 // intrinsic - even when get() is the root
734 // method of the compile - so that, if necessary, the value in
735 // the referent field of the reference object gets recorded by
736 // the pre-barrier code.
737 cg = find_intrinsic(method(), false);
738 }
739 if (cg == NULL) {
740 float past_uses = method()->interpreter_invocation_count();
741 float expected_uses = past_uses;
742 cg = CallGenerator::for_inline(method(), expected_uses);
743 }
744 }
745 if (failing()) return;
746 if (cg == NULL) {
747 record_method_not_compilable("cannot parse method");
748 return;
749 }
750 JVMState* jvms = build_start_state(start(), tf());
751 if ((jvms = cg->generate(jvms)) == NULL) {
752 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
753 record_method_not_compilable("method parse failed");
754 }
755 return;
756 }
757 GraphKit kit(jvms);
758
759 if (!kit.stopped()) {
760 // Accept return values, and transfer control we know not where.
761 // This is done by a special, unique ReturnNode bound to root.
762 return_values(kit.jvms());
763 }
764
765 if (kit.has_exceptions()) {
766 // Any exceptions that escape from this call must be rethrown
767 // to whatever caller is dynamically above us on the stack.
768 // This is done by a special, unique RethrowNode bound to root.
769 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
770 }
771
772 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
773
774 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
775 inline_string_calls(true);
776 }
777
778 if (failing()) return;
779
780 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
781
782 // Remove clutter produced by parsing.
783 if (!failing()) {
784 ResourceMark rm;
785 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
786 }
787 }
788
789 // Note: Large methods are capped off in do_one_bytecode().
790 if (failing()) return;
791
792 // After parsing, node notes are no longer automagic.
793 // They must be propagated by register_new_node_with_optimizer(),
794 // clone(), or the like.
795 set_default_node_notes(NULL);
796
797 #ifndef PRODUCT
798 if (should_print_igv(1)) {
799 _igv_printer->print_inlining();
800 }
801 #endif
802
803 if (failing()) return;
804 NOT_PRODUCT( verify_graph_edges(); )
805
806 // If any phase is randomized for stress testing, seed random number
807 // generation and log the seed for repeatability.
808 if (StressLCM || StressGCM || StressIGVN || StressCCP) {
809 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
810 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
811 FLAG_SET_ERGO(StressSeed, _stress_seed);
812 } else {
813 _stress_seed = StressSeed;
814 }
815 if (_log != NULL) {
816 _log->elem("stress_test seed='%u'", _stress_seed);
817 }
818 }
819
820 // Now optimize
821 Optimize();
822 if (failing()) return;
823 NOT_PRODUCT( verify_graph_edges(); )
824
825 #ifndef PRODUCT
826 if (should_print_ideal()) {
827 print_ideal_ir("print_ideal");
828 }
829 #endif
830
831 #ifdef ASSERT
832 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
833 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
834 #endif
835
836 // Dump compilation data to replay it.
837 if (directive->DumpReplayOption) {
838 env()->dump_replay_data(_compile_id);
839 }
840 if (directive->DumpInlineOption && (ilt() != NULL)) {
841 env()->dump_inline_data(_compile_id);
842 }
843
844 // Now that we know the size of all the monitors we can add a fixed slot
845 // for the original deopt pc.
846 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
847 if (needs_stack_repair()) {
848 // One extra slot for the special stack increment value
849 next_slot += 2;
850 }
851 // TODO 8284443 Only reserve extra slot if needed
852 if (InlineTypeReturnedAsFields) {
853 // One extra slot to hold the IsInit information for a nullable
854 // inline type return if we run out of registers.
855 next_slot += 2;
856 }
857 set_fixed_slots(next_slot);
858
859 // Compute when to use implicit null checks. Used by matching trap based
860 // nodes and NullCheck optimization.
861 set_allowed_deopt_reasons();
862
863 // Now generate code
864 Code_Gen();
865 }
866
867 //------------------------------Compile----------------------------------------
868 // Compile a runtime stub
869 Compile::Compile( ciEnv* ci_env,
870 TypeFunc_generator generator,
871 address stub_function,
872 const char *stub_name,
873 int is_fancy_jump,
874 bool pass_tls,
875 bool return_pc,
876 DirectiveSet* directive)
877 : Phase(Compiler),
878 _compile_id(0),
879 _options(Options::for_runtime_stub()),
880 _method(NULL),
881 _entry_bci(InvocationEntryBci),
882 _stub_function(stub_function),
883 _stub_name(stub_name),
884 _stub_entry_point(NULL),
885 _max_node_limit(MaxNodeLimit),
886 _post_loop_opts_phase(false),
887 _inlining_progress(false),
888 _inlining_incrementally(false),
889 _has_reserved_stack_access(false),
890 #ifndef PRODUCT
891 _igv_idx(0),
892 _trace_opto_output(directive->TraceOptoOutputOption),
893 #endif
894 _has_method_handle_invokes(false),
895 _clinit_barrier_on_entry(false),
896 _stress_seed(0),
897 _comp_arena(mtCompiler),
898 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
899 _env(ci_env),
900 _directive(directive),
901 _log(ci_env->log()),
902 _failure_reason(NULL),
903 _congraph(NULL),
904 NOT_PRODUCT(_igv_printer(NULL) COMMA)
905 _dead_node_list(comp_arena()),
906 _dead_node_count(0),
907 _node_arena(mtCompiler),
908 _old_arena(mtCompiler),
909 _mach_constant_base_node(NULL),
910 _Compile_types(mtCompiler),
911 _initial_gvn(NULL),
912 _for_igvn(NULL),
913 _number_of_mh_late_inlines(0),
914 _native_invokers(),
915 _print_inlining_stream(NULL),
916 _print_inlining_list(NULL),
917 _print_inlining_idx(0),
918 _print_inlining_output(NULL),
919 _replay_inline_data(NULL),
920 _java_calls(0),
921 _inner_loops(0),
922 _interpreter_frame_size(0),
923 #ifndef PRODUCT
924 _in_dump_cnt(0),
925 #endif
926 _allowed_reasons(0) {
927 C = this;
928
929 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
930 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
931
932 #ifndef PRODUCT
933 set_print_assembly(PrintFrameConverterAssembly);
934 set_parsed_irreducible_loop(false);
935 #else
936 set_print_assembly(false); // Must initialize.
937 #endif
938 set_has_irreducible_loop(false); // no loops
939
940 CompileWrapper cw(this);
941 Init(/*AliasLevel=*/ 0);
942 init_tf((*generator)());
943
944 {
945 // The following is a dummy for the sake of GraphKit::gen_stub
946 Unique_Node_List for_igvn(comp_arena());
947 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
948 PhaseGVN gvn(Thread::current()->resource_area(),255);
949 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
950 gvn.transform_no_reclaim(top());
951
952 GraphKit kit;
953 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
954 }
955
956 NOT_PRODUCT( verify_graph_edges(); )
957
958 Code_Gen();
959 }
960
961 //------------------------------Init-------------------------------------------
962 // Prepare for a single compilation
963 void Compile::Init(int aliaslevel) {
964 _unique = 0;
965 _regalloc = NULL;
966
967 _tf = NULL; // filled in later
968 _top = NULL; // cached later
969 _matcher = NULL; // filled in later
970 _cfg = NULL; // filled in later
971
972 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
973
974 _node_note_array = NULL;
975 _default_node_notes = NULL;
976 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
977
978 _immutable_memory = NULL; // filled in at first inquiry
979
980 // Globally visible Nodes
981 // First set TOP to NULL to give safe behavior during creation of RootNode
982 set_cached_top_node(NULL);
983 set_root(new RootNode());
984 // Now that you have a Root to point to, create the real TOP
985 set_cached_top_node( new ConNode(Type::TOP) );
986 set_recent_alloc(NULL, NULL);
987
988 // Create Debug Information Recorder to record scopes, oopmaps, etc.
989 env()->set_oop_recorder(new OopRecorder(env()->arena()));
990 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
991 env()->set_dependencies(new Dependencies(env()));
992
993 _fixed_slots = 0;
994 set_has_split_ifs(false);
995 set_has_loops(false); // first approximation
996 set_has_stringbuilder(false);
997 set_has_boxed_value(false);
998 _trap_can_recompile = false; // no traps emitted yet
999 _major_progress = true; // start out assuming good things will happen
1000 set_has_unsafe_access(false);
1001 set_max_vector_size(0);
1002 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1003 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1004 set_decompile_count(0);
1005
1006 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1007 _loop_opts_cnt = LoopOptsCount;
1008 _has_flattened_accesses = false;
1009 _flattened_accesses_share_alias = true;
1010 _scalarize_in_safepoints = false;
1011
1012 set_do_inlining(Inline);
1013 set_max_inline_size(MaxInlineSize);
1014 set_freq_inline_size(FreqInlineSize);
1015 set_do_scheduling(OptoScheduling);
1016
1017 set_do_vector_loop(false);
1018
1019 if (AllowVectorizeOnDemand) {
1020 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1021 set_do_vector_loop(true);
1022 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());})
1023 } else if (has_method() && method()->name() != 0 &&
1024 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1025 set_do_vector_loop(true);
1026 }
1027 }
1028 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1029 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());})
1030
1031 set_age_code(has_method() && method()->profile_aging());
1032 set_rtm_state(NoRTM); // No RTM lock eliding by default
1033 _max_node_limit = _directive->MaxNodeLimitOption;
1034
1035 #if INCLUDE_RTM_OPT
1036 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1037 int rtm_state = method()->method_data()->rtm_state();
1038 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1039 // Don't generate RTM lock eliding code.
1040 set_rtm_state(NoRTM);
1041 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1042 // Generate RTM lock eliding code without abort ratio calculation code.
1043 set_rtm_state(UseRTM);
1044 } else if (UseRTMDeopt) {
1045 // Generate RTM lock eliding code and include abort ratio calculation
1046 // code if UseRTMDeopt is on.
1047 set_rtm_state(ProfileRTM);
1048 }
1049 }
1050 #endif
1051 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1052 set_clinit_barrier_on_entry(true);
1053 }
1054 if (debug_info()->recording_non_safepoints()) {
1055 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1056 (comp_arena(), 8, 0, NULL));
1057 set_default_node_notes(Node_Notes::make(this));
1058 }
1059
1060 // // -- Initialize types before each compile --
1061 // // Update cached type information
1062 // if( _method && _method->constants() )
1063 // Type::update_loaded_types(_method, _method->constants());
1064
1065 // Init alias_type map.
1066 if (!do_escape_analysis() && aliaslevel == 3) {
1067 aliaslevel = 2; // No unique types without escape analysis
1068 }
1069 _AliasLevel = aliaslevel;
1070 const int grow_ats = 16;
1071 _max_alias_types = grow_ats;
1072 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1073 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1074 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1075 {
1076 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1077 }
1078 // Initialize the first few types.
1079 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1080 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1081 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1082 _num_alias_types = AliasIdxRaw+1;
1083 // Zero out the alias type cache.
1084 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1085 // A NULL adr_type hits in the cache right away. Preload the right answer.
1086 probe_alias_cache(NULL)->_index = AliasIdxTop;
1087
1088 #ifdef ASSERT
1089 _type_verify_symmetry = true;
1090 _phase_optimize_finished = false;
1091 _exception_backedge = false;
1092 #endif
1093 }
1094
1095 //---------------------------init_start----------------------------------------
1096 // Install the StartNode on this compile object.
1097 void Compile::init_start(StartNode* s) {
1098 if (failing())
1099 return; // already failing
1100 assert(s == start(), "");
1101 }
1102
1103 /**
1104 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1105 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1106 * the ideal graph.
1107 */
1108 StartNode* Compile::start() const {
1109 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1110 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1111 Node* start = root()->fast_out(i);
1112 if (start->is_Start()) {
1113 return start->as_Start();
1114 }
1115 }
1116 fatal("Did not find Start node!");
1117 return NULL;
1118 }
1119
1120 //-------------------------------immutable_memory-------------------------------------
1121 // Access immutable memory
1122 Node* Compile::immutable_memory() {
1123 if (_immutable_memory != NULL) {
1124 return _immutable_memory;
1125 }
1126 StartNode* s = start();
1127 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1128 Node *p = s->fast_out(i);
1129 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1130 _immutable_memory = p;
1131 return _immutable_memory;
1132 }
1133 }
1134 ShouldNotReachHere();
1135 return NULL;
1136 }
1137
1138 //----------------------set_cached_top_node------------------------------------
1139 // Install the cached top node, and make sure Node::is_top works correctly.
1140 void Compile::set_cached_top_node(Node* tn) {
1141 if (tn != NULL) verify_top(tn);
1142 Node* old_top = _top;
1143 _top = tn;
1144 // Calling Node::setup_is_top allows the nodes the chance to adjust
1145 // their _out arrays.
1146 if (_top != NULL) _top->setup_is_top();
1147 if (old_top != NULL) old_top->setup_is_top();
1148 assert(_top == NULL || top()->is_top(), "");
1149 }
1150
1151 #ifdef ASSERT
1152 uint Compile::count_live_nodes_by_graph_walk() {
1153 Unique_Node_List useful(comp_arena());
1154 // Get useful node list by walking the graph.
1155 identify_useful_nodes(useful);
1156 return useful.size();
1157 }
1158
1159 void Compile::print_missing_nodes() {
1160
1161 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1162 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1163 return;
1164 }
1165
1166 // This is an expensive function. It is executed only when the user
1167 // specifies VerifyIdealNodeCount option or otherwise knows the
1168 // additional work that needs to be done to identify reachable nodes
1169 // by walking the flow graph and find the missing ones using
1170 // _dead_node_list.
1171
1172 Unique_Node_List useful(comp_arena());
1173 // Get useful node list by walking the graph.
1174 identify_useful_nodes(useful);
1175
1176 uint l_nodes = C->live_nodes();
1177 uint l_nodes_by_walk = useful.size();
1178
1179 if (l_nodes != l_nodes_by_walk) {
1180 if (_log != NULL) {
1181 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1182 _log->stamp();
1183 _log->end_head();
1184 }
1185 VectorSet& useful_member_set = useful.member_set();
1186 int last_idx = l_nodes_by_walk;
1187 for (int i = 0; i < last_idx; i++) {
1188 if (useful_member_set.test(i)) {
1189 if (_dead_node_list.test(i)) {
1190 if (_log != NULL) {
1191 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1192 }
1193 if (PrintIdealNodeCount) {
1194 // Print the log message to tty
1195 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1196 useful.at(i)->dump();
1197 }
1198 }
1199 }
1200 else if (! _dead_node_list.test(i)) {
1201 if (_log != NULL) {
1202 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1203 }
1204 if (PrintIdealNodeCount) {
1205 // Print the log message to tty
1206 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1207 }
1208 }
1209 }
1210 if (_log != NULL) {
1211 _log->tail("mismatched_nodes");
1212 }
1213 }
1214 }
1215 void Compile::record_modified_node(Node* n) {
1216 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1217 _modified_nodes->push(n);
1218 }
1219 }
1220
1221 void Compile::remove_modified_node(Node* n) {
1222 if (_modified_nodes != NULL) {
1223 _modified_nodes->remove(n);
1224 }
1225 }
1226 #endif
1227
1228 #ifndef PRODUCT
1229 void Compile::verify_top(Node* tn) const {
1230 if (tn != NULL) {
1231 assert(tn->is_Con(), "top node must be a constant");
1232 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1233 assert(tn->in(0) != NULL, "must have live top node");
1234 }
1235 }
1236 #endif
1237
1238
1239 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1240
1241 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1242 guarantee(arr != NULL, "");
1243 int num_blocks = arr->length();
1244 if (grow_by < num_blocks) grow_by = num_blocks;
1245 int num_notes = grow_by * _node_notes_block_size;
1246 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1247 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1248 while (num_notes > 0) {
1249 arr->append(notes);
1250 notes += _node_notes_block_size;
1251 num_notes -= _node_notes_block_size;
1252 }
1253 assert(num_notes == 0, "exact multiple, please");
1254 }
1255
1256 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1257 if (source == NULL || dest == NULL) return false;
1258
1259 if (dest->is_Con())
1260 return false; // Do not push debug info onto constants.
1261
1262 #ifdef ASSERT
1263 // Leave a bread crumb trail pointing to the original node:
1264 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1265 dest->set_debug_orig(source);
1266 }
1267 #endif
1268
1269 if (node_note_array() == NULL)
1270 return false; // Not collecting any notes now.
1271
1272 // This is a copy onto a pre-existing node, which may already have notes.
1273 // If both nodes have notes, do not overwrite any pre-existing notes.
1274 Node_Notes* source_notes = node_notes_at(source->_idx);
1275 if (source_notes == NULL || source_notes->is_clear()) return false;
1276 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1277 if (dest_notes == NULL || dest_notes->is_clear()) {
1278 return set_node_notes_at(dest->_idx, source_notes);
1279 }
1280
1281 Node_Notes merged_notes = (*source_notes);
1282 // The order of operations here ensures that dest notes will win...
1283 merged_notes.update_from(dest_notes);
1284 return set_node_notes_at(dest->_idx, &merged_notes);
1285 }
1286
1287
1288 //--------------------------allow_range_check_smearing-------------------------
1289 // Gating condition for coalescing similar range checks.
1290 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1291 // single covering check that is at least as strong as any of them.
1292 // If the optimization succeeds, the simplified (strengthened) range check
1293 // will always succeed. If it fails, we will deopt, and then give up
1294 // on the optimization.
1295 bool Compile::allow_range_check_smearing() const {
1296 // If this method has already thrown a range-check,
1297 // assume it was because we already tried range smearing
1298 // and it failed.
1299 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1300 return !already_trapped;
1301 }
1302
1303
1304 //------------------------------flatten_alias_type-----------------------------
1305 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1306 int offset = tj->offset();
1307 TypePtr::PTR ptr = tj->ptr();
1308
1309 // Known instance (scalarizable allocation) alias only with itself.
1310 bool is_known_inst = tj->isa_oopptr() != NULL &&
1311 tj->is_oopptr()->is_known_instance();
1312
1313 // Process weird unsafe references.
1314 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1315 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass();
1316 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops");
1317 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1318 tj = TypeOopPtr::BOTTOM;
1319 ptr = tj->ptr();
1320 offset = tj->offset();
1321 }
1322
1323 // Array pointers need some flattening
1324 const TypeAryPtr *ta = tj->isa_aryptr();
1325 if (ta && ta->is_stable()) {
1326 // Erase stability property for alias analysis.
1327 tj = ta = ta->cast_to_stable(false);
1328 }
1329 if (ta && ta->is_not_flat()) {
1330 // Erase not flat property for alias analysis.
1331 tj = ta = ta->cast_to_not_flat(false);
1332 }
1333 if (ta && ta->is_not_null_free()) {
1334 // Erase not null free property for alias analysis.
1335 tj = ta = ta->cast_to_not_null_free(false);
1336 }
1337
1338 if( ta && is_known_inst ) {
1339 if ( offset != Type::OffsetBot &&
1340 offset > arrayOopDesc::length_offset_in_bytes() ) {
1341 offset = Type::OffsetBot; // Flatten constant access into array body only
1342 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1343 }
1344 } else if( ta && _AliasLevel >= 2 ) {
1345 // For arrays indexed by constant indices, we flatten the alias
1346 // space to include all of the array body. Only the header, klass
1347 // and array length can be accessed un-aliased.
1348 // For flattened inline type array, each field has its own slice so
1349 // we must include the field offset.
1350 if( offset != Type::OffsetBot ) {
1351 if( ta->const_oop() ) { // MethodData* or Method*
1352 offset = Type::OffsetBot; // Flatten constant access into array body
1353 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1354 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1355 // range is OK as-is.
1356 tj = ta = TypeAryPtr::RANGE;
1357 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1358 tj = TypeInstPtr::KLASS; // all klass loads look alike
1359 ta = TypeAryPtr::RANGE; // generic ignored junk
1360 ptr = TypePtr::BotPTR;
1361 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1362 tj = TypeInstPtr::MARK;
1363 ta = TypeAryPtr::RANGE; // generic ignored junk
1364 ptr = TypePtr::BotPTR;
1365 } else { // Random constant offset into array body
1366 offset = Type::OffsetBot; // Flatten constant access into array body
1367 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1368 }
1369 }
1370 // Arrays of fixed size alias with arrays of unknown size.
1371 if (ta->size() != TypeInt::POS) {
1372 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1373 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1374 }
1375 // Arrays of known objects become arrays of unknown objects.
1376 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1377 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1378 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1379 }
1380 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1381 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1382 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1383 }
1384 // Initially all flattened array accesses share a single slice
1385 if (ta->is_flat() && ta->elem() != TypeInlineType::BOTTOM && _flattened_accesses_share_alias) {
1386 const TypeAry *tary = TypeAry::make(TypeInlineType::BOTTOM, ta->size());
1387 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1388 }
1389 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1390 // cannot be distinguished by bytecode alone.
1391 if (ta->elem() == TypeInt::BOOL) {
1392 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1393 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1394 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1395 }
1396 // During the 2nd round of IterGVN, NotNull castings are removed.
1397 // Make sure the Bottom and NotNull variants alias the same.
1398 // Also, make sure exact and non-exact variants alias the same.
1399 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1400 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1401 }
1402 }
1403
1404 // Oop pointers need some flattening
1405 const TypeInstPtr *to = tj->isa_instptr();
1406 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1407 ciInstanceKlass *k = to->klass()->as_instance_klass();
1408 if( ptr == TypePtr::Constant ) {
1409 if (to->klass() != ciEnv::current()->Class_klass() ||
1410 offset < k->layout_helper_size_in_bytes()) {
1411 // No constant oop pointers (such as Strings); they alias with
1412 // unknown strings.
1413 assert(!is_known_inst, "not scalarizable allocation");
1414 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1415 }
1416 } else if( is_known_inst ) {
1417 tj = to; // Keep NotNull and klass_is_exact for instance type
1418 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1419 // During the 2nd round of IterGVN, NotNull castings are removed.
1420 // Make sure the Bottom and NotNull variants alias the same.
1421 // Also, make sure exact and non-exact variants alias the same.
1422 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1423 }
1424 if (to->speculative() != NULL) {
1425 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->klass()->flatten_array(), to->instance_id());
1426 }
1427 // Canonicalize the holder of this field
1428 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1429 // First handle header references such as a LoadKlassNode, even if the
1430 // object's klass is unloaded at compile time (4965979).
1431 if (!is_known_inst) { // Do it only for non-instance types
1432 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1433 }
1434 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) {
1435 // Static fields are in the space above the normal instance
1436 // fields in the java.lang.Class instance.
1437 if (to->klass() != ciEnv::current()->Class_klass()) {
1438 to = NULL;
1439 tj = TypeOopPtr::BOTTOM;
1440 offset = tj->offset();
1441 }
1442 } else {
1443 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1444 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1445 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1446 if( is_known_inst ) {
1447 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), canonical_holder->flatten_array(), to->instance_id());
1448 } else {
1449 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1450 }
1451 }
1452 }
1453 }
1454
1455 // Klass pointers to object array klasses need some flattening
1456 const TypeKlassPtr *tk = tj->isa_klassptr();
1457 if( tk ) {
1458 // If we are referencing a field within a Klass, we need
1459 // to assume the worst case of an Object. Both exact and
1460 // inexact types must flatten to the same alias class so
1461 // use NotNull as the PTR.
1462 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1463
1464 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1465 TypeInstKlassPtr::OBJECT->klass(),
1466 Type::Offset(offset));
1467 }
1468
1469 ciKlass* klass = tk->klass();
1470 if (klass != NULL && klass->is_obj_array_klass()) {
1471 ciKlass* k = TypeAryPtr::OOPS->klass();
1472 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1473 k = TypeInstPtr::BOTTOM->klass();
1474 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1475 }
1476
1477 // Check for precise loads from the primary supertype array and force them
1478 // to the supertype cache alias index. Check for generic array loads from
1479 // the primary supertype array and also force them to the supertype cache
1480 // alias index. Since the same load can reach both, we need to merge
1481 // these 2 disparate memories into the same alias class. Since the
1482 // primary supertype array is read-only, there's no chance of confusion
1483 // where we bypass an array load and an array store.
1484 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1485 if (offset == Type::OffsetBot ||
1486 (offset >= primary_supers_offset &&
1487 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1488 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1489 offset = in_bytes(Klass::secondary_super_cache_offset());
1490 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1491 }
1492 }
1493
1494 // Flatten all Raw pointers together.
1495 if (tj->base() == Type::RawPtr)
1496 tj = TypeRawPtr::BOTTOM;
1497
1498 if (tj->base() == Type::AnyPtr)
1499 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1500
1501 // Flatten all to bottom for now
1502 switch( _AliasLevel ) {
1503 case 0:
1504 tj = TypePtr::BOTTOM;
1505 break;
1506 case 1: // Flatten to: oop, static, field or array
1507 switch (tj->base()) {
1508 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1509 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1510 case Type::AryPtr: // do not distinguish arrays at all
1511 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1512 case Type::KlassPtr:
1513 case Type::AryKlassPtr:
1514 case Type::InstKlassPtr: tj = TypeInstKlassPtr::OBJECT; break;
1515 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1516 default: ShouldNotReachHere();
1517 }
1518 break;
1519 case 2: // No collapsing at level 2; keep all splits
1520 case 3: // No collapsing at level 3; keep all splits
1521 break;
1522 default:
1523 Unimplemented();
1524 }
1525
1526 offset = tj->offset();
1527 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1528
1529 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1530 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1531 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1532 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1533 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1534 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1535 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1536 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1537 assert( tj->ptr() != TypePtr::TopPTR &&
1538 tj->ptr() != TypePtr::AnyNull &&
1539 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1540 // assert( tj->ptr() != TypePtr::Constant ||
1541 // tj->base() == Type::RawPtr ||
1542 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1543
1544 return tj;
1545 }
1546
1547 void Compile::AliasType::Init(int i, const TypePtr* at) {
1548 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1549 _index = i;
1550 _adr_type = at;
1551 _field = NULL;
1552 _element = NULL;
1553 _is_rewritable = true; // default
1554 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1555 if (atoop != NULL && atoop->is_known_instance()) {
1556 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1557 _general_index = Compile::current()->get_alias_index(gt);
1558 } else {
1559 _general_index = 0;
1560 }
1561 }
1562
1563 BasicType Compile::AliasType::basic_type() const {
1564 if (element() != NULL) {
1565 const Type* element = adr_type()->is_aryptr()->elem();
1566 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1567 } if (field() != NULL) {
1568 return field()->layout_type();
1569 } else {
1570 return T_ILLEGAL; // unknown
1571 }
1572 }
1573
1574 //---------------------------------print_on------------------------------------
1575 #ifndef PRODUCT
1576 void Compile::AliasType::print_on(outputStream* st) {
1577 if (index() < 10)
1578 st->print("@ <%d> ", index());
1579 else st->print("@ <%d>", index());
1580 st->print(is_rewritable() ? " " : " RO");
1581 int offset = adr_type()->offset();
1582 if (offset == Type::OffsetBot)
1583 st->print(" +any");
1584 else st->print(" +%-3d", offset);
1585 st->print(" in ");
1586 adr_type()->dump_on(st);
1587 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1588 if (field() != NULL && tjp) {
1589 if (tjp->klass() != field()->holder() ||
1590 tjp->offset() != field()->offset_in_bytes()) {
1591 st->print(" != ");
1592 field()->print();
1593 st->print(" ***");
1594 }
1595 }
1596 }
1597
1598 void print_alias_types() {
1599 Compile* C = Compile::current();
1600 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1601 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1602 C->alias_type(idx)->print_on(tty);
1603 tty->cr();
1604 }
1605 }
1606 #endif
1607
1608
1609 //----------------------------probe_alias_cache--------------------------------
1610 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1611 intptr_t key = (intptr_t) adr_type;
1612 key ^= key >> logAliasCacheSize;
1613 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1614 }
1615
1616
1617 //-----------------------------grow_alias_types--------------------------------
1618 void Compile::grow_alias_types() {
1619 const int old_ats = _max_alias_types; // how many before?
1620 const int new_ats = old_ats; // how many more?
1621 const int grow_ats = old_ats+new_ats; // how many now?
1622 _max_alias_types = grow_ats;
1623 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1624 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1625 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1626 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1627 }
1628
1629
1630 //--------------------------------find_alias_type------------------------------
1631 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1632 if (_AliasLevel == 0)
1633 return alias_type(AliasIdxBot);
1634
1635 AliasCacheEntry* ace = NULL;
1636 if (!uncached) {
1637 ace = probe_alias_cache(adr_type);
1638 if (ace->_adr_type == adr_type) {
1639 return alias_type(ace->_index);
1640 }
1641 }
1642
1643 // Handle special cases.
1644 if (adr_type == NULL) return alias_type(AliasIdxTop);
1645 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1646
1647 // Do it the slow way.
1648 const TypePtr* flat = flatten_alias_type(adr_type);
1649
1650 #ifdef ASSERT
1651 {
1652 ResourceMark rm;
1653 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1654 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1655 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1656 Type::str(adr_type));
1657 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1658 const TypeOopPtr* foop = flat->is_oopptr();
1659 // Scalarizable allocations have exact klass always.
1660 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1661 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1662 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1663 Type::str(foop), Type::str(xoop));
1664 }
1665 }
1666 #endif
1667
1668 int idx = AliasIdxTop;
1669 for (int i = 0; i < num_alias_types(); i++) {
1670 if (alias_type(i)->adr_type() == flat) {
1671 idx = i;
1672 break;
1673 }
1674 }
1675
1676 if (idx == AliasIdxTop) {
1677 if (no_create) return NULL;
1678 // Grow the array if necessary.
1679 if (_num_alias_types == _max_alias_types) grow_alias_types();
1680 // Add a new alias type.
1681 idx = _num_alias_types++;
1682 _alias_types[idx]->Init(idx, flat);
1683 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1684 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1685 if (flat->isa_instptr()) {
1686 if (flat->offset() == java_lang_Class::klass_offset()
1687 && flat->is_instptr()->klass() == env()->Class_klass())
1688 alias_type(idx)->set_rewritable(false);
1689 }
1690 ciField* field = NULL;
1691 if (flat->isa_aryptr()) {
1692 #ifdef ASSERT
1693 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1694 // (T_BYTE has the weakest alignment and size restrictions...)
1695 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1696 #endif
1697 const Type* elemtype = flat->is_aryptr()->elem();
1698 if (flat->offset() == TypePtr::OffsetBot) {
1699 alias_type(idx)->set_element(elemtype);
1700 }
1701 int field_offset = flat->is_aryptr()->field_offset().get();
1702 if (elemtype->isa_inlinetype() &&
1703 field_offset != Type::OffsetBot) {
1704 ciInlineKlass* vk = elemtype->inline_klass();
1705 field_offset += vk->first_field_offset();
1706 field = vk->get_field_by_offset(field_offset, false);
1707 }
1708 }
1709 if (flat->isa_klassptr()) {
1710 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1711 alias_type(idx)->set_rewritable(false);
1712 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1713 alias_type(idx)->set_rewritable(false);
1714 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1717 alias_type(idx)->set_rewritable(false);
1718 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1719 alias_type(idx)->set_rewritable(false);
1720 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1721 alias_type(idx)->set_rewritable(false);
1722 }
1723 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1724 // but the base pointer type is not distinctive enough to identify
1725 // references into JavaThread.)
1726
1727 // Check for final fields.
1728 const TypeInstPtr* tinst = flat->isa_instptr();
1729 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1730 if (tinst->const_oop() != NULL &&
1731 tinst->klass() == ciEnv::current()->Class_klass() &&
1732 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) {
1733 // static field
1734 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1735 field = k->get_field_by_offset(tinst->offset(), true);
1736 } else if (tinst->klass()->is_inlinetype()) {
1737 // Inline type field
1738 ciInlineKlass* vk = tinst->inline_klass();
1739 field = vk->get_field_by_offset(tinst->offset(), false);
1740 } else {
1741 ciInstanceKlass* k = tinst->klass()->as_instance_klass();
1742 field = k->get_field_by_offset(tinst->offset(), false);
1743 }
1744 }
1745 assert(field == NULL ||
1746 original_field == NULL ||
1747 (field->holder() == original_field->holder() &&
1748 field->offset() == original_field->offset() &&
1749 field->is_static() == original_field->is_static()), "wrong field?");
1750 // Set field() and is_rewritable() attributes.
1751 if (field != NULL) {
1752 alias_type(idx)->set_field(field);
1753 if (flat->isa_aryptr()) {
1754 // Fields of flat arrays are rewritable although they are declared final
1755 assert(flat->is_aryptr()->is_flat(), "must be a flat array");
1756 alias_type(idx)->set_rewritable(true);
1757 }
1758 }
1759 }
1760
1761 // Fill the cache for next time.
1762 if (!uncached) {
1763 ace->_adr_type = adr_type;
1764 ace->_index = idx;
1765 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1766
1767 // Might as well try to fill the cache for the flattened version, too.
1768 AliasCacheEntry* face = probe_alias_cache(flat);
1769 if (face->_adr_type == NULL) {
1770 face->_adr_type = flat;
1771 face->_index = idx;
1772 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1773 }
1774 }
1775
1776 return alias_type(idx);
1777 }
1778
1779
1780 Compile::AliasType* Compile::alias_type(ciField* field) {
1781 const TypeOopPtr* t;
1782 if (field->is_static())
1783 t = TypeInstPtr::make(field->holder()->java_mirror());
1784 else
1785 t = TypeOopPtr::make_from_klass_raw(field->holder());
1786 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1787 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1788 return atp;
1789 }
1790
1791
1792 //------------------------------have_alias_type--------------------------------
1793 bool Compile::have_alias_type(const TypePtr* adr_type) {
1794 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1795 if (ace->_adr_type == adr_type) {
1796 return true;
1797 }
1798
1799 // Handle special cases.
1800 if (adr_type == NULL) return true;
1801 if (adr_type == TypePtr::BOTTOM) return true;
1802
1803 return find_alias_type(adr_type, true, NULL) != NULL;
1804 }
1805
1806 //-----------------------------must_alias--------------------------------------
1807 // True if all values of the given address type are in the given alias category.
1808 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1809 if (alias_idx == AliasIdxBot) return true; // the universal category
1810 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1811 if (alias_idx == AliasIdxTop) return false; // the empty category
1812 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1813
1814 // the only remaining possible overlap is identity
1815 int adr_idx = get_alias_index(adr_type);
1816 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1817 assert(adr_idx == alias_idx ||
1818 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1819 && adr_type != TypeOopPtr::BOTTOM),
1820 "should not be testing for overlap with an unsafe pointer");
1821 return adr_idx == alias_idx;
1822 }
1823
1824 //------------------------------can_alias--------------------------------------
1825 // True if any values of the given address type are in the given alias category.
1826 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1827 if (alias_idx == AliasIdxTop) return false; // the empty category
1828 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1829 // Known instance doesn't alias with bottom memory
1830 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1831 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1832
1833 // the only remaining possible overlap is identity
1834 int adr_idx = get_alias_index(adr_type);
1835 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1836 return adr_idx == alias_idx;
1837 }
1838
1839 //---------------------cleanup_loop_predicates-----------------------
1840 // Remove the opaque nodes that protect the predicates so that all unused
1841 // checks and uncommon_traps will be eliminated from the ideal graph
1842 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1843 if (predicate_count()==0) return;
1844 for (int i = predicate_count(); i > 0; i--) {
1845 Node * n = predicate_opaque1_node(i-1);
1846 assert(n->Opcode() == Op_Opaque1, "must be");
1847 igvn.replace_node(n, n->in(1));
1848 }
1849 assert(predicate_count()==0, "should be clean!");
1850 }
1851
1852 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1853 if (!n->for_post_loop_opts_igvn()) {
1854 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1855 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1856 _for_post_loop_igvn.append(n);
1857 }
1858 }
1859
1860 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1861 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1862 _for_post_loop_igvn.remove(n);
1863 }
1864
1865 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1866 // Verify that all previous optimizations produced a valid graph
1867 // at least to this point, even if no loop optimizations were done.
1868 PhaseIdealLoop::verify(igvn);
1869
1870 C->set_post_loop_opts_phase(); // no more loop opts allowed
1871
1872 assert(!C->major_progress(), "not cleared");
1873
1874 if (_for_post_loop_igvn.length() > 0) {
1875 while (_for_post_loop_igvn.length() > 0) {
1876 Node* n = _for_post_loop_igvn.pop();
1877 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1878 igvn._worklist.push(n);
1879 }
1880 igvn.optimize();
1881 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1882
1883 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1884 if (C->major_progress()) {
1885 C->clear_major_progress(); // ensure that major progress is now clear
1886 }
1887 }
1888 }
1889
1890 void Compile::add_inline_type(Node* n) {
1891 assert(n->is_InlineTypeBase(), "unexpected node");
1892 _inline_type_nodes.push(n);
1893 }
1894
1895 void Compile::remove_inline_type(Node* n) {
1896 assert(n->is_InlineTypeBase(), "unexpected node");
1897 if (_inline_type_nodes.contains(n)) {
1898 _inline_type_nodes.remove(n);
1899 }
1900 }
1901
1902 // Does the return value keep otherwise useless inline type allocations alive?
1903 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1904 ResourceMark rm;
1905 Unique_Node_List wq;
1906 wq.push(ret_val);
1907 bool some_allocations = false;
1908 for (uint i = 0; i < wq.size(); i++) {
1909 Node* n = wq.at(i);
1910 assert(!n->is_InlineType(), "chain of inline type nodes");
1911 if (n->outcnt() > 1) {
1912 // Some other use for the allocation
1913 return false;
1914 } else if (n->is_InlineTypePtr()) {
1915 wq.push(n->in(1));
1916 } else if (n->is_Phi()) {
1917 for (uint j = 1; j < n->req(); j++) {
1918 wq.push(n->in(j));
1919 }
1920 } else if (n->is_CheckCastPP() &&
1921 n->in(1)->is_Proj() &&
1922 n->in(1)->in(0)->is_Allocate()) {
1923 some_allocations = true;
1924 } else if (n->is_CheckCastPP()) {
1925 wq.push(n->in(1));
1926 }
1927 }
1928 return some_allocations;
1929 }
1930
1931 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1932 // Make sure that the return value does not keep an otherwise unused allocation alive
1933 if (tf()->returns_inline_type_as_fields()) {
1934 Node* ret = NULL;
1935 for (uint i = 1; i < root()->req(); i++) {
1936 Node* in = root()->in(i);
1937 if (in->Opcode() == Op_Return) {
1938 assert(ret == NULL, "only one return");
1939 ret = in;
1940 }
1941 }
1942 if (ret != NULL) {
1943 Node* ret_val = ret->in(TypeFunc::Parms);
1944 if (igvn.type(ret_val)->isa_oopptr() &&
1945 return_val_keeps_allocations_alive(ret_val)) {
1946 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
1947 assert(ret_val->outcnt() == 0, "should be dead now");
1948 igvn.remove_dead_node(ret_val);
1949 }
1950 }
1951 }
1952 if (_inline_type_nodes.length() == 0) {
1953 return;
1954 }
1955 // Scalarize inline types in safepoint debug info.
1956 // Delay this until all inlining is over to avoid getting inconsistent debug info.
1957 set_scalarize_in_safepoints(true);
1958 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
1959 _inline_type_nodes.at(i)->as_InlineTypeBase()->make_scalar_in_safepoints(&igvn);
1960 }
1961 if (remove) {
1962 // Remove inline type nodes
1963 while (_inline_type_nodes.length() > 0) {
1964 InlineTypeBaseNode* vt = _inline_type_nodes.pop()->as_InlineTypeBase();
1965 if (vt->outcnt() == 0) {
1966 igvn.remove_dead_node(vt);
1967 } else if (vt->is_InlineTypePtr()) {
1968 igvn.replace_node(vt, vt->get_oop());
1969 } else {
1970 // Check if any users are blackholes. If so, rewrite them to use either the
1971 // allocated buffer, or individual components, instead of the inline type node
1972 // that goes away.
1973 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
1974 if (vt->out(i)->is_Blackhole()) {
1975 BlackholeNode* bh = vt->out(i)->as_Blackhole();
1976
1977 // Unlink the old input
1978 int idx = bh->find_edge(vt);
1979 assert(idx != -1, "The edge should be there");
1980 bh->del_req(idx);
1981 --i;
1982
1983 if (vt->is_allocated(&igvn)) {
1984 // Already has the allocated instance, blackhole that
1985 bh->add_req(vt->get_oop());
1986 } else {
1987 // Not allocated yet, blackhole the components
1988 for (uint c = 0; c < vt->field_count(); c++) {
1989 bh->add_req(vt->field_value(c));
1990 }
1991 }
1992
1993 // Node modified, record for IGVN
1994 igvn.record_for_igvn(bh);
1995 }
1996 }
1997
1998 #ifdef ASSERT
1999 for (DUIterator_Fast imax, i = vt->fast_outs(imax); i < imax; i++) {
2000 assert(vt->fast_out(i)->is_InlineTypeBase(), "Unexpected inline type user");
2001 }
2002 #endif
2003 igvn.replace_node(vt, igvn.C->top());
2004 }
2005 }
2006 }
2007 igvn.optimize();
2008 }
2009
2010 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2011 if (!_has_flattened_accesses) {
2012 return;
2013 }
2014 // Initially, all flattened array accesses share the same slice to
2015 // keep dependencies with Object[] array accesses (that could be
2016 // to a flattened array) correct. We're done with parsing so we
2017 // now know all flattened array accesses in this compile
2018 // unit. Let's move flattened array accesses to their own slice,
2019 // one per element field. This should help memory access
2020 // optimizations.
2021 ResourceMark rm;
2022 Unique_Node_List wq;
2023 wq.push(root());
2024
2025 Node_List mergememnodes;
2026 Node_List memnodes;
2027
2028 // Alias index currently shared by all flattened memory accesses
2029 int index = get_alias_index(TypeAryPtr::INLINES);
2030
2031 // Find MergeMem nodes and flattened array accesses
2032 for (uint i = 0; i < wq.size(); i++) {
2033 Node* n = wq.at(i);
2034 if (n->is_Mem()) {
2035 const TypePtr* adr_type = NULL;
2036 if (n->Opcode() == Op_StoreCM) {
2037 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2038 } else {
2039 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2040 }
2041 if (adr_type == TypeAryPtr::INLINES) {
2042 memnodes.push(n);
2043 }
2044 } else if (n->is_MergeMem()) {
2045 MergeMemNode* mm = n->as_MergeMem();
2046 if (mm->memory_at(index) != mm->base_memory()) {
2047 mergememnodes.push(n);
2048 }
2049 }
2050 for (uint j = 0; j < n->req(); j++) {
2051 Node* m = n->in(j);
2052 if (m != NULL) {
2053 wq.push(m);
2054 }
2055 }
2056 }
2057
2058 if (memnodes.size() > 0) {
2059 _flattened_accesses_share_alias = false;
2060
2061 // We are going to change the slice for the flattened array
2062 // accesses so we need to clear the cache entries that refer to
2063 // them.
2064 for (uint i = 0; i < AliasCacheSize; i++) {
2065 AliasCacheEntry* ace = &_alias_cache[i];
2066 if (ace->_adr_type != NULL &&
2067 ace->_adr_type->isa_aryptr() &&
2068 ace->_adr_type->is_aryptr()->is_flat()) {
2069 ace->_adr_type = NULL;
2070 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the NULL adr_type resolves to AliasIdxTop
2071 }
2072 }
2073
2074 // Find what aliases we are going to add
2075 int start_alias = num_alias_types()-1;
2076 int stop_alias = 0;
2077
2078 for (uint i = 0; i < memnodes.size(); i++) {
2079 Node* m = memnodes.at(i);
2080 const TypePtr* adr_type = NULL;
2081 if (m->Opcode() == Op_StoreCM) {
2082 adr_type = m->in(MemNode::OopStore)->adr_type();
2083 if (adr_type != TypeAryPtr::INLINES) {
2084 // store was optimized out and we lost track of the adr_type
2085 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2086 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2087 get_alias_index(adr_type));
2088 igvn.register_new_node_with_optimizer(clone);
2089 igvn.replace_node(m, clone);
2090 }
2091 } else {
2092 adr_type = m->adr_type();
2093 #ifdef ASSERT
2094 m->as_Mem()->set_adr_type(adr_type);
2095 #endif
2096 }
2097 int idx = get_alias_index(adr_type);
2098 start_alias = MIN2(start_alias, idx);
2099 stop_alias = MAX2(stop_alias, idx);
2100 }
2101
2102 assert(stop_alias >= start_alias, "should have expanded aliases");
2103
2104 Node_Stack stack(0);
2105 #ifdef ASSERT
2106 VectorSet seen(Thread::current()->resource_area());
2107 #endif
2108 // Now let's fix the memory graph so each flattened array access
2109 // is moved to the right slice. Start from the MergeMem nodes.
2110 uint last = unique();
2111 for (uint i = 0; i < mergememnodes.size(); i++) {
2112 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2113 Node* n = current->memory_at(index);
2114 MergeMemNode* mm = NULL;
2115 do {
2116 // Follow memory edges through memory accesses, phis and
2117 // narrow membars and push nodes on the stack. Once we hit
2118 // bottom memory, we pop element off the stack one at a
2119 // time, in reverse order, and move them to the right slice
2120 // by changing their memory edges.
2121 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2122 assert(!seen.test_set(n->_idx), "");
2123 // Uses (a load for instance) will need to be moved to the
2124 // right slice as well and will get a new memory state
2125 // that we don't know yet. The use could also be the
2126 // backedge of a loop. We put a place holder node between
2127 // the memory node and its uses. We replace that place
2128 // holder with the correct memory state once we know it,
2129 // i.e. when nodes are popped off the stack. Using the
2130 // place holder make the logic work in the presence of
2131 // loops.
2132 if (n->outcnt() > 1) {
2133 Node* place_holder = NULL;
2134 assert(!n->has_out_with(Op_Node), "");
2135 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2136 Node* u = n->out(k);
2137 if (u != current && u->_idx < last) {
2138 bool success = false;
2139 for (uint l = 0; l < u->req(); l++) {
2140 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2141 continue;
2142 }
2143 Node* in = u->in(l);
2144 if (in == n) {
2145 if (place_holder == NULL) {
2146 place_holder = new Node(1);
2147 place_holder->init_req(0, n);
2148 }
2149 igvn.replace_input_of(u, l, place_holder);
2150 success = true;
2151 }
2152 }
2153 if (success) {
2154 --k;
2155 }
2156 }
2157 }
2158 }
2159 if (n->is_Phi()) {
2160 stack.push(n, 1);
2161 n = n->in(1);
2162 } else if (n->is_Mem()) {
2163 stack.push(n, n->req());
2164 n = n->in(MemNode::Memory);
2165 } else {
2166 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2167 stack.push(n, n->req());
2168 n = n->in(0)->in(TypeFunc::Memory);
2169 }
2170 } else {
2171 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2172 // Build a new MergeMem node to carry the new memory state
2173 // as we build it. IGVN should fold extraneous MergeMem
2174 // nodes.
2175 mm = MergeMemNode::make(n);
2176 igvn.register_new_node_with_optimizer(mm);
2177 while (stack.size() > 0) {
2178 Node* m = stack.node();
2179 uint idx = stack.index();
2180 if (m->is_Mem()) {
2181 // Move memory node to its new slice
2182 const TypePtr* adr_type = m->adr_type();
2183 int alias = get_alias_index(adr_type);
2184 Node* prev = mm->memory_at(alias);
2185 igvn.replace_input_of(m, MemNode::Memory, prev);
2186 mm->set_memory_at(alias, m);
2187 } else if (m->is_Phi()) {
2188 // We need as many new phis as there are new aliases
2189 igvn.replace_input_of(m, idx, mm);
2190 if (idx == m->req()-1) {
2191 Node* r = m->in(0);
2192 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2193 const Type* adr_type = get_adr_type(j);
2194 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat() || j == (uint)index) {
2195 continue;
2196 }
2197 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2198 igvn.register_new_node_with_optimizer(phi);
2199 for (uint k = 1; k < m->req(); k++) {
2200 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2201 }
2202 mm->set_memory_at(j, phi);
2203 }
2204 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2205 igvn.register_new_node_with_optimizer(base_phi);
2206 for (uint k = 1; k < m->req(); k++) {
2207 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2208 }
2209 mm->set_base_memory(base_phi);
2210 }
2211 } else {
2212 // This is a MemBarCPUOrder node from
2213 // Parse::array_load()/Parse::array_store(), in the
2214 // branch that handles flattened arrays hidden under
2215 // an Object[] array. We also need one new membar per
2216 // new alias to keep the unknown access that the
2217 // membars protect properly ordered with accesses to
2218 // known flattened array.
2219 assert(m->is_Proj(), "projection expected");
2220 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2221 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2222 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2223 const Type* adr_type = get_adr_type(j);
2224 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat() || j == (uint)index) {
2225 continue;
2226 }
2227 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2228 igvn.register_new_node_with_optimizer(mb);
2229 Node* mem = mm->memory_at(j);
2230 mb->init_req(TypeFunc::Control, ctrl);
2231 mb->init_req(TypeFunc::Memory, mem);
2232 ctrl = new ProjNode(mb, TypeFunc::Control);
2233 igvn.register_new_node_with_optimizer(ctrl);
2234 mem = new ProjNode(mb, TypeFunc::Memory);
2235 igvn.register_new_node_with_optimizer(mem);
2236 mm->set_memory_at(j, mem);
2237 }
2238 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2239 }
2240 if (idx < m->req()-1) {
2241 idx += 1;
2242 stack.set_index(idx);
2243 n = m->in(idx);
2244 break;
2245 }
2246 // Take care of place holder nodes
2247 if (m->has_out_with(Op_Node)) {
2248 Node* place_holder = m->find_out_with(Op_Node);
2249 if (place_holder != NULL) {
2250 Node* mm_clone = mm->clone();
2251 igvn.register_new_node_with_optimizer(mm_clone);
2252 Node* hook = new Node(1);
2253 hook->init_req(0, mm);
2254 igvn.replace_node(place_holder, mm_clone);
2255 hook->destruct(&igvn);
2256 }
2257 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2258 }
2259 stack.pop();
2260 }
2261 }
2262 } while(stack.size() > 0);
2263 // Fix the memory state at the MergeMem we started from
2264 igvn.rehash_node_delayed(current);
2265 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2266 const Type* adr_type = get_adr_type(j);
2267 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->is_flat()) {
2268 continue;
2269 }
2270 current->set_memory_at(j, mm);
2271 }
2272 current->set_memory_at(index, current->base_memory());
2273 }
2274 igvn.optimize();
2275 }
2276 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2277 #ifdef ASSERT
2278 if (!_flattened_accesses_share_alias) {
2279 wq.clear();
2280 wq.push(root());
2281 for (uint i = 0; i < wq.size(); i++) {
2282 Node* n = wq.at(i);
2283 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2284 for (uint j = 0; j < n->req(); j++) {
2285 Node* m = n->in(j);
2286 if (m != NULL) {
2287 wq.push(m);
2288 }
2289 }
2290 }
2291 }
2292 #endif
2293 }
2294
2295
2296 // StringOpts and late inlining of string methods
2297 void Compile::inline_string_calls(bool parse_time) {
2298 {
2299 // remove useless nodes to make the usage analysis simpler
2300 ResourceMark rm;
2301 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2302 }
2303
2304 {
2305 ResourceMark rm;
2306 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2307 PhaseStringOpts pso(initial_gvn(), for_igvn());
2308 print_method(PHASE_AFTER_STRINGOPTS, 3);
2309 }
2310
2311 // now inline anything that we skipped the first time around
2312 if (!parse_time) {
2313 _late_inlines_pos = _late_inlines.length();
2314 }
2315
2316 while (_string_late_inlines.length() > 0) {
2317 CallGenerator* cg = _string_late_inlines.pop();
2318 cg->do_late_inline();
2319 if (failing()) return;
2320 }
2321 _string_late_inlines.trunc_to(0);
2322 }
2323
2324 // Late inlining of boxing methods
2325 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2326 if (_boxing_late_inlines.length() > 0) {
2327 assert(has_boxed_value(), "inconsistent");
2328
2329 PhaseGVN* gvn = initial_gvn();
2330 set_inlining_incrementally(true);
2331
2332 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2333 for_igvn()->clear();
2334 gvn->replace_with(&igvn);
2335
2336 _late_inlines_pos = _late_inlines.length();
2337
2338 while (_boxing_late_inlines.length() > 0) {
2339 CallGenerator* cg = _boxing_late_inlines.pop();
2340 cg->do_late_inline();
2341 if (failing()) return;
2342 }
2343 _boxing_late_inlines.trunc_to(0);
2344
2345 inline_incrementally_cleanup(igvn);
2346
2347 set_inlining_incrementally(false);
2348 }
2349 }
2350
2351 bool Compile::inline_incrementally_one() {
2352 assert(IncrementalInline, "incremental inlining should be on");
2353
2354 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2355
2356 set_inlining_progress(false);
2357 set_do_cleanup(false);
2358
2359 for (int i = 0; i < _late_inlines.length(); i++) {
2360 _late_inlines_pos = i+1;
2361 CallGenerator* cg = _late_inlines.at(i);
2362 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2363 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2364 cg->do_late_inline();
2365 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2366 if (failing()) {
2367 return false;
2368 } else if (inlining_progress()) {
2369 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2370 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2371 break; // process one call site at a time
2372 }
2373 } else {
2374 // Ignore late inline direct calls when inlining is not allowed.
2375 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2376 }
2377 }
2378 // Remove processed elements.
2379 _late_inlines.remove_till(_late_inlines_pos);
2380 _late_inlines_pos = 0;
2381
2382 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2383
2384 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2385
2386 set_inlining_progress(false);
2387 set_do_cleanup(false);
2388
2389 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2390 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2391 }
2392
2393 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2394 {
2395 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2396 ResourceMark rm;
2397 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2398 }
2399 {
2400 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2401 igvn = PhaseIterGVN(initial_gvn());
2402 igvn.optimize();
2403 }
2404 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2405 }
2406
2407 // Perform incremental inlining until bound on number of live nodes is reached
2408 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2409 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2410
2411 set_inlining_incrementally(true);
2412 uint low_live_nodes = 0;
2413
2414 while (_late_inlines.length() > 0) {
2415 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2416 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2417 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2418 // PhaseIdealLoop is expensive so we only try it once we are
2419 // out of live nodes and we only try it again if the previous
2420 // helped got the number of nodes down significantly
2421 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2422 if (failing()) return;
2423 low_live_nodes = live_nodes();
2424 _major_progress = true;
2425 }
2426
2427 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2428 bool do_print_inlining = print_inlining() || print_intrinsics();
2429 if (do_print_inlining || log() != NULL) {
2430 // Print inlining message for candidates that we couldn't inline for lack of space.
2431 for (int i = 0; i < _late_inlines.length(); i++) {
2432 CallGenerator* cg = _late_inlines.at(i);
2433 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2434 if (do_print_inlining) {
2435 cg->print_inlining_late(msg);
2436 }
2437 log_late_inline_failure(cg, msg);
2438 }
2439 }
2440 break; // finish
2441 }
2442 }
2443
2444 for_igvn()->clear();
2445 initial_gvn()->replace_with(&igvn);
2446
2447 while (inline_incrementally_one()) {
2448 assert(!failing(), "inconsistent");
2449 }
2450 if (failing()) return;
2451
2452 inline_incrementally_cleanup(igvn);
2453
2454 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2455
2456 if (failing()) return;
2457
2458 if (_late_inlines.length() == 0) {
2459 break; // no more progress
2460 }
2461 }
2462 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2463
2464 if (_string_late_inlines.length() > 0) {
2465 assert(has_stringbuilder(), "inconsistent");
2466 for_igvn()->clear();
2467 initial_gvn()->replace_with(&igvn);
2468
2469 inline_string_calls(false);
2470
2471 if (failing()) return;
2472
2473 inline_incrementally_cleanup(igvn);
2474 }
2475
2476 set_inlining_incrementally(false);
2477 }
2478
2479 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2480 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2481 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2482 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2483 // as if "inlining_incrementally() == true" were set.
2484 assert(inlining_incrementally() == false, "not allowed");
2485 #ifdef ASSERT
2486 Unique_Node_List* modified_nodes = _modified_nodes;
2487 _modified_nodes = NULL;
2488 #endif
2489 assert(_late_inlines.length() > 0, "sanity");
2490
2491 while (_late_inlines.length() > 0) {
2492 for_igvn()->clear();
2493 initial_gvn()->replace_with(&igvn);
2494
2495 while (inline_incrementally_one()) {
2496 assert(!failing(), "inconsistent");
2497 }
2498 if (failing()) return;
2499
2500 inline_incrementally_cleanup(igvn);
2501 }
2502 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2503 }
2504
2505 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2506 if (_loop_opts_cnt > 0) {
2507 debug_only( int cnt = 0; );
2508 while (major_progress() && (_loop_opts_cnt > 0)) {
2509 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2510 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2511 PhaseIdealLoop::optimize(igvn, mode);
2512 _loop_opts_cnt--;
2513 if (failing()) return false;
2514 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2515 }
2516 }
2517 return true;
2518 }
2519
2520 // Remove edges from "root" to each SafePoint at a backward branch.
2521 // They were inserted during parsing (see add_safepoint()) to make
2522 // infinite loops without calls or exceptions visible to root, i.e.,
2523 // useful.
2524 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2525 Node *r = root();
2526 if (r != NULL) {
2527 for (uint i = r->req(); i < r->len(); ++i) {
2528 Node *n = r->in(i);
2529 if (n != NULL && n->is_SafePoint()) {
2530 r->rm_prec(i);
2531 if (n->outcnt() == 0) {
2532 igvn.remove_dead_node(n);
2533 }
2534 --i;
2535 }
2536 }
2537 // Parsing may have added top inputs to the root node (Path
2538 // leading to the Halt node proven dead). Make sure we get a
2539 // chance to clean them up.
2540 igvn._worklist.push(r);
2541 igvn.optimize();
2542 }
2543 }
2544
2545 //------------------------------Optimize---------------------------------------
2546 // Given a graph, optimize it.
2547 void Compile::Optimize() {
2548 TracePhase tp("optimizer", &timers[_t_optimizer]);
2549
2550 #ifndef PRODUCT
2551 if (env()->break_at_compile()) {
2552 BREAKPOINT;
2553 }
2554
2555 #endif
2556
2557 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2558 #ifdef ASSERT
2559 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2560 #endif
2561
2562 ResourceMark rm;
2563
2564 print_inlining_reinit();
2565
2566 NOT_PRODUCT( verify_graph_edges(); )
2567
2568 print_method(PHASE_AFTER_PARSING, 1);
2569
2570 {
2571 // Iterative Global Value Numbering, including ideal transforms
2572 // Initialize IterGVN with types and values from parse-time GVN
2573 PhaseIterGVN igvn(initial_gvn());
2574 #ifdef ASSERT
2575 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2576 #endif
2577 {
2578 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2579 igvn.optimize();
2580 }
2581
2582 if (failing()) return;
2583
2584 print_method(PHASE_ITER_GVN1, 2);
2585
2586 inline_incrementally(igvn);
2587
2588 print_method(PHASE_INCREMENTAL_INLINE, 2);
2589
2590 if (failing()) return;
2591
2592 if (eliminate_boxing()) {
2593 // Inline valueOf() methods now.
2594 inline_boxing_calls(igvn);
2595
2596 if (AlwaysIncrementalInline) {
2597 inline_incrementally(igvn);
2598 }
2599
2600 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2601
2602 if (failing()) return;
2603 }
2604
2605 // Remove the speculative part of types and clean up the graph from
2606 // the extra CastPP nodes whose only purpose is to carry them. Do
2607 // that early so that optimizations are not disrupted by the extra
2608 // CastPP nodes.
2609 remove_speculative_types(igvn);
2610
2611 // No more new expensive nodes will be added to the list from here
2612 // so keep only the actual candidates for optimizations.
2613 cleanup_expensive_nodes(igvn);
2614
2615 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2616 if (EnableVectorSupport && has_vbox_nodes()) {
2617 TracePhase tp("", &timers[_t_vector]);
2618 PhaseVector pv(igvn);
2619 pv.optimize_vector_boxes();
2620
2621 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2622 }
2623 assert(!has_vbox_nodes(), "sanity");
2624
2625 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2626 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2627 initial_gvn()->replace_with(&igvn);
2628 Unique_Node_List* old_worklist = for_igvn();
2629 old_worklist->clear();
2630 Unique_Node_List new_worklist(C->comp_arena());
2631 {
2632 ResourceMark rm;
2633 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2634 }
2635 Unique_Node_List* save_for_igvn = for_igvn();
2636 set_for_igvn(&new_worklist);
2637 igvn = PhaseIterGVN(initial_gvn());
2638 igvn.optimize();
2639 set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2640 }
2641
2642 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2643 // safepoints
2644 remove_root_to_sfpts_edges(igvn);
2645
2646 // Process inline type nodes now that all inlining is over
2647 process_inline_types(igvn);
2648
2649 adjust_flattened_array_access_aliases(igvn);
2650
2651 // Perform escape analysis
2652 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2653 if (has_loops()) {
2654 // Cleanup graph (remove dead nodes).
2655 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2656 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2657 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2658 if (failing()) return;
2659 }
2660 bool progress;
2661 do {
2662 ConnectionGraph::do_analysis(this, &igvn);
2663
2664 if (failing()) return;
2665
2666 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2667
2668 // Optimize out fields loads from scalar replaceable allocations.
2669 igvn.optimize();
2670 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2671
2672 if (failing()) return;
2673
2674 if (congraph() != NULL && macro_count() > 0) {
2675 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2676 PhaseMacroExpand mexp(igvn);
2677 mexp.eliminate_macro_nodes();
2678 igvn.set_delay_transform(false);
2679
2680 igvn.optimize();
2681 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2682
2683 if (failing()) return;
2684 }
2685 progress = do_iterative_escape_analysis() &&
2686 (macro_count() < mcount) &&
2687 ConnectionGraph::has_candidates(this);
2688 // Try again if candidates exist and made progress
2689 // by removing some allocations and/or locks.
2690 } while (progress);
2691 }
2692
2693 // Loop transforms on the ideal graph. Range Check Elimination,
2694 // peeling, unrolling, etc.
2695
2696 // Set loop opts counter
2697 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2698 {
2699 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2700 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2701 _loop_opts_cnt--;
2702 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2703 if (failing()) return;
2704 }
2705 // Loop opts pass if partial peeling occurred in previous pass
2706 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2707 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2708 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2709 _loop_opts_cnt--;
2710 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2711 if (failing()) return;
2712 }
2713 // Loop opts pass for loop-unrolling before CCP
2714 if(major_progress() && (_loop_opts_cnt > 0)) {
2715 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2716 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2717 _loop_opts_cnt--;
2718 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2719 }
2720 if (!failing()) {
2721 // Verify that last round of loop opts produced a valid graph
2722 PhaseIdealLoop::verify(igvn);
2723 }
2724 }
2725 if (failing()) return;
2726
2727 // Conditional Constant Propagation;
2728 PhaseCCP ccp( &igvn );
2729 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2730 {
2731 TracePhase tp("ccp", &timers[_t_ccp]);
2732 ccp.do_transform();
2733 }
2734 print_method(PHASE_CCP1, 2);
2735
2736 assert( true, "Break here to ccp.dump_old2new_map()");
2737
2738 // Iterative Global Value Numbering, including ideal transforms
2739 {
2740 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2741 igvn = ccp;
2742 igvn.optimize();
2743 }
2744 print_method(PHASE_ITER_GVN2, 2);
2745
2746 if (failing()) return;
2747
2748 // Loop transforms on the ideal graph. Range Check Elimination,
2749 // peeling, unrolling, etc.
2750 if (!optimize_loops(igvn, LoopOptsDefault)) {
2751 return;
2752 }
2753
2754 if (failing()) return;
2755
2756 C->clear_major_progress(); // ensure that major progress is now clear
2757
2758 process_for_post_loop_opts_igvn(igvn);
2759
2760 #ifdef ASSERT
2761 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2762 #endif
2763
2764 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2765
2766 if (_late_inlines.length() > 0) {
2767 // More opportunities to optimize virtual and MH calls.
2768 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2769 process_late_inline_calls_no_inline(igvn);
2770 }
2771
2772 {
2773 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2774 PhaseMacroExpand mex(igvn);
2775 if (mex.expand_macro_nodes()) {
2776 assert(failing(), "must bail out w/ explicit message");
2777 return;
2778 }
2779 print_method(PHASE_MACRO_EXPANSION, 2);
2780 }
2781
2782 // Process inline type nodes again and remove them. From here
2783 // on we don't need to keep track of field values anymore.
2784 process_inline_types(igvn, /* remove= */ true);
2785
2786 {
2787 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2788 if (bs->expand_barriers(this, igvn)) {
2789 assert(failing(), "must bail out w/ explicit message");
2790 return;
2791 }
2792 print_method(PHASE_BARRIER_EXPANSION, 2);
2793 }
2794
2795 if (C->max_vector_size() > 0) {
2796 C->optimize_logic_cones(igvn);
2797 igvn.optimize();
2798 }
2799
2800 DEBUG_ONLY( _modified_nodes = NULL; )
2801 DEBUG_ONLY( _late_inlines.clear(); )
2802
2803 assert(igvn._worklist.size() == 0, "not empty");
2804 } // (End scope of igvn; run destructor if necessary for asserts.)
2805
2806 check_no_dead_use();
2807
2808 process_print_inlining();
2809
2810 // A method with only infinite loops has no edges entering loops from root
2811 {
2812 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2813 if (final_graph_reshaping()) {
2814 assert(failing(), "must bail out w/ explicit message");
2815 return;
2816 }
2817 }
2818
2819 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2820 DEBUG_ONLY(set_phase_optimize_finished();)
2821 }
2822
2823 #ifdef ASSERT
2824 void Compile::check_no_dead_use() const {
2825 ResourceMark rm;
2826 Unique_Node_List wq;
2827 wq.push(root());
2828 for (uint i = 0; i < wq.size(); ++i) {
2829 Node* n = wq.at(i);
2830 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2831 Node* u = n->fast_out(j);
2832 if (u->outcnt() == 0 && !u->is_Con()) {
2833 u->dump();
2834 fatal("no reachable node should have no use");
2835 }
2836 wq.push(u);
2837 }
2838 }
2839 }
2840 #endif
2841
2842 void Compile::inline_vector_reboxing_calls() {
2843 if (C->_vector_reboxing_late_inlines.length() > 0) {
2844 _late_inlines_pos = C->_late_inlines.length();
2845 while (_vector_reboxing_late_inlines.length() > 0) {
2846 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2847 cg->do_late_inline();
2848 if (failing()) return;
2849 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2850 }
2851 _vector_reboxing_late_inlines.trunc_to(0);
2852 }
2853 }
2854
2855 bool Compile::has_vbox_nodes() {
2856 if (C->_vector_reboxing_late_inlines.length() > 0) {
2857 return true;
2858 }
2859 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2860 Node * n = C->macro_node(macro_idx);
2861 assert(n->is_macro(), "only macro nodes expected here");
2862 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2863 return true;
2864 }
2865 }
2866 return false;
2867 }
2868
2869 //---------------------------- Bitwise operation packing optimization ---------------------------
2870
2871 static bool is_vector_unary_bitwise_op(Node* n) {
2872 return n->Opcode() == Op_XorV &&
2873 VectorNode::is_vector_bitwise_not_pattern(n);
2874 }
2875
2876 static bool is_vector_binary_bitwise_op(Node* n) {
2877 switch (n->Opcode()) {
2878 case Op_AndV:
2879 case Op_OrV:
2880 return true;
2881
2882 case Op_XorV:
2883 return !is_vector_unary_bitwise_op(n);
2884
2885 default:
2886 return false;
2887 }
2888 }
2889
2890 static bool is_vector_ternary_bitwise_op(Node* n) {
2891 return n->Opcode() == Op_MacroLogicV;
2892 }
2893
2894 static bool is_vector_bitwise_op(Node* n) {
2895 return is_vector_unary_bitwise_op(n) ||
2896 is_vector_binary_bitwise_op(n) ||
2897 is_vector_ternary_bitwise_op(n);
2898 }
2899
2900 static bool is_vector_bitwise_cone_root(Node* n) {
2901 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2902 return false;
2903 }
2904 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2905 if (is_vector_bitwise_op(n->fast_out(i))) {
2906 return false;
2907 }
2908 }
2909 return true;
2910 }
2911
2912 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2913 uint cnt = 0;
2914 if (is_vector_bitwise_op(n)) {
2915 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2916 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2917 for (uint i = 1; i < inp_cnt; i++) {
2918 Node* in = n->in(i);
2919 bool skip = VectorNode::is_all_ones_vector(in);
2920 if (!skip && !inputs.member(in)) {
2921 inputs.push(in);
2922 cnt++;
2923 }
2924 }
2925 assert(cnt <= 1, "not unary");
2926 } else {
2927 uint last_req = inp_cnt;
2928 if (is_vector_ternary_bitwise_op(n)) {
2929 last_req = inp_cnt - 1; // skip last input
2930 }
2931 for (uint i = 1; i < last_req; i++) {
2932 Node* def = n->in(i);
2933 if (!inputs.member(def)) {
2934 inputs.push(def);
2935 cnt++;
2936 }
2937 }
2938 }
2939 } else { // not a bitwise operations
2940 if (!inputs.member(n)) {
2941 inputs.push(n);
2942 cnt++;
2943 }
2944 }
2945 return cnt;
2946 }
2947
2948 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2949 Unique_Node_List useful_nodes;
2950 C->identify_useful_nodes(useful_nodes);
2951
2952 for (uint i = 0; i < useful_nodes.size(); i++) {
2953 Node* n = useful_nodes.at(i);
2954 if (is_vector_bitwise_cone_root(n)) {
2955 list.push(n);
2956 }
2957 }
2958 }
2959
2960 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2961 const TypeVect* vt,
2962 Unique_Node_List& partition,
2963 Unique_Node_List& inputs) {
2964 assert(partition.size() == 2 || partition.size() == 3, "not supported");
2965 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
2966 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2967
2968 Node* in1 = inputs.at(0);
2969 Node* in2 = inputs.at(1);
2970 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2971
2972 uint func = compute_truth_table(partition, inputs);
2973
2974 Node* pn = partition.at(partition.size() - 1);
2975 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
2976 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2977 }
2978
2979 static uint extract_bit(uint func, uint pos) {
2980 return (func & (1 << pos)) >> pos;
2981 }
2982
2983 //
2984 // A macro logic node represents a truth table. It has 4 inputs,
2985 // First three inputs corresponds to 3 columns of a truth table
2986 // and fourth input captures the logic function.
2987 //
2988 // eg. fn = (in1 AND in2) OR in3;
2989 //
2990 // MacroNode(in1,in2,in3,fn)
2991 //
2992 // -----------------
2993 // in1 in2 in3 fn
2994 // -----------------
2995 // 0 0 0 0
2996 // 0 0 1 1
2997 // 0 1 0 0
2998 // 0 1 1 1
2999 // 1 0 0 0
3000 // 1 0 1 1
3001 // 1 1 0 1
3002 // 1 1 1 1
3003 //
3004
3005 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3006 int res = 0;
3007 for (int i = 0; i < 8; i++) {
3008 int bit1 = extract_bit(in1, i);
3009 int bit2 = extract_bit(in2, i);
3010 int bit3 = extract_bit(in3, i);
3011
3012 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3013 int func_bit = extract_bit(func, func_bit_pos);
3014
3015 res |= func_bit << i;
3016 }
3017 return res;
3018 }
3019
3020 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3021 assert(n != NULL, "");
3022 assert(eval_map.contains(n), "absent");
3023 return *(eval_map.get(n));
3024 }
3025
3026 static void eval_operands(Node* n,
3027 uint& func1, uint& func2, uint& func3,
3028 ResourceHashtable<Node*,uint>& eval_map) {
3029 assert(is_vector_bitwise_op(n), "");
3030
3031 if (is_vector_unary_bitwise_op(n)) {
3032 Node* opnd = n->in(1);
3033 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3034 opnd = n->in(2);
3035 }
3036 func1 = eval_operand(opnd, eval_map);
3037 } else if (is_vector_binary_bitwise_op(n)) {
3038 func1 = eval_operand(n->in(1), eval_map);
3039 func2 = eval_operand(n->in(2), eval_map);
3040 } else {
3041 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3042 func1 = eval_operand(n->in(1), eval_map);
3043 func2 = eval_operand(n->in(2), eval_map);
3044 func3 = eval_operand(n->in(3), eval_map);
3045 }
3046 }
3047
3048 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3049 assert(inputs.size() <= 3, "sanity");
3050 ResourceMark rm;
3051 uint res = 0;
3052 ResourceHashtable<Node*,uint> eval_map;
3053
3054 // Populate precomputed functions for inputs.
3055 // Each input corresponds to one column of 3 input truth-table.
3056 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3057 0xCC, // (_, b, _) -> b
3058 0xF0 }; // (a, _, _) -> a
3059 for (uint i = 0; i < inputs.size(); i++) {
3060 eval_map.put(inputs.at(i), input_funcs[2-i]);
3061 }
3062
3063 for (uint i = 0; i < partition.size(); i++) {
3064 Node* n = partition.at(i);
3065
3066 uint func1 = 0, func2 = 0, func3 = 0;
3067 eval_operands(n, func1, func2, func3, eval_map);
3068
3069 switch (n->Opcode()) {
3070 case Op_OrV:
3071 assert(func3 == 0, "not binary");
3072 res = func1 | func2;
3073 break;
3074 case Op_AndV:
3075 assert(func3 == 0, "not binary");
3076 res = func1 & func2;
3077 break;
3078 case Op_XorV:
3079 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3080 assert(func2 == 0 && func3 == 0, "not unary");
3081 res = (~func1) & 0xFF;
3082 } else {
3083 assert(func3 == 0, "not binary");
3084 res = func1 ^ func2;
3085 }
3086 break;
3087 case Op_MacroLogicV:
3088 // Ordering of inputs may change during evaluation of sub-tree
3089 // containing MacroLogic node as a child node, thus a re-evaluation
3090 // makes sure that function is evaluated in context of current
3091 // inputs.
3092 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3093 break;
3094
3095 default: assert(false, "not supported: %s", n->Name());
3096 }
3097 assert(res <= 0xFF, "invalid");
3098 eval_map.put(n, res);
3099 }
3100 return res;
3101 }
3102
3103 // Criteria under which nodes gets packed into a macro logic node:-
3104 // 1) Parent and both child nodes are all unmasked or masked with
3105 // same predicates.
3106 // 2) Masked parent can be packed with left child if it is predicated
3107 // and both have same predicates.
3108 // 3) Masked parent can be packed with right child if its un-predicated
3109 // or has matching predication condition.
3110 // 4) An unmasked parent can be packed with an unmasked child.
3111 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3112 assert(partition.size() == 0, "not empty");
3113 assert(inputs.size() == 0, "not empty");
3114 if (is_vector_ternary_bitwise_op(n)) {
3115 return false;
3116 }
3117
3118 bool is_unary_op = is_vector_unary_bitwise_op(n);
3119 if (is_unary_op) {
3120 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3121 return false; // too few inputs
3122 }
3123
3124 bool pack_left_child = true;
3125 bool pack_right_child = true;
3126
3127 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3128 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3129
3130 int left_child_input_cnt = 0;
3131 int right_child_input_cnt = 0;
3132
3133 bool parent_is_predicated = n->is_predicated_vector();
3134 bool left_child_predicated = n->in(1)->is_predicated_vector();
3135 bool right_child_predicated = n->in(2)->is_predicated_vector();
3136
3137 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : NULL;
3138 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
3139 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
3140
3141 do {
3142 if (pack_left_child && left_child_LOP &&
3143 ((!parent_is_predicated && !left_child_predicated) ||
3144 ((parent_is_predicated && left_child_predicated &&
3145 parent_pred == left_child_pred)))) {
3146 partition.push(n->in(1));
3147 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3148 } else {
3149 inputs.push(n->in(1));
3150 left_child_input_cnt = 1;
3151 }
3152
3153 if (pack_right_child && right_child_LOP &&
3154 (!right_child_predicated ||
3155 (right_child_predicated && parent_is_predicated &&
3156 parent_pred == right_child_pred))) {
3157 partition.push(n->in(2));
3158 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3159 } else {
3160 inputs.push(n->in(2));
3161 right_child_input_cnt = 1;
3162 }
3163
3164 if (inputs.size() > 3) {
3165 assert(partition.size() > 0, "");
3166 inputs.clear();
3167 partition.clear();
3168 if (left_child_input_cnt > right_child_input_cnt) {
3169 pack_left_child = false;
3170 } else {
3171 pack_right_child = false;
3172 }
3173 } else {
3174 break;
3175 }
3176 } while(true);
3177
3178 if(partition.size()) {
3179 partition.push(n);
3180 }
3181
3182 return (partition.size() == 2 || partition.size() == 3) &&
3183 (inputs.size() == 2 || inputs.size() == 3);
3184 }
3185
3186 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3187 assert(is_vector_bitwise_op(n), "not a root");
3188
3189 visited.set(n->_idx);
3190
3191 // 1) Do a DFS walk over the logic cone.
3192 for (uint i = 1; i < n->req(); i++) {
3193 Node* in = n->in(i);
3194 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3195 process_logic_cone_root(igvn, in, visited);
3196 }
3197 }
3198
3199 // 2) Bottom up traversal: Merge node[s] with
3200 // the parent to form macro logic node.
3201 Unique_Node_List partition;
3202 Unique_Node_List inputs;
3203 if (compute_logic_cone(n, partition, inputs)) {
3204 const TypeVect* vt = n->bottom_type()->is_vect();
3205 Node* pn = partition.at(partition.size() - 1);
3206 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
3207 if (mask == NULL ||
3208 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3209 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3210 #ifdef ASSERT
3211 if (TraceNewVectors) {
3212 tty->print("new Vector node: ");
3213 macro_logic->dump();
3214 }
3215 #endif
3216 igvn.replace_node(n, macro_logic);
3217 }
3218 }
3219 }
3220
3221 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3222 ResourceMark rm;
3223 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3224 Unique_Node_List list;
3225 collect_logic_cone_roots(list);
3226
3227 while (list.size() > 0) {
3228 Node* n = list.pop();
3229 const TypeVect* vt = n->bottom_type()->is_vect();
3230 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3231 if (supported) {
3232 VectorSet visited(comp_arena());
3233 process_logic_cone_root(igvn, n, visited);
3234 }
3235 }
3236 }
3237 }
3238
3239 //------------------------------Code_Gen---------------------------------------
3240 // Given a graph, generate code for it
3241 void Compile::Code_Gen() {
3242 if (failing()) {
3243 return;
3244 }
3245
3246 // Perform instruction selection. You might think we could reclaim Matcher
3247 // memory PDQ, but actually the Matcher is used in generating spill code.
3248 // Internals of the Matcher (including some VectorSets) must remain live
3249 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3250 // set a bit in reclaimed memory.
3251
3252 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3253 // nodes. Mapping is only valid at the root of each matched subtree.
3254 NOT_PRODUCT( verify_graph_edges(); )
3255
3256 Matcher matcher;
3257 _matcher = &matcher;
3258 {
3259 TracePhase tp("matcher", &timers[_t_matcher]);
3260 matcher.match();
3261 if (failing()) {
3262 return;
3263 }
3264 }
3265 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3266 // nodes. Mapping is only valid at the root of each matched subtree.
3267 NOT_PRODUCT( verify_graph_edges(); )
3268
3269 // If you have too many nodes, or if matching has failed, bail out
3270 check_node_count(0, "out of nodes matching instructions");
3271 if (failing()) {
3272 return;
3273 }
3274
3275 print_method(PHASE_MATCHING, 2);
3276
3277 // Build a proper-looking CFG
3278 PhaseCFG cfg(node_arena(), root(), matcher);
3279 _cfg = &cfg;
3280 {
3281 TracePhase tp("scheduler", &timers[_t_scheduler]);
3282 bool success = cfg.do_global_code_motion();
3283 if (!success) {
3284 return;
3285 }
3286
3287 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3288 NOT_PRODUCT( verify_graph_edges(); )
3289 cfg.verify();
3290 }
3291
3292 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3293 _regalloc = ®alloc;
3294 {
3295 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
3296 // Perform register allocation. After Chaitin, use-def chains are
3297 // no longer accurate (at spill code) and so must be ignored.
3298 // Node->LRG->reg mappings are still accurate.
3299 _regalloc->Register_Allocate();
3300
3301 // Bail out if the allocator builds too many nodes
3302 if (failing()) {
3303 return;
3304 }
3305 }
3306
3307 // Prior to register allocation we kept empty basic blocks in case the
3308 // the allocator needed a place to spill. After register allocation we
3309 // are not adding any new instructions. If any basic block is empty, we
3310 // can now safely remove it.
3311 {
3312 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3313 cfg.remove_empty_blocks();
3314 if (do_freq_based_layout()) {
3315 PhaseBlockLayout layout(cfg);
3316 } else {
3317 cfg.set_loop_alignment();
3318 }
3319 cfg.fixup_flow();
3320 }
3321
3322 // Apply peephole optimizations
3323 if( OptoPeephole ) {
3324 TracePhase tp("peephole", &timers[_t_peephole]);
3325 PhasePeephole peep( _regalloc, cfg);
3326 peep.do_transform();
3327 }
3328
3329 // Do late expand if CPU requires this.
3330 if (Matcher::require_postalloc_expand) {
3331 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3332 cfg.postalloc_expand(_regalloc);
3333 }
3334
3335 // Convert Nodes to instruction bits in a buffer
3336 {
3337 TracePhase tp("output", &timers[_t_output]);
3338 PhaseOutput output;
3339 output.Output();
3340 if (failing()) return;
3341 output.install();
3342 }
3343
3344 print_method(PHASE_FINAL_CODE, 1);
3345
3346 // He's dead, Jim.
3347 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3348 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3349 }
3350
3351 //------------------------------Final_Reshape_Counts---------------------------
3352 // This class defines counters to help identify when a method
3353 // may/must be executed using hardware with only 24-bit precision.
3354 struct Final_Reshape_Counts : public StackObj {
3355 int _call_count; // count non-inlined 'common' calls
3356 int _float_count; // count float ops requiring 24-bit precision
3357 int _double_count; // count double ops requiring more precision
3358 int _java_call_count; // count non-inlined 'java' calls
3359 int _inner_loop_count; // count loops which need alignment
3360 VectorSet _visited; // Visitation flags
3361 Node_List _tests; // Set of IfNodes & PCTableNodes
3362
3363 Final_Reshape_Counts() :
3364 _call_count(0), _float_count(0), _double_count(0),
3365 _java_call_count(0), _inner_loop_count(0) { }
3366
3367 void inc_call_count () { _call_count ++; }
3368 void inc_float_count () { _float_count ++; }
3369 void inc_double_count() { _double_count++; }
3370 void inc_java_call_count() { _java_call_count++; }
3371 void inc_inner_loop_count() { _inner_loop_count++; }
3372
3373 int get_call_count () const { return _call_count ; }
3374 int get_float_count () const { return _float_count ; }
3375 int get_double_count() const { return _double_count; }
3376 int get_java_call_count() const { return _java_call_count; }
3377 int get_inner_loop_count() const { return _inner_loop_count; }
3378 };
3379
3380 // Eliminate trivially redundant StoreCMs and accumulate their
3381 // precedence edges.
3382 void Compile::eliminate_redundant_card_marks(Node* n) {
3383 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3384 if (n->in(MemNode::Address)->outcnt() > 1) {
3385 // There are multiple users of the same address so it might be
3386 // possible to eliminate some of the StoreCMs
3387 Node* mem = n->in(MemNode::Memory);
3388 Node* adr = n->in(MemNode::Address);
3389 Node* val = n->in(MemNode::ValueIn);
3390 Node* prev = n;
3391 bool done = false;
3392 // Walk the chain of StoreCMs eliminating ones that match. As
3393 // long as it's a chain of single users then the optimization is
3394 // safe. Eliminating partially redundant StoreCMs would require
3395 // cloning copies down the other paths.
3396 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3397 if (adr == mem->in(MemNode::Address) &&
3398 val == mem->in(MemNode::ValueIn)) {
3399 // redundant StoreCM
3400 if (mem->req() > MemNode::OopStore) {
3401 // Hasn't been processed by this code yet.
3402 n->add_prec(mem->in(MemNode::OopStore));
3403 } else {
3404 // Already converted to precedence edge
3405 for (uint i = mem->req(); i < mem->len(); i++) {
3406 // Accumulate any precedence edges
3407 if (mem->in(i) != NULL) {
3408 n->add_prec(mem->in(i));
3409 }
3410 }
3411 // Everything above this point has been processed.
3412 done = true;
3413 }
3414 // Eliminate the previous StoreCM
3415 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3416 assert(mem->outcnt() == 0, "should be dead");
3417 mem->disconnect_inputs(this);
3418 } else {
3419 prev = mem;
3420 }
3421 mem = prev->in(MemNode::Memory);
3422 }
3423 }
3424 }
3425
3426
3427 //------------------------------final_graph_reshaping_impl----------------------
3428 // Implement items 1-5 from final_graph_reshaping below.
3429 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3430
3431 if ( n->outcnt() == 0 ) return; // dead node
3432 uint nop = n->Opcode();
3433
3434 // Check for 2-input instruction with "last use" on right input.
3435 // Swap to left input. Implements item (2).
3436 if( n->req() == 3 && // two-input instruction
3437 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3438 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3439 n->in(2)->outcnt() == 1 &&// right use IS a last use
3440 !n->in(2)->is_Con() ) { // right use is not a constant
3441 // Check for commutative opcode
3442 switch( nop ) {
3443 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3444 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3445 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3446 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3447 case Op_AndL: case Op_XorL: case Op_OrL:
3448 case Op_AndI: case Op_XorI: case Op_OrI: {
3449 // Move "last use" input to left by swapping inputs
3450 n->swap_edges(1, 2);
3451 break;
3452 }
3453 default:
3454 break;
3455 }
3456 }
3457
3458 #ifdef ASSERT
3459 if( n->is_Mem() ) {
3460 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3461 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3462 // oop will be recorded in oop map if load crosses safepoint
3463 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3464 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3465 "raw memory operations should have control edge");
3466 }
3467 if (n->is_MemBar()) {
3468 MemBarNode* mb = n->as_MemBar();
3469 if (mb->trailing_store() || mb->trailing_load_store()) {
3470 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3471 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3472 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3473 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3474 } else if (mb->leading()) {
3475 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3476 }
3477 }
3478 #endif
3479 // Count FPU ops and common calls, implements item (3)
3480 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3481 if (!gc_handled) {
3482 final_graph_reshaping_main_switch(n, frc, nop);
3483 }
3484
3485 // Collect CFG split points
3486 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3487 frc._tests.push(n);
3488 }
3489 }
3490
3491 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3492 switch( nop ) {
3493 // Count all float operations that may use FPU
3494 case Op_AddF:
3495 case Op_SubF:
3496 case Op_MulF:
3497 case Op_DivF:
3498 case Op_NegF:
3499 case Op_ModF:
3500 case Op_ConvI2F:
3501 case Op_ConF:
3502 case Op_CmpF:
3503 case Op_CmpF3:
3504 case Op_StoreF:
3505 case Op_LoadF:
3506 // case Op_ConvL2F: // longs are split into 32-bit halves
3507 frc.inc_float_count();
3508 break;
3509
3510 case Op_ConvF2D:
3511 case Op_ConvD2F:
3512 frc.inc_float_count();
3513 frc.inc_double_count();
3514 break;
3515
3516 // Count all double operations that may use FPU
3517 case Op_AddD:
3518 case Op_SubD:
3519 case Op_MulD:
3520 case Op_DivD:
3521 case Op_NegD:
3522 case Op_ModD:
3523 case Op_ConvI2D:
3524 case Op_ConvD2I:
3525 // case Op_ConvL2D: // handled by leaf call
3526 // case Op_ConvD2L: // handled by leaf call
3527 case Op_ConD:
3528 case Op_CmpD:
3529 case Op_CmpD3:
3530 case Op_StoreD:
3531 case Op_LoadD:
3532 case Op_LoadD_unaligned:
3533 frc.inc_double_count();
3534 break;
3535 case Op_Opaque1: // Remove Opaque Nodes before matching
3536 case Op_Opaque2: // Remove Opaque Nodes before matching
3537 case Op_Opaque3:
3538 n->subsume_by(n->in(1), this);
3539 break;
3540 case Op_CallStaticJava:
3541 case Op_CallJava:
3542 case Op_CallDynamicJava:
3543 frc.inc_java_call_count(); // Count java call site;
3544 case Op_CallRuntime:
3545 case Op_CallLeaf:
3546 case Op_CallLeafVector:
3547 case Op_CallNative:
3548 case Op_CallLeafNoFP: {
3549 assert (n->is_Call(), "");
3550 CallNode *call = n->as_Call();
3551 // Count call sites where the FP mode bit would have to be flipped.
3552 // Do not count uncommon runtime calls:
3553 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3554 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3555 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3556 frc.inc_call_count(); // Count the call site
3557 } else { // See if uncommon argument is shared
3558 Node *n = call->in(TypeFunc::Parms);
3559 int nop = n->Opcode();
3560 // Clone shared simple arguments to uncommon calls, item (1).
3561 if (n->outcnt() > 1 &&
3562 !n->is_Proj() &&
3563 nop != Op_CreateEx &&
3564 nop != Op_CheckCastPP &&
3565 nop != Op_DecodeN &&
3566 nop != Op_DecodeNKlass &&
3567 !n->is_Mem() &&
3568 !n->is_Phi()) {
3569 Node *x = n->clone();
3570 call->set_req(TypeFunc::Parms, x);
3571 }
3572 }
3573 break;
3574 }
3575
3576 case Op_StoreCM:
3577 {
3578 // Convert OopStore dependence into precedence edge
3579 Node* prec = n->in(MemNode::OopStore);
3580 n->del_req(MemNode::OopStore);
3581 if (prec->is_MergeMem()) {
3582 MergeMemNode* mm = prec->as_MergeMem();
3583 Node* base = mm->base_memory();
3584 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) {
3585 const Type* adr_type = get_adr_type(i);
3586 if (adr_type->isa_aryptr() && adr_type->is_aryptr()->is_flat()) {
3587 Node* m = mm->memory_at(i);
3588 n->add_prec(m);
3589 }
3590 }
3591 if (mm->outcnt() == 0) {
3592 mm->disconnect_inputs(this);
3593 }
3594 } else {
3595 n->add_prec(prec);
3596 }
3597 eliminate_redundant_card_marks(n);
3598 }
3599
3600 // fall through
3601
3602 case Op_StoreB:
3603 case Op_StoreC:
3604 case Op_StorePConditional:
3605 case Op_StoreI:
3606 case Op_StoreL:
3607 case Op_StoreIConditional:
3608 case Op_StoreLConditional:
3609 case Op_CompareAndSwapB:
3610 case Op_CompareAndSwapS:
3611 case Op_CompareAndSwapI:
3612 case Op_CompareAndSwapL:
3613 case Op_CompareAndSwapP:
3614 case Op_CompareAndSwapN:
3615 case Op_WeakCompareAndSwapB:
3616 case Op_WeakCompareAndSwapS:
3617 case Op_WeakCompareAndSwapI:
3618 case Op_WeakCompareAndSwapL:
3619 case Op_WeakCompareAndSwapP:
3620 case Op_WeakCompareAndSwapN:
3621 case Op_CompareAndExchangeB:
3622 case Op_CompareAndExchangeS:
3623 case Op_CompareAndExchangeI:
3624 case Op_CompareAndExchangeL:
3625 case Op_CompareAndExchangeP:
3626 case Op_CompareAndExchangeN:
3627 case Op_GetAndAddS:
3628 case Op_GetAndAddB:
3629 case Op_GetAndAddI:
3630 case Op_GetAndAddL:
3631 case Op_GetAndSetS:
3632 case Op_GetAndSetB:
3633 case Op_GetAndSetI:
3634 case Op_GetAndSetL:
3635 case Op_GetAndSetP:
3636 case Op_GetAndSetN:
3637 case Op_StoreP:
3638 case Op_StoreN:
3639 case Op_StoreNKlass:
3640 case Op_LoadB:
3641 case Op_LoadUB:
3642 case Op_LoadUS:
3643 case Op_LoadI:
3644 case Op_LoadKlass:
3645 case Op_LoadNKlass:
3646 case Op_LoadL:
3647 case Op_LoadL_unaligned:
3648 case Op_LoadPLocked:
3649 case Op_LoadP:
3650 case Op_LoadN:
3651 case Op_LoadRange:
3652 case Op_LoadS:
3653 break;
3654
3655 case Op_AddP: { // Assert sane base pointers
3656 Node *addp = n->in(AddPNode::Address);
3657 assert( !addp->is_AddP() ||
3658 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3659 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3660 "Base pointers must match (addp %u)", addp->_idx );
3661 #ifdef _LP64
3662 if ((UseCompressedOops || UseCompressedClassPointers) &&
3663 addp->Opcode() == Op_ConP &&
3664 addp == n->in(AddPNode::Base) &&
3665 n->in(AddPNode::Offset)->is_Con()) {
3666 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3667 // on the platform and on the compressed oops mode.
3668 // Use addressing with narrow klass to load with offset on x86.
3669 // Some platforms can use the constant pool to load ConP.
3670 // Do this transformation here since IGVN will convert ConN back to ConP.
3671 const Type* t = addp->bottom_type();
3672 bool is_oop = t->isa_oopptr() != NULL;
3673 bool is_klass = t->isa_klassptr() != NULL;
3674
3675 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3676 (is_klass && Matcher::const_klass_prefer_decode())) {
3677 Node* nn = NULL;
3678
3679 int op = is_oop ? Op_ConN : Op_ConNKlass;
3680
3681 // Look for existing ConN node of the same exact type.
3682 Node* r = root();
3683 uint cnt = r->outcnt();
3684 for (uint i = 0; i < cnt; i++) {
3685 Node* m = r->raw_out(i);
3686 if (m!= NULL && m->Opcode() == op &&
3687 m->bottom_type()->make_ptr() == t) {
3688 nn = m;
3689 break;
3690 }
3691 }
3692 if (nn != NULL) {
3693 // Decode a narrow oop to match address
3694 // [R12 + narrow_oop_reg<<3 + offset]
3695 if (is_oop) {
3696 nn = new DecodeNNode(nn, t);
3697 } else {
3698 nn = new DecodeNKlassNode(nn, t);
3699 }
3700 // Check for succeeding AddP which uses the same Base.
3701 // Otherwise we will run into the assertion above when visiting that guy.
3702 for (uint i = 0; i < n->outcnt(); ++i) {
3703 Node *out_i = n->raw_out(i);
3704 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3705 out_i->set_req(AddPNode::Base, nn);
3706 #ifdef ASSERT
3707 for (uint j = 0; j < out_i->outcnt(); ++j) {
3708 Node *out_j = out_i->raw_out(j);
3709 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3710 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3711 }
3712 #endif
3713 }
3714 }
3715 n->set_req(AddPNode::Base, nn);
3716 n->set_req(AddPNode::Address, nn);
3717 if (addp->outcnt() == 0) {
3718 addp->disconnect_inputs(this);
3719 }
3720 }
3721 }
3722 }
3723 #endif
3724 break;
3725 }
3726
3727 case Op_CastPP: {
3728 // Remove CastPP nodes to gain more freedom during scheduling but
3729 // keep the dependency they encode as control or precedence edges
3730 // (if control is set already) on memory operations. Some CastPP
3731 // nodes don't have a control (don't carry a dependency): skip
3732 // those.
3733 if (n->in(0) != NULL) {
3734 ResourceMark rm;
3735 Unique_Node_List wq;
3736 wq.push(n);
3737 for (uint next = 0; next < wq.size(); ++next) {
3738 Node *m = wq.at(next);
3739 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3740 Node* use = m->fast_out(i);
3741 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3742 use->ensure_control_or_add_prec(n->in(0));
3743 } else {
3744 switch(use->Opcode()) {
3745 case Op_AddP:
3746 case Op_DecodeN:
3747 case Op_DecodeNKlass:
3748 case Op_CheckCastPP:
3749 case Op_CastPP:
3750 wq.push(use);
3751 break;
3752 }
3753 }
3754 }
3755 }
3756 }
3757 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3758 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3759 Node* in1 = n->in(1);
3760 const Type* t = n->bottom_type();
3761 Node* new_in1 = in1->clone();
3762 new_in1->as_DecodeN()->set_type(t);
3763
3764 if (!Matcher::narrow_oop_use_complex_address()) {
3765 //
3766 // x86, ARM and friends can handle 2 adds in addressing mode
3767 // and Matcher can fold a DecodeN node into address by using
3768 // a narrow oop directly and do implicit NULL check in address:
3769 //
3770 // [R12 + narrow_oop_reg<<3 + offset]
3771 // NullCheck narrow_oop_reg
3772 //
3773 // On other platforms (Sparc) we have to keep new DecodeN node and
3774 // use it to do implicit NULL check in address:
3775 //
3776 // decode_not_null narrow_oop_reg, base_reg
3777 // [base_reg + offset]
3778 // NullCheck base_reg
3779 //
3780 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3781 // to keep the information to which NULL check the new DecodeN node
3782 // corresponds to use it as value in implicit_null_check().
3783 //
3784 new_in1->set_req(0, n->in(0));
3785 }
3786
3787 n->subsume_by(new_in1, this);
3788 if (in1->outcnt() == 0) {
3789 in1->disconnect_inputs(this);
3790 }
3791 } else {
3792 n->subsume_by(n->in(1), this);
3793 if (n->outcnt() == 0) {
3794 n->disconnect_inputs(this);
3795 }
3796 }
3797 break;
3798 }
3799 #ifdef _LP64
3800 case Op_CmpP:
3801 // Do this transformation here to preserve CmpPNode::sub() and
3802 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3803 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3804 Node* in1 = n->in(1);
3805 Node* in2 = n->in(2);
3806 if (!in1->is_DecodeNarrowPtr()) {
3807 in2 = in1;
3808 in1 = n->in(2);
3809 }
3810 assert(in1->is_DecodeNarrowPtr(), "sanity");
3811
3812 Node* new_in2 = NULL;
3813 if (in2->is_DecodeNarrowPtr()) {
3814 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3815 new_in2 = in2->in(1);
3816 } else if (in2->Opcode() == Op_ConP) {
3817 const Type* t = in2->bottom_type();
3818 if (t == TypePtr::NULL_PTR) {
3819 assert(in1->is_DecodeN(), "compare klass to null?");
3820 // Don't convert CmpP null check into CmpN if compressed
3821 // oops implicit null check is not generated.
3822 // This will allow to generate normal oop implicit null check.
3823 if (Matcher::gen_narrow_oop_implicit_null_checks())
3824 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3825 //
3826 // This transformation together with CastPP transformation above
3827 // will generated code for implicit NULL checks for compressed oops.
3828 //
3829 // The original code after Optimize()
3830 //
3831 // LoadN memory, narrow_oop_reg
3832 // decode narrow_oop_reg, base_reg
3833 // CmpP base_reg, NULL
3834 // CastPP base_reg // NotNull
3835 // Load [base_reg + offset], val_reg
3836 //
3837 // after these transformations will be
3838 //
3839 // LoadN memory, narrow_oop_reg
3840 // CmpN narrow_oop_reg, NULL
3841 // decode_not_null narrow_oop_reg, base_reg
3842 // Load [base_reg + offset], val_reg
3843 //
3844 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3845 // since narrow oops can be used in debug info now (see the code in
3846 // final_graph_reshaping_walk()).
3847 //
3848 // At the end the code will be matched to
3849 // on x86:
3850 //
3851 // Load_narrow_oop memory, narrow_oop_reg
3852 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3853 // NullCheck narrow_oop_reg
3854 //
3855 // and on sparc:
3856 //
3857 // Load_narrow_oop memory, narrow_oop_reg
3858 // decode_not_null narrow_oop_reg, base_reg
3859 // Load [base_reg + offset], val_reg
3860 // NullCheck base_reg
3861 //
3862 } else if (t->isa_oopptr()) {
3863 new_in2 = ConNode::make(t->make_narrowoop());
3864 } else if (t->isa_klassptr()) {
3865 new_in2 = ConNode::make(t->make_narrowklass());
3866 }
3867 }
3868 if (new_in2 != NULL) {
3869 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3870 n->subsume_by(cmpN, this);
3871 if (in1->outcnt() == 0) {
3872 in1->disconnect_inputs(this);
3873 }
3874 if (in2->outcnt() == 0) {
3875 in2->disconnect_inputs(this);
3876 }
3877 }
3878 }
3879 break;
3880
3881 case Op_DecodeN:
3882 case Op_DecodeNKlass:
3883 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3884 // DecodeN could be pinned when it can't be fold into
3885 // an address expression, see the code for Op_CastPP above.
3886 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3887 break;
3888
3889 case Op_EncodeP:
3890 case Op_EncodePKlass: {
3891 Node* in1 = n->in(1);
3892 if (in1->is_DecodeNarrowPtr()) {
3893 n->subsume_by(in1->in(1), this);
3894 } else if (in1->Opcode() == Op_ConP) {
3895 const Type* t = in1->bottom_type();
3896 if (t == TypePtr::NULL_PTR) {
3897 assert(t->isa_oopptr(), "null klass?");
3898 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3899 } else if (t->isa_oopptr()) {
3900 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3901 } else if (t->isa_klassptr()) {
3902 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3903 }
3904 }
3905 if (in1->outcnt() == 0) {
3906 in1->disconnect_inputs(this);
3907 }
3908 break;
3909 }
3910
3911 case Op_Proj: {
3912 if (OptimizeStringConcat || IncrementalInline) {
3913 ProjNode* proj = n->as_Proj();
3914 if (proj->_is_io_use) {
3915 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3916 // Separate projections were used for the exception path which
3917 // are normally removed by a late inline. If it wasn't inlined
3918 // then they will hang around and should just be replaced with
3919 // the original one. Merge them.
3920 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3921 if (non_io_proj != NULL) {
3922 proj->subsume_by(non_io_proj , this);
3923 }
3924 }
3925 }
3926 break;
3927 }
3928
3929 case Op_Phi:
3930 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3931 // The EncodeP optimization may create Phi with the same edges
3932 // for all paths. It is not handled well by Register Allocator.
3933 Node* unique_in = n->in(1);
3934 assert(unique_in != NULL, "");
3935 uint cnt = n->req();
3936 for (uint i = 2; i < cnt; i++) {
3937 Node* m = n->in(i);
3938 assert(m != NULL, "");
3939 if (unique_in != m)
3940 unique_in = NULL;
3941 }
3942 if (unique_in != NULL) {
3943 n->subsume_by(unique_in, this);
3944 }
3945 }
3946 break;
3947
3948 #endif
3949
3950 #ifdef ASSERT
3951 case Op_CastII:
3952 // Verify that all range check dependent CastII nodes were removed.
3953 if (n->isa_CastII()->has_range_check()) {
3954 n->dump(3);
3955 assert(false, "Range check dependent CastII node was not removed");
3956 }
3957 break;
3958 #endif
3959
3960 case Op_ModI:
3961 if (UseDivMod) {
3962 // Check if a%b and a/b both exist
3963 Node* d = n->find_similar(Op_DivI);
3964 if (d) {
3965 // Replace them with a fused divmod if supported
3966 if (Matcher::has_match_rule(Op_DivModI)) {
3967 DivModINode* divmod = DivModINode::make(n);
3968 d->subsume_by(divmod->div_proj(), this);
3969 n->subsume_by(divmod->mod_proj(), this);
3970 } else {
3971 // replace a%b with a-((a/b)*b)
3972 Node* mult = new MulINode(d, d->in(2));
3973 Node* sub = new SubINode(d->in(1), mult);
3974 n->subsume_by(sub, this);
3975 }
3976 }
3977 }
3978 break;
3979
3980 case Op_ModL:
3981 if (UseDivMod) {
3982 // Check if a%b and a/b both exist
3983 Node* d = n->find_similar(Op_DivL);
3984 if (d) {
3985 // Replace them with a fused divmod if supported
3986 if (Matcher::has_match_rule(Op_DivModL)) {
3987 DivModLNode* divmod = DivModLNode::make(n);
3988 d->subsume_by(divmod->div_proj(), this);
3989 n->subsume_by(divmod->mod_proj(), this);
3990 } else {
3991 // replace a%b with a-((a/b)*b)
3992 Node* mult = new MulLNode(d, d->in(2));
3993 Node* sub = new SubLNode(d->in(1), mult);
3994 n->subsume_by(sub, this);
3995 }
3996 }
3997 }
3998 break;
3999
4000 case Op_LoadVector:
4001 case Op_StoreVector:
4002 case Op_LoadVectorGather:
4003 case Op_StoreVectorScatter:
4004 case Op_LoadVectorGatherMasked:
4005 case Op_StoreVectorScatterMasked:
4006 case Op_VectorCmpMasked:
4007 case Op_VectorMaskGen:
4008 case Op_LoadVectorMasked:
4009 case Op_StoreVectorMasked:
4010 break;
4011
4012 case Op_AddReductionVI:
4013 case Op_AddReductionVL:
4014 case Op_AddReductionVF:
4015 case Op_AddReductionVD:
4016 case Op_MulReductionVI:
4017 case Op_MulReductionVL:
4018 case Op_MulReductionVF:
4019 case Op_MulReductionVD:
4020 case Op_MinReductionV:
4021 case Op_MaxReductionV:
4022 case Op_AndReductionV:
4023 case Op_OrReductionV:
4024 case Op_XorReductionV:
4025 break;
4026
4027 case Op_PackB:
4028 case Op_PackS:
4029 case Op_PackI:
4030 case Op_PackF:
4031 case Op_PackL:
4032 case Op_PackD:
4033 if (n->req()-1 > 2) {
4034 // Replace many operand PackNodes with a binary tree for matching
4035 PackNode* p = (PackNode*) n;
4036 Node* btp = p->binary_tree_pack(1, n->req());
4037 n->subsume_by(btp, this);
4038 }
4039 break;
4040 case Op_Loop:
4041 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4042 case Op_CountedLoop:
4043 case Op_LongCountedLoop:
4044 case Op_OuterStripMinedLoop:
4045 if (n->as_Loop()->is_inner_loop()) {
4046 frc.inc_inner_loop_count();
4047 }
4048 n->as_Loop()->verify_strip_mined(0);
4049 break;
4050 case Op_LShiftI:
4051 case Op_RShiftI:
4052 case Op_URShiftI:
4053 case Op_LShiftL:
4054 case Op_RShiftL:
4055 case Op_URShiftL:
4056 if (Matcher::need_masked_shift_count) {
4057 // The cpu's shift instructions don't restrict the count to the
4058 // lower 5/6 bits. We need to do the masking ourselves.
4059 Node* in2 = n->in(2);
4060 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4061 const TypeInt* t = in2->find_int_type();
4062 if (t != NULL && t->is_con()) {
4063 juint shift = t->get_con();
4064 if (shift > mask) { // Unsigned cmp
4065 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4066 }
4067 } else {
4068 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
4069 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4070 n->set_req(2, shift);
4071 }
4072 }
4073 if (in2->outcnt() == 0) { // Remove dead node
4074 in2->disconnect_inputs(this);
4075 }
4076 }
4077 break;
4078 case Op_MemBarStoreStore:
4079 case Op_MemBarRelease:
4080 // Break the link with AllocateNode: it is no longer useful and
4081 // confuses register allocation.
4082 if (n->req() > MemBarNode::Precedent) {
4083 n->set_req(MemBarNode::Precedent, top());
4084 }
4085 break;
4086 case Op_MemBarAcquire: {
4087 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4088 // At parse time, the trailing MemBarAcquire for a volatile load
4089 // is created with an edge to the load. After optimizations,
4090 // that input may be a chain of Phis. If those phis have no
4091 // other use, then the MemBarAcquire keeps them alive and
4092 // register allocation can be confused.
4093 ResourceMark rm;
4094 Unique_Node_List wq;
4095 wq.push(n->in(MemBarNode::Precedent));
4096 n->set_req(MemBarNode::Precedent, top());
4097 while (wq.size() > 0) {
4098 Node* m = wq.pop();
4099 if (m->outcnt() == 0) {
4100 for (uint j = 0; j < m->req(); j++) {
4101 Node* in = m->in(j);
4102 if (in != NULL) {
4103 wq.push(in);
4104 }
4105 }
4106 m->disconnect_inputs(this);
4107 }
4108 }
4109 }
4110 break;
4111 }
4112 case Op_Blackhole:
4113 break;
4114 case Op_RangeCheck: {
4115 RangeCheckNode* rc = n->as_RangeCheck();
4116 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4117 n->subsume_by(iff, this);
4118 frc._tests.push(iff);
4119 break;
4120 }
4121 case Op_ConvI2L: {
4122 if (!Matcher::convi2l_type_required) {
4123 // Code generation on some platforms doesn't need accurate
4124 // ConvI2L types. Widening the type can help remove redundant
4125 // address computations.
4126 n->as_Type()->set_type(TypeLong::INT);
4127 ResourceMark rm;
4128 Unique_Node_List wq;
4129 wq.push(n);
4130 for (uint next = 0; next < wq.size(); next++) {
4131 Node *m = wq.at(next);
4132
4133 for(;;) {
4134 // Loop over all nodes with identical inputs edges as m
4135 Node* k = m->find_similar(m->Opcode());
4136 if (k == NULL) {
4137 break;
4138 }
4139 // Push their uses so we get a chance to remove node made
4140 // redundant
4141 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4142 Node* u = k->fast_out(i);
4143 if (u->Opcode() == Op_LShiftL ||
4144 u->Opcode() == Op_AddL ||
4145 u->Opcode() == Op_SubL ||
4146 u->Opcode() == Op_AddP) {
4147 wq.push(u);
4148 }
4149 }
4150 // Replace all nodes with identical edges as m with m
4151 k->subsume_by(m, this);
4152 }
4153 }
4154 }
4155 break;
4156 }
4157 case Op_CmpUL: {
4158 if (!Matcher::has_match_rule(Op_CmpUL)) {
4159 // No support for unsigned long comparisons
4160 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4161 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4162 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4163 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4164 Node* andl = new AndLNode(orl, remove_sign_mask);
4165 Node* cmp = new CmpLNode(andl, n->in(2));
4166 n->subsume_by(cmp, this);
4167 }
4168 break;
4169 }
4170 #ifdef ASSERT
4171 case Op_InlineTypePtr:
4172 case Op_InlineType: {
4173 n->dump(-1);
4174 assert(false, "inline type node was not removed");
4175 break;
4176 }
4177 #endif
4178 default:
4179 assert(!n->is_Call(), "");
4180 assert(!n->is_Mem(), "");
4181 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4182 break;
4183 }
4184 }
4185
4186 //------------------------------final_graph_reshaping_walk---------------------
4187 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4188 // requires that the walk visits a node's inputs before visiting the node.
4189 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4190 Unique_Node_List sfpt;
4191
4192 frc._visited.set(root->_idx); // first, mark node as visited
4193 uint cnt = root->req();
4194 Node *n = root;
4195 uint i = 0;
4196 while (true) {
4197 if (i < cnt) {
4198 // Place all non-visited non-null inputs onto stack
4199 Node* m = n->in(i);
4200 ++i;
4201 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4202 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4203 // compute worst case interpreter size in case of a deoptimization
4204 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4205
4206 sfpt.push(m);
4207 }
4208 cnt = m->req();
4209 nstack.push(n, i); // put on stack parent and next input's index
4210 n = m;
4211 i = 0;
4212 }
4213 } else {
4214 // Now do post-visit work
4215 final_graph_reshaping_impl( n, frc );
4216 if (nstack.is_empty())
4217 break; // finished
4218 n = nstack.node(); // Get node from stack
4219 cnt = n->req();
4220 i = nstack.index();
4221 nstack.pop(); // Shift to the next node on stack
4222 }
4223 }
4224
4225 // Skip next transformation if compressed oops are not used.
4226 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4227 (!UseCompressedOops && !UseCompressedClassPointers))
4228 return;
4229
4230 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4231 // It could be done for an uncommon traps or any safepoints/calls
4232 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4233 while (sfpt.size() > 0) {
4234 n = sfpt.pop();
4235 JVMState *jvms = n->as_SafePoint()->jvms();
4236 assert(jvms != NULL, "sanity");
4237 int start = jvms->debug_start();
4238 int end = n->req();
4239 bool is_uncommon = (n->is_CallStaticJava() &&
4240 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4241 for (int j = start; j < end; j++) {
4242 Node* in = n->in(j);
4243 if (in->is_DecodeNarrowPtr()) {
4244 bool safe_to_skip = true;
4245 if (!is_uncommon ) {
4246 // Is it safe to skip?
4247 for (uint i = 0; i < in->outcnt(); i++) {
4248 Node* u = in->raw_out(i);
4249 if (!u->is_SafePoint() ||
4250 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4251 safe_to_skip = false;
4252 }
4253 }
4254 }
4255 if (safe_to_skip) {
4256 n->set_req(j, in->in(1));
4257 }
4258 if (in->outcnt() == 0) {
4259 in->disconnect_inputs(this);
4260 }
4261 }
4262 }
4263 }
4264 }
4265
4266 //------------------------------final_graph_reshaping--------------------------
4267 // Final Graph Reshaping.
4268 //
4269 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4270 // and not commoned up and forced early. Must come after regular
4271 // optimizations to avoid GVN undoing the cloning. Clone constant
4272 // inputs to Loop Phis; these will be split by the allocator anyways.
4273 // Remove Opaque nodes.
4274 // (2) Move last-uses by commutative operations to the left input to encourage
4275 // Intel update-in-place two-address operations and better register usage
4276 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4277 // calls canonicalizing them back.
4278 // (3) Count the number of double-precision FP ops, single-precision FP ops
4279 // and call sites. On Intel, we can get correct rounding either by
4280 // forcing singles to memory (requires extra stores and loads after each
4281 // FP bytecode) or we can set a rounding mode bit (requires setting and
4282 // clearing the mode bit around call sites). The mode bit is only used
4283 // if the relative frequency of single FP ops to calls is low enough.
4284 // This is a key transform for SPEC mpeg_audio.
4285 // (4) Detect infinite loops; blobs of code reachable from above but not
4286 // below. Several of the Code_Gen algorithms fail on such code shapes,
4287 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4288 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4289 // Detection is by looking for IfNodes where only 1 projection is
4290 // reachable from below or CatchNodes missing some targets.
4291 // (5) Assert for insane oop offsets in debug mode.
4292
4293 bool Compile::final_graph_reshaping() {
4294 // an infinite loop may have been eliminated by the optimizer,
4295 // in which case the graph will be empty.
4296 if (root()->req() == 1) {
4297 record_method_not_compilable("trivial infinite loop");
4298 return true;
4299 }
4300
4301 // Expensive nodes have their control input set to prevent the GVN
4302 // from freely commoning them. There's no GVN beyond this point so
4303 // no need to keep the control input. We want the expensive nodes to
4304 // be freely moved to the least frequent code path by gcm.
4305 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4306 for (int i = 0; i < expensive_count(); i++) {
4307 _expensive_nodes.at(i)->set_req(0, NULL);
4308 }
4309
4310 Final_Reshape_Counts frc;
4311
4312 // Visit everybody reachable!
4313 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4314 Node_Stack nstack(live_nodes() >> 1);
4315 final_graph_reshaping_walk(nstack, root(), frc);
4316
4317 // Check for unreachable (from below) code (i.e., infinite loops).
4318 for( uint i = 0; i < frc._tests.size(); i++ ) {
4319 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4320 // Get number of CFG targets.
4321 // Note that PCTables include exception targets after calls.
4322 uint required_outcnt = n->required_outcnt();
4323 if (n->outcnt() != required_outcnt) {
4324 // Check for a few special cases. Rethrow Nodes never take the
4325 // 'fall-thru' path, so expected kids is 1 less.
4326 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4327 if (n->in(0)->in(0)->is_Call()) {
4328 CallNode* call = n->in(0)->in(0)->as_Call();
4329 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4330 required_outcnt--; // Rethrow always has 1 less kid
4331 } else if (call->req() > TypeFunc::Parms &&
4332 call->is_CallDynamicJava()) {
4333 // Check for null receiver. In such case, the optimizer has
4334 // detected that the virtual call will always result in a null
4335 // pointer exception. The fall-through projection of this CatchNode
4336 // will not be populated.
4337 Node* arg0 = call->in(TypeFunc::Parms);
4338 if (arg0->is_Type() &&
4339 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4340 required_outcnt--;
4341 }
4342 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4343 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4344 // Check for illegal array length. In such case, the optimizer has
4345 // detected that the allocation attempt will always result in an
4346 // exception. There is no fall-through projection of this CatchNode .
4347 assert(call->is_CallStaticJava(), "static call expected");
4348 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4349 Node* valid_length_test = call->in(call->req()-1);
4350 call->del_req(call->req()-1);
4351 if (valid_length_test->find_int_con(1) == 0) {
4352 required_outcnt--;
4353 }
4354 assert(n->outcnt() == required_outcnt, "malformed control flow");
4355 continue;
4356 }
4357 }
4358 }
4359 // Recheck with a better notion of 'required_outcnt'
4360 if (n->outcnt() != required_outcnt) {
4361 record_method_not_compilable("malformed control flow");
4362 return true; // Not all targets reachable!
4363 }
4364 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4365 CallNode* call = n->in(0)->in(0)->as_Call();
4366 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4367 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4368 assert(call->is_CallStaticJava(), "static call expected");
4369 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4370 call->del_req(call->req()-1); // valid length test useless now
4371 }
4372 }
4373 // Check that I actually visited all kids. Unreached kids
4374 // must be infinite loops.
4375 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4376 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4377 record_method_not_compilable("infinite loop");
4378 return true; // Found unvisited kid; must be unreach
4379 }
4380
4381 // Here so verification code in final_graph_reshaping_walk()
4382 // always see an OuterStripMinedLoopEnd
4383 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4384 IfNode* init_iff = n->as_If();
4385 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4386 n->subsume_by(iff, this);
4387 }
4388 }
4389
4390 #ifdef IA32
4391 // If original bytecodes contained a mixture of floats and doubles
4392 // check if the optimizer has made it homogenous, item (3).
4393 if (UseSSE == 0 &&
4394 frc.get_float_count() > 32 &&
4395 frc.get_double_count() == 0 &&
4396 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4397 set_24_bit_selection_and_mode(false, true);
4398 }
4399 #endif // IA32
4400
4401 set_java_calls(frc.get_java_call_count());
4402 set_inner_loops(frc.get_inner_loop_count());
4403
4404 // No infinite loops, no reason to bail out.
4405 return false;
4406 }
4407
4408 //-----------------------------too_many_traps----------------------------------
4409 // Report if there are too many traps at the current method and bci.
4410 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4411 bool Compile::too_many_traps(ciMethod* method,
4412 int bci,
4413 Deoptimization::DeoptReason reason) {
4414 ciMethodData* md = method->method_data();
4415 if (md->is_empty()) {
4416 // Assume the trap has not occurred, or that it occurred only
4417 // because of a transient condition during start-up in the interpreter.
4418 return false;
4419 }
4420 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4421 if (md->has_trap_at(bci, m, reason) != 0) {
4422 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4423 // Also, if there are multiple reasons, or if there is no per-BCI record,
4424 // assume the worst.
4425 if (log())
4426 log()->elem("observe trap='%s' count='%d'",
4427 Deoptimization::trap_reason_name(reason),
4428 md->trap_count(reason));
4429 return true;
4430 } else {
4431 // Ignore method/bci and see if there have been too many globally.
4432 return too_many_traps(reason, md);
4433 }
4434 }
4435
4436 // Less-accurate variant which does not require a method and bci.
4437 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4438 ciMethodData* logmd) {
4439 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4440 // Too many traps globally.
4441 // Note that we use cumulative trap_count, not just md->trap_count.
4442 if (log()) {
4443 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4444 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4445 Deoptimization::trap_reason_name(reason),
4446 mcount, trap_count(reason));
4447 }
4448 return true;
4449 } else {
4450 // The coast is clear.
4451 return false;
4452 }
4453 }
4454
4455 //--------------------------too_many_recompiles--------------------------------
4456 // Report if there are too many recompiles at the current method and bci.
4457 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4458 // Is not eager to return true, since this will cause the compiler to use
4459 // Action_none for a trap point, to avoid too many recompilations.
4460 bool Compile::too_many_recompiles(ciMethod* method,
4461 int bci,
4462 Deoptimization::DeoptReason reason) {
4463 ciMethodData* md = method->method_data();
4464 if (md->is_empty()) {
4465 // Assume the trap has not occurred, or that it occurred only
4466 // because of a transient condition during start-up in the interpreter.
4467 return false;
4468 }
4469 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4470 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4471 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4472 Deoptimization::DeoptReason per_bc_reason
4473 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4474 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4475 if ((per_bc_reason == Deoptimization::Reason_none
4476 || md->has_trap_at(bci, m, reason) != 0)
4477 // The trap frequency measure we care about is the recompile count:
4478 && md->trap_recompiled_at(bci, m)
4479 && md->overflow_recompile_count() >= bc_cutoff) {
4480 // Do not emit a trap here if it has already caused recompilations.
4481 // Also, if there are multiple reasons, or if there is no per-BCI record,
4482 // assume the worst.
4483 if (log())
4484 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4485 Deoptimization::trap_reason_name(reason),
4486 md->trap_count(reason),
4487 md->overflow_recompile_count());
4488 return true;
4489 } else if (trap_count(reason) != 0
4490 && decompile_count() >= m_cutoff) {
4491 // Too many recompiles globally, and we have seen this sort of trap.
4492 // Use cumulative decompile_count, not just md->decompile_count.
4493 if (log())
4494 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4495 Deoptimization::trap_reason_name(reason),
4496 md->trap_count(reason), trap_count(reason),
4497 md->decompile_count(), decompile_count());
4498 return true;
4499 } else {
4500 // The coast is clear.
4501 return false;
4502 }
4503 }
4504
4505 // Compute when not to trap. Used by matching trap based nodes and
4506 // NullCheck optimization.
4507 void Compile::set_allowed_deopt_reasons() {
4508 _allowed_reasons = 0;
4509 if (is_method_compilation()) {
4510 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4511 assert(rs < BitsPerInt, "recode bit map");
4512 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4513 _allowed_reasons |= nth_bit(rs);
4514 }
4515 }
4516 }
4517 }
4518
4519 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4520 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4521 }
4522
4523 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4524 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4525 }
4526
4527 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4528 if (holder->is_initialized()) {
4529 return false;
4530 }
4531 if (holder->is_being_initialized()) {
4532 if (accessing_method->holder() == holder) {
4533 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4534 // <init>, or a static method. In all those cases, there was an initialization
4535 // barrier on the holder klass passed.
4536 if (accessing_method->is_class_initializer() ||
4537 accessing_method->is_object_constructor() ||
4538 accessing_method->is_static()) {
4539 return false;
4540 }
4541 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4542 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4543 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4544 // child class can become fully initialized while its parent class is still being initialized.
4545 if (accessing_method->is_class_initializer()) {
4546 return false;
4547 }
4548 }
4549 ciMethod* root = method(); // the root method of compilation
4550 if (root != accessing_method) {
4551 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4552 }
4553 }
4554 return true;
4555 }
4556
4557 #ifndef PRODUCT
4558 //------------------------------verify_graph_edges---------------------------
4559 // Walk the Graph and verify that there is a one-to-one correspondence
4560 // between Use-Def edges and Def-Use edges in the graph.
4561 void Compile::verify_graph_edges(bool no_dead_code) {
4562 if (VerifyGraphEdges) {
4563 Unique_Node_List visited;
4564 // Call recursive graph walk to check edges
4565 _root->verify_edges(visited);
4566 if (no_dead_code) {
4567 // Now make sure that no visited node is used by an unvisited node.
4568 bool dead_nodes = false;
4569 Unique_Node_List checked;
4570 while (visited.size() > 0) {
4571 Node* n = visited.pop();
4572 checked.push(n);
4573 for (uint i = 0; i < n->outcnt(); i++) {
4574 Node* use = n->raw_out(i);
4575 if (checked.member(use)) continue; // already checked
4576 if (visited.member(use)) continue; // already in the graph
4577 if (use->is_Con()) continue; // a dead ConNode is OK
4578 // At this point, we have found a dead node which is DU-reachable.
4579 if (!dead_nodes) {
4580 tty->print_cr("*** Dead nodes reachable via DU edges:");
4581 dead_nodes = true;
4582 }
4583 use->dump(2);
4584 tty->print_cr("---");
4585 checked.push(use); // No repeats; pretend it is now checked.
4586 }
4587 }
4588 assert(!dead_nodes, "using nodes must be reachable from root");
4589 }
4590 }
4591 }
4592 #endif
4593
4594 // The Compile object keeps track of failure reasons separately from the ciEnv.
4595 // This is required because there is not quite a 1-1 relation between the
4596 // ciEnv and its compilation task and the Compile object. Note that one
4597 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4598 // to backtrack and retry without subsuming loads. Other than this backtracking
4599 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4600 // by the logic in C2Compiler.
4601 void Compile::record_failure(const char* reason) {
4602 if (log() != NULL) {
4603 log()->elem("failure reason='%s' phase='compile'", reason);
4604 }
4605 if (_failure_reason == NULL) {
4606 // Record the first failure reason.
4607 _failure_reason = reason;
4608 }
4609
4610 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4611 C->print_method(PHASE_FAILURE, 1);
4612 }
4613 _root = NULL; // flush the graph, too
4614 }
4615
4616 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4617 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4618 _phase_name(name), _dolog(CITimeVerbose)
4619 {
4620 if (_dolog) {
4621 C = Compile::current();
4622 _log = C->log();
4623 } else {
4624 C = NULL;
4625 _log = NULL;
4626 }
4627 if (_log != NULL) {
4628 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4629 _log->stamp();
4630 _log->end_head();
4631 }
4632 }
4633
4634 Compile::TracePhase::~TracePhase() {
4635
4636 C = Compile::current();
4637 if (_dolog) {
4638 _log = C->log();
4639 } else {
4640 _log = NULL;
4641 }
4642
4643 #ifdef ASSERT
4644 if (PrintIdealNodeCount) {
4645 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4646 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4647 }
4648
4649 if (VerifyIdealNodeCount) {
4650 Compile::current()->print_missing_nodes();
4651 }
4652 #endif
4653
4654 if (_log != NULL) {
4655 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4656 }
4657 }
4658
4659 //----------------------------static_subtype_check-----------------------------
4660 // Shortcut important common cases when superklass is exact:
4661 // (0) superklass is java.lang.Object (can occur in reflective code)
4662 // (1) subklass is already limited to a subtype of superklass => always ok
4663 // (2) subklass does not overlap with superklass => always fail
4664 // (3) superklass has NO subtypes and we can check with a simple compare.
4665 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4666 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4667 return SSC_full_test; // Let caller generate the general case.
4668 }
4669
4670 if (superk == env()->Object_klass()) {
4671 return SSC_always_true; // (0) this test cannot fail
4672 }
4673
4674 ciType* superelem = superk;
4675 ciType* subelem = subk;
4676 if (superelem->is_array_klass()) {
4677 superelem = superelem->as_array_klass()->base_element_type();
4678 }
4679 if (subelem->is_array_klass()) {
4680 subelem = subelem->as_array_klass()->base_element_type();
4681 }
4682
4683 if (!subk->is_interface()) { // cannot trust static interface types yet
4684 if (subk->is_subtype_of(superk)) {
4685 return SSC_always_true; // (1) false path dead; no dynamic test needed
4686 }
4687 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4688 !(subelem->is_klass() && subelem->as_klass()->is_interface()) &&
4689 !superk->is_subtype_of(subk)) {
4690 return SSC_always_false; // (2) true path dead; no dynamic test needed
4691 }
4692 }
4693
4694 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4695 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to
4696 // the klass for [LMyValue. Perform a full test.
4697 if (superk->is_obj_array_klass() && !superk->as_array_klass()->is_elem_null_free() &&
4698 superk->as_array_klass()->element_klass()->is_inlinetype()) {
4699 return SSC_full_test;
4700 }
4701 // If casting to an instance klass, it must have no subtypes
4702 if (superk->is_interface()) {
4703 // Cannot trust interfaces yet.
4704 // %%% S.B. superk->nof_implementors() == 1
4705 } else if (superelem->is_instance_klass()) {
4706 ciInstanceKlass* ik = superelem->as_instance_klass();
4707 if (!ik->has_subklass() && !ik->is_interface()) {
4708 if (!ik->is_final()) {
4709 // Add a dependency if there is a chance of a later subclass.
4710 dependencies()->assert_leaf_type(ik);
4711 }
4712 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4713 }
4714 } else {
4715 // A primitive array type has no subtypes.
4716 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4717 }
4718
4719 return SSC_full_test;
4720 }
4721
4722 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4723 #ifdef _LP64
4724 // The scaled index operand to AddP must be a clean 64-bit value.
4725 // Java allows a 32-bit int to be incremented to a negative
4726 // value, which appears in a 64-bit register as a large
4727 // positive number. Using that large positive number as an
4728 // operand in pointer arithmetic has bad consequences.
4729 // On the other hand, 32-bit overflow is rare, and the possibility
4730 // can often be excluded, if we annotate the ConvI2L node with
4731 // a type assertion that its value is known to be a small positive
4732 // number. (The prior range check has ensured this.)
4733 // This assertion is used by ConvI2LNode::Ideal.
4734 int index_max = max_jint - 1; // array size is max_jint, index is one less
4735 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4736 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4737 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4738 #endif
4739 return idx;
4740 }
4741
4742 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4743 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4744 if (ctrl != NULL) {
4745 // Express control dependency by a CastII node with a narrow type.
4746 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4747 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4748 // node from floating above the range check during loop optimizations. Otherwise, the
4749 // ConvI2L node may be eliminated independently of the range check, causing the data path
4750 // to become TOP while the control path is still there (although it's unreachable).
4751 value->set_req(0, ctrl);
4752 value = phase->transform(value);
4753 }
4754 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4755 return phase->transform(new ConvI2LNode(value, ltype));
4756 }
4757
4758 void Compile::print_inlining_stream_free() {
4759 if (_print_inlining_stream != NULL) {
4760 _print_inlining_stream->~stringStream();
4761 _print_inlining_stream = NULL;
4762 }
4763 }
4764
4765 // The message about the current inlining is accumulated in
4766 // _print_inlining_stream and transfered into the _print_inlining_list
4767 // once we know whether inlining succeeds or not. For regular
4768 // inlining, messages are appended to the buffer pointed by
4769 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4770 // a new buffer is added after _print_inlining_idx in the list. This
4771 // way we can update the inlining message for late inlining call site
4772 // when the inlining is attempted again.
4773 void Compile::print_inlining_init() {
4774 if (print_inlining() || print_intrinsics()) {
4775 // print_inlining_init is actually called several times.
4776 print_inlining_stream_free();
4777 _print_inlining_stream = new stringStream();
4778 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4779 }
4780 }
4781
4782 void Compile::print_inlining_reinit() {
4783 if (print_inlining() || print_intrinsics()) {
4784 print_inlining_stream_free();
4785 // Re allocate buffer when we change ResourceMark
4786 _print_inlining_stream = new stringStream();
4787 }
4788 }
4789
4790 void Compile::print_inlining_reset() {
4791 _print_inlining_stream->reset();
4792 }
4793
4794 void Compile::print_inlining_commit() {
4795 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4796 // Transfer the message from _print_inlining_stream to the current
4797 // _print_inlining_list buffer and clear _print_inlining_stream.
4798 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4799 print_inlining_reset();
4800 }
4801
4802 void Compile::print_inlining_push() {
4803 // Add new buffer to the _print_inlining_list at current position
4804 _print_inlining_idx++;
4805 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4806 }
4807
4808 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4809 return _print_inlining_list->at(_print_inlining_idx);
4810 }
4811
4812 void Compile::print_inlining_update(CallGenerator* cg) {
4813 if (print_inlining() || print_intrinsics()) {
4814 if (cg->is_late_inline()) {
4815 if (print_inlining_current()->cg() != cg &&
4816 (print_inlining_current()->cg() != NULL ||
4817 print_inlining_current()->ss()->size() != 0)) {
4818 print_inlining_push();
4819 }
4820 print_inlining_commit();
4821 print_inlining_current()->set_cg(cg);
4822 } else {
4823 if (print_inlining_current()->cg() != NULL) {
4824 print_inlining_push();
4825 }
4826 print_inlining_commit();
4827 }
4828 }
4829 }
4830
4831 void Compile::print_inlining_move_to(CallGenerator* cg) {
4832 // We resume inlining at a late inlining call site. Locate the
4833 // corresponding inlining buffer so that we can update it.
4834 if (print_inlining() || print_intrinsics()) {
4835 for (int i = 0; i < _print_inlining_list->length(); i++) {
4836 if (_print_inlining_list->at(i)->cg() == cg) {
4837 _print_inlining_idx = i;
4838 return;
4839 }
4840 }
4841 ShouldNotReachHere();
4842 }
4843 }
4844
4845 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4846 if (print_inlining() || print_intrinsics()) {
4847 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4848 assert(print_inlining_current()->cg() == cg, "wrong entry");
4849 // replace message with new message
4850 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4851 print_inlining_commit();
4852 print_inlining_current()->set_cg(cg);
4853 }
4854 }
4855
4856 void Compile::print_inlining_assert_ready() {
4857 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4858 }
4859
4860 void Compile::process_print_inlining() {
4861 assert(_late_inlines.length() == 0, "not drained yet");
4862 if (print_inlining() || print_intrinsics()) {
4863 ResourceMark rm;
4864 stringStream ss;
4865 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4866 for (int i = 0; i < _print_inlining_list->length(); i++) {
4867 PrintInliningBuffer* pib = _print_inlining_list->at(i);
4868 ss.print("%s", pib->ss()->as_string());
4869 delete pib;
4870 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
4871 }
4872 // Reset _print_inlining_list, it only contains destructed objects.
4873 // It is on the arena, so it will be freed when the arena is reset.
4874 _print_inlining_list = NULL;
4875 // _print_inlining_stream won't be used anymore, either.
4876 print_inlining_stream_free();
4877 size_t end = ss.size();
4878 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4879 strncpy(_print_inlining_output, ss.base(), end+1);
4880 _print_inlining_output[end] = 0;
4881 }
4882 }
4883
4884 void Compile::dump_print_inlining() {
4885 if (_print_inlining_output != NULL) {
4886 tty->print_raw(_print_inlining_output);
4887 }
4888 }
4889
4890 void Compile::log_late_inline(CallGenerator* cg) {
4891 if (log() != NULL) {
4892 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4893 cg->unique_id());
4894 JVMState* p = cg->call_node()->jvms();
4895 while (p != NULL) {
4896 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4897 p = p->caller();
4898 }
4899 log()->tail("late_inline");
4900 }
4901 }
4902
4903 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4904 log_late_inline(cg);
4905 if (log() != NULL) {
4906 log()->inline_fail(msg);
4907 }
4908 }
4909
4910 void Compile::log_inline_id(CallGenerator* cg) {
4911 if (log() != NULL) {
4912 // The LogCompilation tool needs a unique way to identify late
4913 // inline call sites. This id must be unique for this call site in
4914 // this compilation. Try to have it unique across compilations as
4915 // well because it can be convenient when grepping through the log
4916 // file.
4917 // Distinguish OSR compilations from others in case CICountOSR is
4918 // on.
4919 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4920 cg->set_unique_id(id);
4921 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4922 }
4923 }
4924
4925 void Compile::log_inline_failure(const char* msg) {
4926 if (C->log() != NULL) {
4927 C->log()->inline_fail(msg);
4928 }
4929 }
4930
4931
4932 // Dump inlining replay data to the stream.
4933 // Don't change thread state and acquire any locks.
4934 void Compile::dump_inline_data(outputStream* out) {
4935 InlineTree* inl_tree = ilt();
4936 if (inl_tree != NULL) {
4937 out->print(" inline %d", inl_tree->count());
4938 inl_tree->dump_replay_data(out);
4939 }
4940 }
4941
4942 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4943 if (n1->Opcode() < n2->Opcode()) return -1;
4944 else if (n1->Opcode() > n2->Opcode()) return 1;
4945
4946 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4947 for (uint i = 1; i < n1->req(); i++) {
4948 if (n1->in(i) < n2->in(i)) return -1;
4949 else if (n1->in(i) > n2->in(i)) return 1;
4950 }
4951
4952 return 0;
4953 }
4954
4955 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4956 Node* n1 = *n1p;
4957 Node* n2 = *n2p;
4958
4959 return cmp_expensive_nodes(n1, n2);
4960 }
4961
4962 void Compile::sort_expensive_nodes() {
4963 if (!expensive_nodes_sorted()) {
4964 _expensive_nodes.sort(cmp_expensive_nodes);
4965 }
4966 }
4967
4968 bool Compile::expensive_nodes_sorted() const {
4969 for (int i = 1; i < _expensive_nodes.length(); i++) {
4970 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4971 return false;
4972 }
4973 }
4974 return true;
4975 }
4976
4977 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4978 if (_expensive_nodes.length() == 0) {
4979 return false;
4980 }
4981
4982 assert(OptimizeExpensiveOps, "optimization off?");
4983
4984 // Take this opportunity to remove dead nodes from the list
4985 int j = 0;
4986 for (int i = 0; i < _expensive_nodes.length(); i++) {
4987 Node* n = _expensive_nodes.at(i);
4988 if (!n->is_unreachable(igvn)) {
4989 assert(n->is_expensive(), "should be expensive");
4990 _expensive_nodes.at_put(j, n);
4991 j++;
4992 }
4993 }
4994 _expensive_nodes.trunc_to(j);
4995
4996 // Then sort the list so that similar nodes are next to each other
4997 // and check for at least two nodes of identical kind with same data
4998 // inputs.
4999 sort_expensive_nodes();
5000
5001 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5002 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5003 return true;
5004 }
5005 }
5006
5007 return false;
5008 }
5009
5010 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5011 if (_expensive_nodes.length() == 0) {
5012 return;
5013 }
5014
5015 assert(OptimizeExpensiveOps, "optimization off?");
5016
5017 // Sort to bring similar nodes next to each other and clear the
5018 // control input of nodes for which there's only a single copy.
5019 sort_expensive_nodes();
5020
5021 int j = 0;
5022 int identical = 0;
5023 int i = 0;
5024 bool modified = false;
5025 for (; i < _expensive_nodes.length()-1; i++) {
5026 assert(j <= i, "can't write beyond current index");
5027 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5028 identical++;
5029 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5030 continue;
5031 }
5032 if (identical > 0) {
5033 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5034 identical = 0;
5035 } else {
5036 Node* n = _expensive_nodes.at(i);
5037 igvn.replace_input_of(n, 0, NULL);
5038 igvn.hash_insert(n);
5039 modified = true;
5040 }
5041 }
5042 if (identical > 0) {
5043 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5044 } else if (_expensive_nodes.length() >= 1) {
5045 Node* n = _expensive_nodes.at(i);
5046 igvn.replace_input_of(n, 0, NULL);
5047 igvn.hash_insert(n);
5048 modified = true;
5049 }
5050 _expensive_nodes.trunc_to(j);
5051 if (modified) {
5052 igvn.optimize();
5053 }
5054 }
5055
5056 void Compile::add_expensive_node(Node * n) {
5057 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5058 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5059 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5060 if (OptimizeExpensiveOps) {
5061 _expensive_nodes.append(n);
5062 } else {
5063 // Clear control input and let IGVN optimize expensive nodes if
5064 // OptimizeExpensiveOps is off.
5065 n->set_req(0, NULL);
5066 }
5067 }
5068
5069 /**
5070 * Track coarsened Lock and Unlock nodes.
5071 */
5072
5073 class Lock_List : public Node_List {
5074 uint _origin_cnt;
5075 public:
5076 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5077 uint origin_cnt() const { return _origin_cnt; }
5078 };
5079
5080 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5081 int length = locks.length();
5082 if (length > 0) {
5083 // Have to keep this list until locks elimination during Macro nodes elimination.
5084 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5085 for (int i = 0; i < length; i++) {
5086 AbstractLockNode* lock = locks.at(i);
5087 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5088 locks_list->push(lock);
5089 }
5090 _coarsened_locks.append(locks_list);
5091 }
5092 }
5093
5094 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5095 int count = coarsened_count();
5096 for (int i = 0; i < count; i++) {
5097 Node_List* locks_list = _coarsened_locks.at(i);
5098 for (uint j = 0; j < locks_list->size(); j++) {
5099 Node* lock = locks_list->at(j);
5100 assert(lock->is_AbstractLock(), "sanity");
5101 if (!useful.member(lock)) {
5102 locks_list->yank(lock);
5103 }
5104 }
5105 }
5106 }
5107
5108 void Compile::remove_coarsened_lock(Node* n) {
5109 if (n->is_AbstractLock()) {
5110 int count = coarsened_count();
5111 for (int i = 0; i < count; i++) {
5112 Node_List* locks_list = _coarsened_locks.at(i);
5113 locks_list->yank(n);
5114 }
5115 }
5116 }
5117
5118 bool Compile::coarsened_locks_consistent() {
5119 int count = coarsened_count();
5120 for (int i = 0; i < count; i++) {
5121 bool unbalanced = false;
5122 bool modified = false; // track locks kind modifications
5123 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5124 uint size = locks_list->size();
5125 if (size == 0) {
5126 unbalanced = false; // All locks were eliminated - good
5127 } else if (size != locks_list->origin_cnt()) {
5128 unbalanced = true; // Some locks were removed from list
5129 } else {
5130 for (uint j = 0; j < size; j++) {
5131 Node* lock = locks_list->at(j);
5132 // All nodes in group should have the same state (modified or not)
5133 if (!lock->as_AbstractLock()->is_coarsened()) {
5134 if (j == 0) {
5135 // first on list was modified, the rest should be too for consistency
5136 modified = true;
5137 } else if (!modified) {
5138 // this lock was modified but previous locks on the list were not
5139 unbalanced = true;
5140 break;
5141 }
5142 } else if (modified) {
5143 // previous locks on list were modified but not this lock
5144 unbalanced = true;
5145 break;
5146 }
5147 }
5148 }
5149 if (unbalanced) {
5150 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5151 #ifdef ASSERT
5152 if (PrintEliminateLocks) {
5153 tty->print_cr("=== unbalanced coarsened locks ===");
5154 for (uint l = 0; l < size; l++) {
5155 locks_list->at(l)->dump();
5156 }
5157 }
5158 #endif
5159 record_failure(C2Compiler::retry_no_locks_coarsening());
5160 return false;
5161 }
5162 }
5163 return true;
5164 }
5165
5166 /**
5167 * Remove the speculative part of types and clean up the graph
5168 */
5169 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5170 if (UseTypeSpeculation) {
5171 Unique_Node_List worklist;
5172 worklist.push(root());
5173 int modified = 0;
5174 // Go over all type nodes that carry a speculative type, drop the
5175 // speculative part of the type and enqueue the node for an igvn
5176 // which may optimize it out.
5177 for (uint next = 0; next < worklist.size(); ++next) {
5178 Node *n = worklist.at(next);
5179 if (n->is_Type()) {
5180 TypeNode* tn = n->as_Type();
5181 const Type* t = tn->type();
5182 const Type* t_no_spec = t->remove_speculative();
5183 if (t_no_spec != t) {
5184 bool in_hash = igvn.hash_delete(n);
5185 assert(in_hash, "node should be in igvn hash table");
5186 tn->set_type(t_no_spec);
5187 igvn.hash_insert(n);
5188 igvn._worklist.push(n); // give it a chance to go away
5189 modified++;
5190 }
5191 }
5192 // Iterate over outs - endless loops is unreachable from below
5193 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5194 Node *m = n->fast_out(i);
5195 if (not_a_node(m)) {
5196 continue;
5197 }
5198 worklist.push(m);
5199 }
5200 }
5201 // Drop the speculative part of all types in the igvn's type table
5202 igvn.remove_speculative_types();
5203 if (modified > 0) {
5204 igvn.optimize();
5205 }
5206 #ifdef ASSERT
5207 // Verify that after the IGVN is over no speculative type has resurfaced
5208 worklist.clear();
5209 worklist.push(root());
5210 for (uint next = 0; next < worklist.size(); ++next) {
5211 Node *n = worklist.at(next);
5212 const Type* t = igvn.type_or_null(n);
5213 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5214 if (n->is_Type()) {
5215 t = n->as_Type()->type();
5216 assert(t == t->remove_speculative(), "no more speculative types");
5217 }
5218 // Iterate over outs - endless loops is unreachable from below
5219 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5220 Node *m = n->fast_out(i);
5221 if (not_a_node(m)) {
5222 continue;
5223 }
5224 worklist.push(m);
5225 }
5226 }
5227 igvn.check_no_speculative_types();
5228 #endif
5229 }
5230 }
5231
5232 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5233 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5234 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5235 if (!EnableValhalla || ta == NULL || tb == NULL ||
5236 ta->is_zero_type() || tb->is_zero_type() ||
5237 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5238 // Use old acmp if one operand is null or not an inline type
5239 return new CmpPNode(a, b);
5240 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5241 // We know that one operand is an inline type. Therefore,
5242 // new acmp will only return true if both operands are NULL.
5243 // Check if both operands are null by or'ing the oops.
5244 a = phase->transform(new CastP2XNode(NULL, a));
5245 b = phase->transform(new CastP2XNode(NULL, b));
5246 a = phase->transform(new OrXNode(a, b));
5247 return new CmpXNode(a, phase->MakeConX(0));
5248 }
5249 // Use new acmp
5250 return NULL;
5251 }
5252
5253 // Auxiliary methods to support randomized stressing/fuzzing.
5254
5255 int Compile::random() {
5256 _stress_seed = os::next_random(_stress_seed);
5257 return static_cast<int>(_stress_seed);
5258 }
5259
5260 // This method can be called the arbitrary number of times, with current count
5261 // as the argument. The logic allows selecting a single candidate from the
5262 // running list of candidates as follows:
5263 // int count = 0;
5264 // Cand* selected = null;
5265 // while(cand = cand->next()) {
5266 // if (randomized_select(++count)) {
5267 // selected = cand;
5268 // }
5269 // }
5270 //
5271 // Including count equalizes the chances any candidate is "selected".
5272 // This is useful when we don't have the complete list of candidates to choose
5273 // from uniformly. In this case, we need to adjust the randomicity of the
5274 // selection, or else we will end up biasing the selection towards the latter
5275 // candidates.
5276 //
5277 // Quick back-envelope calculation shows that for the list of n candidates
5278 // the equal probability for the candidate to persist as "best" can be
5279 // achieved by replacing it with "next" k-th candidate with the probability
5280 // of 1/k. It can be easily shown that by the end of the run, the
5281 // probability for any candidate is converged to 1/n, thus giving the
5282 // uniform distribution among all the candidates.
5283 //
5284 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5285 #define RANDOMIZED_DOMAIN_POW 29
5286 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5287 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5288 bool Compile::randomized_select(int count) {
5289 assert(count > 0, "only positive");
5290 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5291 }
5292
5293 CloneMap& Compile::clone_map() { return _clone_map; }
5294 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5295
5296 void NodeCloneInfo::dump() const {
5297 tty->print(" {%d:%d} ", idx(), gen());
5298 }
5299
5300 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5301 uint64_t val = value(old->_idx);
5302 NodeCloneInfo cio(val);
5303 assert(val != 0, "old node should be in the map");
5304 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5305 insert(nnn->_idx, cin.get());
5306 #ifndef PRODUCT
5307 if (is_debug()) {
5308 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5309 }
5310 #endif
5311 }
5312
5313 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5314 NodeCloneInfo cio(value(old->_idx));
5315 if (cio.get() == 0) {
5316 cio.set(old->_idx, 0);
5317 insert(old->_idx, cio.get());
5318 #ifndef PRODUCT
5319 if (is_debug()) {
5320 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5321 }
5322 #endif
5323 }
5324 clone(old, nnn, gen);
5325 }
5326
5327 int CloneMap::max_gen() const {
5328 int g = 0;
5329 DictI di(_dict);
5330 for(; di.test(); ++di) {
5331 int t = gen(di._key);
5332 if (g < t) {
5333 g = t;
5334 #ifndef PRODUCT
5335 if (is_debug()) {
5336 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5337 }
5338 #endif
5339 }
5340 }
5341 return g;
5342 }
5343
5344 void CloneMap::dump(node_idx_t key) const {
5345 uint64_t val = value(key);
5346 if (val != 0) {
5347 NodeCloneInfo ni(val);
5348 ni.dump();
5349 }
5350 }
5351
5352 // Move Allocate nodes to the start of the list
5353 void Compile::sort_macro_nodes() {
5354 int count = macro_count();
5355 int allocates = 0;
5356 for (int i = 0; i < count; i++) {
5357 Node* n = macro_node(i);
5358 if (n->is_Allocate()) {
5359 if (i != allocates) {
5360 Node* tmp = macro_node(allocates);
5361 _macro_nodes.at_put(allocates, n);
5362 _macro_nodes.at_put(i, tmp);
5363 }
5364 allocates++;
5365 }
5366 }
5367 }
5368
5369 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5370 EventCompilerPhase event;
5371 if (event.should_commit()) {
5372 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5373 }
5374 #ifndef PRODUCT
5375 ResourceMark rm;
5376 stringStream ss;
5377 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5378 if (n != nullptr) {
5379 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5380 }
5381
5382 const char* name = ss.as_string();
5383 if (should_print_igv(level)) {
5384 _igv_printer->print_method(name, level);
5385 }
5386 if (should_print_phase(cpt)) {
5387 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5388 }
5389 #endif
5390 C->_latest_stage_start_counter.stamp();
5391 }
5392
5393 // Only used from CompileWrapper
5394 void Compile::begin_method() {
5395 #ifndef PRODUCT
5396 if (_method != NULL && should_print_igv(1)) {
5397 _igv_printer->begin_method();
5398 }
5399 #endif
5400 C->_latest_stage_start_counter.stamp();
5401 }
5402
5403 // Only used from CompileWrapper
5404 void Compile::end_method() {
5405 EventCompilerPhase event;
5406 if (event.should_commit()) {
5407 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5408 }
5409
5410 #ifndef PRODUCT
5411 if (_method != NULL && should_print_igv(1)) {
5412 _igv_printer->end_method();
5413 }
5414 #endif
5415 }
5416
5417 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5418 #ifndef PRODUCT
5419 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5420 return true;
5421 }
5422 #endif
5423 return false;
5424 }
5425
5426 bool Compile::should_print_igv(int level) {
5427 #ifndef PRODUCT
5428 if (PrintIdealGraphLevel < 0) { // disabled by the user
5429 return false;
5430 }
5431
5432 bool need = directive()->IGVPrintLevelOption >= level;
5433 if (need && !_igv_printer) {
5434 _igv_printer = IdealGraphPrinter::printer();
5435 _igv_printer->set_compile(this);
5436 }
5437 return need;
5438 #else
5439 return false;
5440 #endif
5441 }
5442
5443 #ifndef PRODUCT
5444 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
5445 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
5446
5447 // Called from debugger. Prints method to the default file with the default phase name.
5448 // This works regardless of any Ideal Graph Visualizer flags set or not.
5449 void igv_print() {
5450 Compile::current()->igv_print_method_to_file();
5451 }
5452
5453 // Same as igv_print() above but with a specified phase name.
5454 void igv_print(const char* phase_name) {
5455 Compile::current()->igv_print_method_to_file(phase_name);
5456 }
5457
5458 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5459 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5460 // This works regardless of any Ideal Graph Visualizer flags set or not.
5461 void igv_print(bool network) {
5462 if (network) {
5463 Compile::current()->igv_print_method_to_network();
5464 } else {
5465 Compile::current()->igv_print_method_to_file();
5466 }
5467 }
5468
5469 // Same as igv_print(bool network) above but with a specified phase name.
5470 void igv_print(bool network, const char* phase_name) {
5471 if (network) {
5472 Compile::current()->igv_print_method_to_network(phase_name);
5473 } else {
5474 Compile::current()->igv_print_method_to_file(phase_name);
5475 }
5476 }
5477
5478 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5479 void igv_print_default() {
5480 Compile::current()->print_method(PHASE_DEBUG, 0);
5481 }
5482
5483 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5484 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5485 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5486 void igv_append() {
5487 Compile::current()->igv_print_method_to_file("Debug", true);
5488 }
5489
5490 // Same as igv_append() above but with a specified phase name.
5491 void igv_append(const char* phase_name) {
5492 Compile::current()->igv_print_method_to_file(phase_name, true);
5493 }
5494
5495 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5496 const char* file_name = "custom_debug.xml";
5497 if (_debug_file_printer == NULL) {
5498 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5499 } else {
5500 _debug_file_printer->update_compiled_method(C->method());
5501 }
5502 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5503 _debug_file_printer->print(phase_name, (Node*)C->root());
5504 }
5505
5506 void Compile::igv_print_method_to_network(const char* phase_name) {
5507 if (_debug_network_printer == NULL) {
5508 _debug_network_printer = new IdealGraphPrinter(C);
5509 } else {
5510 _debug_network_printer->update_compiled_method(C->method());
5511 }
5512 tty->print_cr("Method printed over network stream to IGV");
5513 _debug_network_printer->print(phase_name, (Node*)C->root());
5514 }
5515 #endif
5516
5517 void Compile::add_native_invoker(RuntimeStub* stub) {
5518 _native_invokers.append(stub);
5519 }
5520
5521 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5522 if (type != NULL && phase->type(value)->higher_equal(type)) {
5523 return value;
5524 }
5525 Node* result = NULL;
5526 if (bt == T_BYTE) {
5527 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5528 result = new RShiftINode(result, phase->intcon(24));
5529 } else if (bt == T_BOOLEAN) {
5530 result = new AndINode(value, phase->intcon(0xFF));
5531 } else if (bt == T_CHAR) {
5532 result = new AndINode(value,phase->intcon(0xFFFF));
5533 } else {
5534 assert(bt == T_SHORT, "unexpected narrow type");
5535 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5536 result = new RShiftINode(result, phase->intcon(16));
5537 }
5538 if (transform_res) {
5539 result = phase->transform(result);
5540 }
5541 return result;
5542 }
5543