1 /*
2 * Copyright (c) 1997, 2025, 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 "asm/macroAssembler.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "ci/ciReplay.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compilationFailureInfo.hpp"
32 #include "compiler/compilationMemoryStatistic.hpp"
33 #include "compiler/compileBroker.hpp"
34 #include "compiler/compileLog.hpp"
35 #include "compiler/compilerOracle.hpp"
36 #include "compiler/compiler_globals.hpp"
37 #include "compiler/disassembler.hpp"
38 #include "compiler/oopMap.hpp"
39 #include "gc/shared/barrierSet.hpp"
40 #include "gc/shared/c2/barrierSetC2.hpp"
41 #include "jfr/jfrEvents.hpp"
42 #include "jvm_io.h"
43 #include "memory/allocation.hpp"
44 #include "memory/arena.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "opto/addnode.hpp"
47 #include "opto/block.hpp"
48 #include "opto/c2compiler.hpp"
49 #include "opto/callGenerator.hpp"
50 #include "opto/callnode.hpp"
51 #include "opto/castnode.hpp"
52 #include "opto/cfgnode.hpp"
53 #include "opto/chaitin.hpp"
54 #include "opto/compile.hpp"
55 #include "opto/connode.hpp"
56 #include "opto/convertnode.hpp"
57 #include "opto/divnode.hpp"
58 #include "opto/escape.hpp"
59 #include "opto/idealGraphPrinter.hpp"
60 #include "opto/inlinetypenode.hpp"
61 #include "opto/locknode.hpp"
62 #include "opto/loopnode.hpp"
63 #include "opto/machnode.hpp"
64 #include "opto/macro.hpp"
65 #include "opto/matcher.hpp"
66 #include "opto/mathexactnode.hpp"
67 #include "opto/memnode.hpp"
68 #include "opto/movenode.hpp"
69 #include "opto/mulnode.hpp"
70 #include "opto/narrowptrnode.hpp"
71 #include "opto/node.hpp"
72 #include "opto/opaquenode.hpp"
73 #include "opto/opcodes.hpp"
74 #include "opto/output.hpp"
75 #include "opto/parse.hpp"
76 #include "opto/phaseX.hpp"
77 #include "opto/rootnode.hpp"
78 #include "opto/runtime.hpp"
79 #include "opto/stringopts.hpp"
80 #include "opto/type.hpp"
81 #include "opto/vector.hpp"
82 #include "opto/vectornode.hpp"
83 #include "runtime/globals_extension.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stubRoutines.hpp"
87 #include "runtime/timer.hpp"
88 #include "utilities/align.hpp"
89 #include "utilities/copy.hpp"
90 #include "utilities/macros.hpp"
91 #include "utilities/resourceHash.hpp"
92
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == nullptr) {
97 _mach_constant_base_node = new MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
102
103
104 /// Support for intrinsics.
105
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 class IntrinsicDescPair {
109 private:
110 ciMethod* _m;
111 bool _is_virtual;
112 public:
113 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
114 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
115 ciMethod* m= elt->method();
116 ciMethod* key_m = key->_m;
117 if (key_m < m) return -1;
118 else if (key_m > m) return 1;
119 else {
120 bool is_virtual = elt->is_virtual();
121 bool key_virtual = key->_is_virtual;
122 if (key_virtual < is_virtual) return -1;
123 else if (key_virtual > is_virtual) return 1;
124 else return 0;
125 }
126 }
127 };
128 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
129 #ifdef ASSERT
130 for (int i = 1; i < _intrinsics.length(); i++) {
131 CallGenerator* cg1 = _intrinsics.at(i-1);
132 CallGenerator* cg2 = _intrinsics.at(i);
133 assert(cg1->method() != cg2->method()
134 ? cg1->method() < cg2->method()
135 : cg1->is_virtual() < cg2->is_virtual(),
136 "compiler intrinsics list must stay sorted");
137 }
138 #endif
139 IntrinsicDescPair pair(m, is_virtual);
140 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
141 }
142
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 bool found = false;
145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
146 assert(!found, "registering twice");
147 _intrinsics.insert_before(index, cg);
148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
149 }
150
151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
152 assert(m->is_loaded(), "don't try this on unloaded methods");
153 if (_intrinsics.length() > 0) {
154 bool found = false;
155 int index = intrinsic_insertion_index(m, is_virtual, found);
156 if (found) {
157 return _intrinsics.at(index);
158 }
159 }
160 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
161 if (m->intrinsic_id() != vmIntrinsics::_none &&
162 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
163 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
164 if (cg != nullptr) {
165 // Save it for next time:
166 register_intrinsic(cg);
167 return cg;
168 } else {
169 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
170 }
171 }
172 return nullptr;
173 }
174
175 // Compile::make_vm_intrinsic is defined in library_call.cpp.
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
182
183 inline int as_int(vmIntrinsics::ID id) {
184 return vmIntrinsics::as_int(id);
185 }
186
187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
188 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
189 int oflags = _intrinsic_hist_flags[as_int(id)];
190 assert(flags != 0, "what happened?");
191 if (is_virtual) {
192 flags |= _intrinsic_virtual;
193 }
194 bool changed = (flags != oflags);
195 if ((flags & _intrinsic_worked) != 0) {
196 juint count = (_intrinsic_hist_count[as_int(id)] += 1);
197 if (count == 1) {
198 changed = true; // first time
199 }
200 // increment the overall count also:
201 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
202 }
203 if (changed) {
204 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
205 // Something changed about the intrinsic's virtuality.
206 if ((flags & _intrinsic_virtual) != 0) {
207 // This is the first use of this intrinsic as a virtual call.
208 if (oflags != 0) {
209 // We already saw it as a non-virtual, so note both cases.
210 flags |= _intrinsic_both;
211 }
212 } else if ((oflags & _intrinsic_both) == 0) {
213 // This is the first use of this intrinsic as a non-virtual
214 flags |= _intrinsic_both;
215 }
216 }
217 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
218 }
219 // update the overall flags also:
220 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
221 return changed;
222 }
223
224 static char* format_flags(int flags, char* buf) {
225 buf[0] = 0;
226 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
227 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
228 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
229 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
230 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
231 if (buf[0] == 0) strcat(buf, ",");
232 assert(buf[0] == ',', "must be");
233 return &buf[1];
234 }
235
236 void Compile::print_intrinsic_statistics() {
237 char flagsbuf[100];
238 ttyLocker ttyl;
239 if (xtty != nullptr) xtty->head("statistics type='intrinsic'");
240 tty->print_cr("Compiler intrinsic usage:");
241 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
242 if (total == 0) total = 1; // avoid div0 in case of no successes
243 #define PRINT_STAT_LINE(name, c, f) \
244 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
245 for (auto id : EnumRange<vmIntrinsicID>{}) {
246 int flags = _intrinsic_hist_flags[as_int(id)];
247 juint count = _intrinsic_hist_count[as_int(id)];
248 if ((flags | count) != 0) {
249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
250 }
251 }
252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
253 if (xtty != nullptr) xtty->tail("statistics");
254 }
255
256 void Compile::print_statistics() {
257 { ttyLocker ttyl;
258 if (xtty != nullptr) xtty->head("statistics type='opto'");
259 Parse::print_statistics();
260 PhaseStringOpts::print_statistics();
261 PhaseCCP::print_statistics();
262 PhaseRegAlloc::print_statistics();
263 PhaseOutput::print_statistics();
264 PhasePeephole::print_statistics();
265 PhaseIdealLoop::print_statistics();
266 ConnectionGraph::print_statistics();
267 PhaseMacroExpand::print_statistics();
268 if (xtty != nullptr) xtty->tail("statistics");
269 }
270 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
271 // put this under its own <statistics> element.
272 print_intrinsic_statistics();
273 }
274 }
275 #endif //PRODUCT
276
277 void Compile::gvn_replace_by(Node* n, Node* nn) {
278 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
279 Node* use = n->last_out(i);
280 bool is_in_table = initial_gvn()->hash_delete(use);
281 uint uses_found = 0;
282 for (uint j = 0; j < use->len(); j++) {
283 if (use->in(j) == n) {
284 if (j < use->req())
285 use->set_req(j, nn);
286 else
287 use->set_prec(j, nn);
288 uses_found++;
289 }
290 }
291 if (is_in_table) {
292 // reinsert into table
293 initial_gvn()->hash_find_insert(use);
294 }
295 record_for_igvn(use);
296 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist);
297 i -= uses_found; // we deleted 1 or more copies of this edge
298 }
299 }
300
301
302 // Identify all nodes that are reachable from below, useful.
303 // Use breadth-first pass that records state in a Unique_Node_List,
304 // recursive traversal is slower.
305 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
306 int estimated_worklist_size = live_nodes();
307 useful.map( estimated_worklist_size, nullptr ); // preallocate space
308
309 // Initialize worklist
310 if (root() != nullptr) { useful.push(root()); }
311 // If 'top' is cached, declare it useful to preserve cached node
312 if (cached_top_node()) { useful.push(cached_top_node()); }
313
314 // Push all useful nodes onto the list, breadthfirst
315 for( uint next = 0; next < useful.size(); ++next ) {
316 assert( next < unique(), "Unique useful nodes < total nodes");
317 Node *n = useful.at(next);
318 uint max = n->len();
319 for( uint i = 0; i < max; ++i ) {
320 Node *m = n->in(i);
321 if (not_a_node(m)) continue;
322 useful.push(m);
323 }
324 }
325 }
326
327 // Update dead_node_list with any missing dead nodes using useful
328 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
329 void Compile::update_dead_node_list(Unique_Node_List &useful) {
330 uint max_idx = unique();
331 VectorSet& useful_node_set = useful.member_set();
332
333 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
334 // If node with index node_idx is not in useful set,
335 // mark it as dead in dead node list.
336 if (!useful_node_set.test(node_idx)) {
337 record_dead_node(node_idx);
338 }
339 }
340 }
341
342 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
343 int shift = 0;
344 for (int i = 0; i < inlines->length(); i++) {
345 CallGenerator* cg = inlines->at(i);
346 if (useful.member(cg->call_node())) {
347 if (shift > 0) {
348 inlines->at_put(i - shift, cg);
349 }
350 } else {
351 shift++; // skip over the dead element
352 }
353 }
354 if (shift > 0) {
355 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
356 }
357 }
358
359 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
360 assert(dead != nullptr && dead->is_Call(), "sanity");
361 int found = 0;
362 for (int i = 0; i < inlines->length(); i++) {
363 if (inlines->at(i)->call_node() == dead) {
364 inlines->remove_at(i);
365 found++;
366 NOT_DEBUG( break; ) // elements are unique, so exit early
367 }
368 }
369 assert(found <= 1, "not unique");
370 }
371
372 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)>
373 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) {
374 for (int i = node_list.length() - 1; i >= 0; i--) {
375 N* node = node_list.at(i);
376 if (!useful.member(node)) {
377 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
378 }
379 }
380 }
381
382 void Compile::remove_useless_node(Node* dead) {
383 remove_modified_node(dead);
384
385 // Constant node that has no out-edges and has only one in-edge from
386 // root is usually dead. However, sometimes reshaping walk makes
387 // it reachable by adding use edges. So, we will NOT count Con nodes
388 // as dead to be conservative about the dead node count at any
389 // given time.
390 if (!dead->is_Con()) {
391 record_dead_node(dead->_idx);
392 }
393 if (dead->is_macro()) {
394 remove_macro_node(dead);
395 }
396 if (dead->is_expensive()) {
397 remove_expensive_node(dead);
398 }
399 if (dead->is_OpaqueTemplateAssertionPredicate()) {
400 remove_template_assertion_predicate_opaque(dead->as_OpaqueTemplateAssertionPredicate());
401 }
402 if (dead->is_ParsePredicate()) {
403 remove_parse_predicate(dead->as_ParsePredicate());
404 }
405 if (dead->for_post_loop_opts_igvn()) {
406 remove_from_post_loop_opts_igvn(dead);
407 }
408 if (dead->is_InlineType()) {
409 remove_inline_type(dead);
410 }
411 if (dead->for_merge_stores_igvn()) {
412 remove_from_merge_stores_igvn(dead);
413 }
414 if (dead->is_Call()) {
415 remove_useless_late_inlines( &_late_inlines, dead);
416 remove_useless_late_inlines( &_string_late_inlines, dead);
417 remove_useless_late_inlines( &_boxing_late_inlines, dead);
418 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
419
420 if (dead->is_CallStaticJava()) {
421 remove_unstable_if_trap(dead->as_CallStaticJava(), false);
422 }
423 }
424 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
425 bs->unregister_potential_barrier_node(dead);
426 }
427
428 // Disconnect all useless nodes by disconnecting those at the boundary.
429 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist, const Unique_Node_List* root_and_safepoints) {
430 uint next = 0;
431 while (next < useful.size()) {
432 Node *n = useful.at(next++);
433 if (n->is_SafePoint()) {
434 // We're done with a parsing phase. Replaced nodes are not valid
435 // beyond that point.
436 n->as_SafePoint()->delete_replaced_nodes();
437 }
438 // Use raw traversal of out edges since this code removes out edges
439 int max = n->outcnt();
440 for (int j = 0; j < max; ++j) {
441 Node* child = n->raw_out(j);
442 if (!useful.member(child)) {
443 assert(!child->is_top() || child != top(),
444 "If top is cached in Compile object it is in useful list");
445 // Only need to remove this out-edge to the useless node
446 n->raw_del_out(j);
447 --j;
448 --max;
449 if (child->is_data_proj_of_pure_function(n)) {
450 worklist.push(n);
451 }
452 }
453 }
454 if (n->outcnt() == 1 && n->has_special_unique_user()) {
455 assert(useful.member(n->unique_out()), "do not push a useless node");
456 worklist.push(n->unique_out());
457 }
458 if (n->outcnt() == 0) {
459 worklist.push(n);
460 }
461 }
462
463 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes
464 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes
465 // Remove useless Template Assertion Predicate opaque nodes
466 remove_useless_nodes(_template_assertion_predicate_opaques, useful);
467 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes
468 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
469 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes
470 #ifdef ASSERT
471 if (_modified_nodes != nullptr) {
472 _modified_nodes->remove_useless_nodes(useful.member_set());
473 }
474 #endif
475 remove_useless_nodes(_for_merge_stores_igvn, useful); // remove useless node recorded for merge stores IGVN pass
476 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps
477 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes
478 #ifdef ASSERT
479 if (_modified_nodes != nullptr) {
480 _modified_nodes->remove_useless_nodes(useful.member_set());
481 }
482 #endif
483
484 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
485 bs->eliminate_useless_gc_barriers(useful, this);
486 // clean up the late inline lists
487 remove_useless_late_inlines( &_late_inlines, useful);
488 remove_useless_late_inlines( &_string_late_inlines, useful);
489 remove_useless_late_inlines( &_boxing_late_inlines, useful);
490 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
491 debug_only(verify_graph_edges(true /*check for no_dead_code*/, root_and_safepoints);)
492 }
493
494 // ============================================================================
495 //------------------------------CompileWrapper---------------------------------
496 class CompileWrapper : public StackObj {
497 Compile *const _compile;
498 public:
499 CompileWrapper(Compile* compile);
500
501 ~CompileWrapper();
502 };
503
504 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
505 // the Compile* pointer is stored in the current ciEnv:
506 ciEnv* env = compile->env();
507 assert(env == ciEnv::current(), "must already be a ciEnv active");
508 assert(env->compiler_data() == nullptr, "compile already active?");
509 env->set_compiler_data(compile);
510 assert(compile == Compile::current(), "sanity");
511
512 compile->set_type_dict(nullptr);
513 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
514 compile->clone_map().set_clone_idx(0);
515 compile->set_type_last_size(0);
516 compile->set_last_tf(nullptr, nullptr);
517 compile->set_indexSet_arena(nullptr);
518 compile->set_indexSet_free_block_list(nullptr);
519 compile->init_type_arena();
520 Type::Initialize(compile);
521 _compile->begin_method();
522 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
523 }
524 CompileWrapper::~CompileWrapper() {
525 // simulate crash during compilation
526 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
527
528 _compile->end_method();
529 _compile->env()->set_compiler_data(nullptr);
530 }
531
532
533 //----------------------------print_compile_messages---------------------------
534 void Compile::print_compile_messages() {
535 #ifndef PRODUCT
536 // Check if recompiling
537 if (!subsume_loads() && PrintOpto) {
538 // Recompiling without allowing machine instructions to subsume loads
539 tty->print_cr("*********************************************************");
540 tty->print_cr("** Bailout: Recompile without subsuming loads **");
541 tty->print_cr("*********************************************************");
542 }
543 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
544 // Recompiling without escape analysis
545 tty->print_cr("*********************************************************");
546 tty->print_cr("** Bailout: Recompile without escape analysis **");
547 tty->print_cr("*********************************************************");
548 }
549 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
550 // Recompiling without iterative escape analysis
551 tty->print_cr("*********************************************************");
552 tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
553 tty->print_cr("*********************************************************");
554 }
555 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) {
556 // Recompiling without reducing allocation merges
557 tty->print_cr("*********************************************************");
558 tty->print_cr("** Bailout: Recompile without reduce allocation merges **");
559 tty->print_cr("*********************************************************");
560 }
561 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
562 // Recompiling without boxing elimination
563 tty->print_cr("*********************************************************");
564 tty->print_cr("** Bailout: Recompile without boxing elimination **");
565 tty->print_cr("*********************************************************");
566 }
567 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
568 // Recompiling without locks coarsening
569 tty->print_cr("*********************************************************");
570 tty->print_cr("** Bailout: Recompile without locks coarsening **");
571 tty->print_cr("*********************************************************");
572 }
573 if (env()->break_at_compile()) {
574 // Open the debugger when compiling this method.
575 tty->print("### Breaking when compiling: ");
576 method()->print_short_name();
577 tty->cr();
578 BREAKPOINT;
579 }
580
581 if( PrintOpto ) {
582 if (is_osr_compilation()) {
583 tty->print("[OSR]%3d", _compile_id);
584 } else {
585 tty->print("%3d", _compile_id);
586 }
587 }
588 #endif
589 }
590
591 #ifndef PRODUCT
592 void Compile::print_ideal_ir(const char* phase_name) {
593 // keep the following output all in one block
594 // This output goes directly to the tty, not the compiler log.
595 // To enable tools to match it up with the compilation activity,
596 // be sure to tag this tty output with the compile ID.
597
598 // Node dumping can cause a safepoint, which can break the tty lock.
599 // Buffer all node dumps, so that all safepoints happen before we lock.
600 ResourceMark rm;
601 stringStream ss;
602
603 if (_output == nullptr) {
604 ss.print_cr("AFTER: %s", phase_name);
605 // Print out all nodes in ascending order of index.
606 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss);
607 } else {
608 // Dump the node blockwise if we have a scheduling
609 _output->print_scheduling(&ss);
610 }
611
612 // Check that the lock is not broken by a safepoint.
613 NoSafepointVerifier nsv;
614 ttyLocker ttyl;
615 if (xtty != nullptr) {
616 xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
617 compile_id(),
618 is_osr_compilation() ? " compile_kind='osr'" : "",
619 phase_name);
620 }
621
622 tty->print("%s", ss.as_string());
623
624 if (xtty != nullptr) {
625 xtty->tail("ideal");
626 }
627 }
628 #endif
629
630 // ============================================================================
631 //------------------------------Compile standard-------------------------------
632
633 // Compile a method. entry_bci is -1 for normal compilations and indicates
634 // the continuation bci for on stack replacement.
635
636
637 Compile::Compile(ciEnv* ci_env, ciMethod* target, int osr_bci,
638 Options options, DirectiveSet* directive)
639 : Phase(Compiler),
640 _compile_id(ci_env->compile_id()),
641 _options(options),
642 _method(target),
643 _entry_bci(osr_bci),
644 _ilt(nullptr),
645 _stub_function(nullptr),
646 _stub_name(nullptr),
647 _stub_entry_point(nullptr),
648 _max_node_limit(MaxNodeLimit),
649 _post_loop_opts_phase(false),
650 _merge_stores_phase(false),
651 _allow_macro_nodes(true),
652 _inlining_progress(false),
653 _inlining_incrementally(false),
654 _do_cleanup(false),
655 _has_reserved_stack_access(target->has_reserved_stack_access()),
656 _has_circular_inline_type(false),
657 #ifndef PRODUCT
658 _igv_idx(0),
659 _trace_opto_output(directive->TraceOptoOutputOption),
660 #endif
661 _has_method_handle_invokes(false),
662 _clinit_barrier_on_entry(false),
663 _stress_seed(0),
664 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
665 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
666 _env(ci_env),
667 _directive(directive),
668 _log(ci_env->log()),
669 _first_failure_details(nullptr),
670 _intrinsics(comp_arena(), 0, 0, nullptr),
671 _macro_nodes(comp_arena(), 8, 0, nullptr),
672 _parse_predicates(comp_arena(), 8, 0, nullptr),
673 _template_assertion_predicate_opaques(comp_arena(), 8, 0, nullptr),
674 _expensive_nodes(comp_arena(), 8, 0, nullptr),
675 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
676 _inline_type_nodes (comp_arena(), 8, 0, nullptr),
677 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
678 _unstable_if_traps(comp_arena(), 8, 0, nullptr),
679 _coarsened_locks(comp_arena(), 8, 0, nullptr),
680 _congraph(nullptr),
681 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
682 _unique(0),
683 _dead_node_count(0),
684 _dead_node_list(comp_arena()),
685 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
686 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
687 _node_arena(&_node_arena_one),
688 _mach_constant_base_node(nullptr),
689 _Compile_types(mtCompiler, Arena::Tag::tag_type),
690 _initial_gvn(nullptr),
691 _igvn_worklist(nullptr),
692 _types(nullptr),
693 _node_hash(nullptr),
694 _late_inlines(comp_arena(), 2, 0, nullptr),
695 _string_late_inlines(comp_arena(), 2, 0, nullptr),
696 _boxing_late_inlines(comp_arena(), 2, 0, nullptr),
697 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr),
698 _late_inlines_pos(0),
699 _number_of_mh_late_inlines(0),
700 _oom(false),
701 _replay_inline_data(nullptr),
702 _inline_printer(this),
703 _java_calls(0),
704 _inner_loops(0),
705 _interpreter_frame_size(0),
706 _output(nullptr)
707 #ifndef PRODUCT
708 ,
709 _in_dump_cnt(0)
710 #endif
711 {
712 C = this;
713 CompileWrapper cw(this);
714
715 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
716 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false);
717
718 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
719 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
720 // We can always print a disassembly, either abstract (hex dump) or
721 // with the help of a suitable hsdis library. Thus, we should not
722 // couple print_assembly and print_opto_assembly controls.
723 // But: always print opto and regular assembly on compile command 'print'.
724 bool print_assembly = directive->PrintAssemblyOption;
725 set_print_assembly(print_opto_assembly || print_assembly);
726 #else
727 set_print_assembly(false); // must initialize.
728 #endif
729
730 #ifndef PRODUCT
731 set_parsed_irreducible_loop(false);
732 #endif
733
734 if (directive->ReplayInlineOption) {
735 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
736 }
737 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
738 set_print_intrinsics(directive->PrintIntrinsicsOption);
739 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
740
741 if (ProfileTraps) {
742 // Make sure the method being compiled gets its own MDO,
743 // so we can at least track the decompile_count().
744 method()->ensure_method_data();
745 }
746
747 if (StressLCM || StressGCM || StressIGVN || StressCCP ||
748 StressIncrementalInlining || StressMacroExpansion || StressUnstableIfTraps || StressBailout) {
749 initialize_stress_seed(directive);
750 }
751
752 Init(/*do_aliasing=*/ true);
753
754 print_compile_messages();
755
756 _ilt = InlineTree::build_inline_tree_root();
757
758 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
759 assert(num_alias_types() >= AliasIdxRaw, "");
760
761 #define MINIMUM_NODE_HASH 1023
762
763 // GVN that will be run immediately on new nodes
764 uint estimated_size = method()->code_size()*4+64;
765 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
766 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
767 _types = new (comp_arena()) Type_Array(comp_arena());
768 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size);
769 PhaseGVN gvn;
770 set_initial_gvn(&gvn);
771
772 { // Scope for timing the parser
773 TracePhase tp(_t_parser);
774
775 // Put top into the hash table ASAP.
776 initial_gvn()->transform(top());
777
778 // Set up tf(), start(), and find a CallGenerator.
779 CallGenerator* cg = nullptr;
780 if (is_osr_compilation()) {
781 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true));
782 StartNode* s = new StartOSRNode(root(), tf()->domain_sig());
783 initial_gvn()->set_type_bottom(s);
784 verify_start(s);
785 cg = CallGenerator::for_osr(method(), entry_bci());
786 } else {
787 // Normal case.
788 init_tf(TypeFunc::make(method()));
789 StartNode* s = new StartNode(root(), tf()->domain_cc());
790 initial_gvn()->set_type_bottom(s);
791 verify_start(s);
792 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
793 // With java.lang.ref.reference.get() we must go through the
794 // intrinsic - even when get() is the root
795 // method of the compile - so that, if necessary, the value in
796 // the referent field of the reference object gets recorded by
797 // the pre-barrier code.
798 cg = find_intrinsic(method(), false);
799 }
800 if (cg == nullptr) {
801 float past_uses = method()->interpreter_invocation_count();
802 float expected_uses = past_uses;
803 cg = CallGenerator::for_inline(method(), expected_uses);
804 }
805 }
806 if (failing()) return;
807 if (cg == nullptr) {
808 const char* reason = InlineTree::check_can_parse(method());
809 assert(reason != nullptr, "expect reason for parse failure");
810 stringStream ss;
811 ss.print("cannot parse method: %s", reason);
812 record_method_not_compilable(ss.as_string());
813 return;
814 }
815
816 gvn.set_type(root(), root()->bottom_type());
817
818 JVMState* jvms = build_start_state(start(), tf());
819 if ((jvms = cg->generate(jvms)) == nullptr) {
820 assert(failure_reason() != nullptr, "expect reason for parse failure");
821 stringStream ss;
822 ss.print("method parse failed: %s", failure_reason());
823 record_method_not_compilable(ss.as_string() DEBUG_ONLY(COMMA true));
824 return;
825 }
826 GraphKit kit(jvms);
827
828 if (!kit.stopped()) {
829 // Accept return values, and transfer control we know not where.
830 // This is done by a special, unique ReturnNode bound to root.
831 return_values(kit.jvms());
832 }
833
834 if (kit.has_exceptions()) {
835 // Any exceptions that escape from this call must be rethrown
836 // to whatever caller is dynamically above us on the stack.
837 // This is done by a special, unique RethrowNode bound to root.
838 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
839 }
840
841 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
842
843 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
844 inline_string_calls(true);
845 }
846
847 if (failing()) return;
848
849 // Remove clutter produced by parsing.
850 if (!failing()) {
851 ResourceMark rm;
852 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
853 }
854 }
855
856 // Note: Large methods are capped off in do_one_bytecode().
857 if (failing()) return;
858
859 // After parsing, node notes are no longer automagic.
860 // They must be propagated by register_new_node_with_optimizer(),
861 // clone(), or the like.
862 set_default_node_notes(nullptr);
863
864 #ifndef PRODUCT
865 if (should_print_igv(1)) {
866 _igv_printer->print_inlining();
867 }
868 #endif
869
870 if (failing()) return;
871 NOT_PRODUCT( verify_graph_edges(); )
872
873 // Now optimize
874 Optimize();
875 if (failing()) return;
876 NOT_PRODUCT( verify_graph_edges(); )
877
878 #ifndef PRODUCT
879 if (should_print_ideal()) {
880 print_ideal_ir("print_ideal");
881 }
882 #endif
883
884 #ifdef ASSERT
885 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
886 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
887 #endif
888
889 // Dump compilation data to replay it.
890 if (directive->DumpReplayOption) {
891 env()->dump_replay_data(_compile_id);
892 }
893 if (directive->DumpInlineOption && (ilt() != nullptr)) {
894 env()->dump_inline_data(_compile_id);
895 }
896
897 // Now that we know the size of all the monitors we can add a fixed slot
898 // for the original deopt pc.
899 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
900 if (needs_stack_repair()) {
901 // One extra slot for the special stack increment value
902 next_slot += 2;
903 }
904 // TODO 8284443 Only reserve extra slot if needed
905 if (InlineTypeReturnedAsFields) {
906 // One extra slot to hold the IsInit information for a nullable
907 // inline type return if we run out of registers.
908 next_slot += 2;
909 }
910 set_fixed_slots(next_slot);
911
912 // Compute when to use implicit null checks. Used by matching trap based
913 // nodes and NullCheck optimization.
914 set_allowed_deopt_reasons();
915
916 // Now generate code
917 Code_Gen();
918 }
919
920 //------------------------------Compile----------------------------------------
921 // Compile a runtime stub
922 Compile::Compile(ciEnv* ci_env,
923 TypeFunc_generator generator,
924 address stub_function,
925 const char *stub_name,
926 int is_fancy_jump,
927 bool pass_tls,
928 bool return_pc,
929 DirectiveSet* directive)
930 : Phase(Compiler),
931 _compile_id(0),
932 _options(Options::for_runtime_stub()),
933 _method(nullptr),
934 _entry_bci(InvocationEntryBci),
935 _stub_function(stub_function),
936 _stub_name(stub_name),
937 _stub_entry_point(nullptr),
938 _max_node_limit(MaxNodeLimit),
939 _post_loop_opts_phase(false),
940 _merge_stores_phase(false),
941 _allow_macro_nodes(true),
942 _inlining_progress(false),
943 _inlining_incrementally(false),
944 _has_reserved_stack_access(false),
945 _has_circular_inline_type(false),
946 #ifndef PRODUCT
947 _igv_idx(0),
948 _trace_opto_output(directive->TraceOptoOutputOption),
949 #endif
950 _has_method_handle_invokes(false),
951 _clinit_barrier_on_entry(false),
952 _stress_seed(0),
953 _comp_arena(mtCompiler, Arena::Tag::tag_comp),
954 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
955 _env(ci_env),
956 _directive(directive),
957 _log(ci_env->log()),
958 _first_failure_details(nullptr),
959 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr),
960 _for_merge_stores_igvn(comp_arena(), 8, 0, nullptr),
961 _congraph(nullptr),
962 NOT_PRODUCT(_igv_printer(nullptr) COMMA)
963 _unique(0),
964 _dead_node_count(0),
965 _dead_node_list(comp_arena()),
966 _node_arena_one(mtCompiler, Arena::Tag::tag_node),
967 _node_arena_two(mtCompiler, Arena::Tag::tag_node),
968 _node_arena(&_node_arena_one),
969 _mach_constant_base_node(nullptr),
970 _Compile_types(mtCompiler, Arena::Tag::tag_type),
971 _initial_gvn(nullptr),
972 _igvn_worklist(nullptr),
973 _types(nullptr),
974 _node_hash(nullptr),
975 _number_of_mh_late_inlines(0),
976 _oom(false),
977 _replay_inline_data(nullptr),
978 _inline_printer(this),
979 _java_calls(0),
980 _inner_loops(0),
981 _interpreter_frame_size(0),
982 _output(nullptr),
983 #ifndef PRODUCT
984 _in_dump_cnt(0),
985 #endif
986 _allowed_reasons(0) {
987 C = this;
988
989 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false);
990 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false);
991
992 #ifndef PRODUCT
993 set_print_assembly(PrintFrameConverterAssembly);
994 set_parsed_irreducible_loop(false);
995 #else
996 set_print_assembly(false); // Must initialize.
997 #endif
998 set_has_irreducible_loop(false); // no loops
999
1000 CompileWrapper cw(this);
1001 Init(/*do_aliasing=*/ false);
1002 init_tf((*generator)());
1003
1004 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena());
1005 _types = new (comp_arena()) Type_Array(comp_arena());
1006 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255);
1007
1008 if (StressLCM || StressGCM || StressBailout) {
1009 initialize_stress_seed(directive);
1010 }
1011
1012 {
1013 PhaseGVN gvn;
1014 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1015 gvn.transform(top());
1016
1017 GraphKit kit;
1018 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1019 }
1020
1021 NOT_PRODUCT( verify_graph_edges(); )
1022
1023 Code_Gen();
1024 }
1025
1026 Compile::~Compile() {
1027 delete _first_failure_details;
1028 };
1029
1030 //------------------------------Init-------------------------------------------
1031 // Prepare for a single compilation
1032 void Compile::Init(bool aliasing) {
1033 _do_aliasing = aliasing;
1034 _unique = 0;
1035 _regalloc = nullptr;
1036
1037 _tf = nullptr; // filled in later
1038 _top = nullptr; // cached later
1039 _matcher = nullptr; // filled in later
1040 _cfg = nullptr; // filled in later
1041
1042 IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
1043
1044 _node_note_array = nullptr;
1045 _default_node_notes = nullptr;
1046 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize()
1047
1048 _immutable_memory = nullptr; // filled in at first inquiry
1049
1050 #ifdef ASSERT
1051 _phase_optimize_finished = false;
1052 _phase_verify_ideal_loop = false;
1053 _exception_backedge = false;
1054 _type_verify = nullptr;
1055 #endif
1056
1057 // Globally visible Nodes
1058 // First set TOP to null to give safe behavior during creation of RootNode
1059 set_cached_top_node(nullptr);
1060 set_root(new RootNode());
1061 // Now that you have a Root to point to, create the real TOP
1062 set_cached_top_node( new ConNode(Type::TOP) );
1063 set_recent_alloc(nullptr, nullptr);
1064
1065 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1066 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1067 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1068 env()->set_dependencies(new Dependencies(env()));
1069
1070 _fixed_slots = 0;
1071 set_has_split_ifs(false);
1072 set_has_loops(false); // first approximation
1073 set_has_stringbuilder(false);
1074 set_has_boxed_value(false);
1075 _trap_can_recompile = false; // no traps emitted yet
1076 _major_progress = true; // start out assuming good things will happen
1077 set_has_unsafe_access(false);
1078 set_max_vector_size(0);
1079 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1080 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1081 set_decompile_count(0);
1082
1083 #ifndef PRODUCT
1084 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter));
1085 #endif
1086
1087 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1088 _loop_opts_cnt = LoopOptsCount;
1089 _has_flat_accesses = false;
1090 _flat_accesses_share_alias = true;
1091 _scalarize_in_safepoints = false;
1092
1093 set_do_inlining(Inline);
1094 set_max_inline_size(MaxInlineSize);
1095 set_freq_inline_size(FreqInlineSize);
1096 set_do_scheduling(OptoScheduling);
1097
1098 set_do_vector_loop(false);
1099 set_has_monitors(false);
1100 set_has_scoped_access(false);
1101
1102 if (AllowVectorizeOnDemand) {
1103 if (has_method() && _directive->VectorizeOption) {
1104 set_do_vector_loop(true);
1105 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());})
1106 } else if (has_method() && method()->name() != nullptr &&
1107 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1108 set_do_vector_loop(true);
1109 }
1110 }
1111 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1112 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());})
1113
1114 _max_node_limit = _directive->MaxNodeLimitOption;
1115
1116 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1117 set_clinit_barrier_on_entry(true);
1118 }
1119 if (debug_info()->recording_non_safepoints()) {
1120 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1121 (comp_arena(), 8, 0, nullptr));
1122 set_default_node_notes(Node_Notes::make(this));
1123 }
1124
1125 const int grow_ats = 16;
1126 _max_alias_types = grow_ats;
1127 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1128 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1129 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1130 {
1131 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1132 }
1133 // Initialize the first few types.
1134 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr);
1135 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1136 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1137 _num_alias_types = AliasIdxRaw+1;
1138 // Zero out the alias type cache.
1139 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1140 // A null adr_type hits in the cache right away. Preload the right answer.
1141 probe_alias_cache(nullptr)->_index = AliasIdxTop;
1142 }
1143
1144 #ifdef ASSERT
1145 // Verify that the current StartNode is valid.
1146 void Compile::verify_start(StartNode* s) const {
1147 assert(failing_internal() || s == start(), "should be StartNode");
1148 }
1149 #endif
1150
1151 /**
1152 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1153 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1154 * the ideal graph.
1155 */
1156 StartNode* Compile::start() const {
1157 assert (!failing_internal() || C->failure_is_artificial(), "Must not have pending failure. Reason is: %s", failure_reason());
1158 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1159 Node* start = root()->fast_out(i);
1160 if (start->is_Start()) {
1161 return start->as_Start();
1162 }
1163 }
1164 fatal("Did not find Start node!");
1165 return nullptr;
1166 }
1167
1168 //-------------------------------immutable_memory-------------------------------------
1169 // Access immutable memory
1170 Node* Compile::immutable_memory() {
1171 if (_immutable_memory != nullptr) {
1172 return _immutable_memory;
1173 }
1174 StartNode* s = start();
1175 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1176 Node *p = s->fast_out(i);
1177 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1178 _immutable_memory = p;
1179 return _immutable_memory;
1180 }
1181 }
1182 ShouldNotReachHere();
1183 return nullptr;
1184 }
1185
1186 //----------------------set_cached_top_node------------------------------------
1187 // Install the cached top node, and make sure Node::is_top works correctly.
1188 void Compile::set_cached_top_node(Node* tn) {
1189 if (tn != nullptr) verify_top(tn);
1190 Node* old_top = _top;
1191 _top = tn;
1192 // Calling Node::setup_is_top allows the nodes the chance to adjust
1193 // their _out arrays.
1194 if (_top != nullptr) _top->setup_is_top();
1195 if (old_top != nullptr) old_top->setup_is_top();
1196 assert(_top == nullptr || top()->is_top(), "");
1197 }
1198
1199 #ifdef ASSERT
1200 uint Compile::count_live_nodes_by_graph_walk() {
1201 Unique_Node_List useful(comp_arena());
1202 // Get useful node list by walking the graph.
1203 identify_useful_nodes(useful);
1204 return useful.size();
1205 }
1206
1207 void Compile::print_missing_nodes() {
1208
1209 // Return if CompileLog is null and PrintIdealNodeCount is false.
1210 if ((_log == nullptr) && (! PrintIdealNodeCount)) {
1211 return;
1212 }
1213
1214 // This is an expensive function. It is executed only when the user
1215 // specifies VerifyIdealNodeCount option or otherwise knows the
1216 // additional work that needs to be done to identify reachable nodes
1217 // by walking the flow graph and find the missing ones using
1218 // _dead_node_list.
1219
1220 Unique_Node_List useful(comp_arena());
1221 // Get useful node list by walking the graph.
1222 identify_useful_nodes(useful);
1223
1224 uint l_nodes = C->live_nodes();
1225 uint l_nodes_by_walk = useful.size();
1226
1227 if (l_nodes != l_nodes_by_walk) {
1228 if (_log != nullptr) {
1229 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1230 _log->stamp();
1231 _log->end_head();
1232 }
1233 VectorSet& useful_member_set = useful.member_set();
1234 int last_idx = l_nodes_by_walk;
1235 for (int i = 0; i < last_idx; i++) {
1236 if (useful_member_set.test(i)) {
1237 if (_dead_node_list.test(i)) {
1238 if (_log != nullptr) {
1239 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1240 }
1241 if (PrintIdealNodeCount) {
1242 // Print the log message to tty
1243 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1244 useful.at(i)->dump();
1245 }
1246 }
1247 }
1248 else if (! _dead_node_list.test(i)) {
1249 if (_log != nullptr) {
1250 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1251 }
1252 if (PrintIdealNodeCount) {
1253 // Print the log message to tty
1254 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1255 }
1256 }
1257 }
1258 if (_log != nullptr) {
1259 _log->tail("mismatched_nodes");
1260 }
1261 }
1262 }
1263 void Compile::record_modified_node(Node* n) {
1264 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) {
1265 _modified_nodes->push(n);
1266 }
1267 }
1268
1269 void Compile::remove_modified_node(Node* n) {
1270 if (_modified_nodes != nullptr) {
1271 _modified_nodes->remove(n);
1272 }
1273 }
1274 #endif
1275
1276 #ifndef PRODUCT
1277 void Compile::verify_top(Node* tn) const {
1278 if (tn != nullptr) {
1279 assert(tn->is_Con(), "top node must be a constant");
1280 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1281 assert(tn->in(0) != nullptr, "must have live top node");
1282 }
1283 }
1284 #endif
1285
1286
1287 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1288
1289 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1290 guarantee(arr != nullptr, "");
1291 int num_blocks = arr->length();
1292 if (grow_by < num_blocks) grow_by = num_blocks;
1293 int num_notes = grow_by * _node_notes_block_size;
1294 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1295 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1296 while (num_notes > 0) {
1297 arr->append(notes);
1298 notes += _node_notes_block_size;
1299 num_notes -= _node_notes_block_size;
1300 }
1301 assert(num_notes == 0, "exact multiple, please");
1302 }
1303
1304 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1305 if (source == nullptr || dest == nullptr) return false;
1306
1307 if (dest->is_Con())
1308 return false; // Do not push debug info onto constants.
1309
1310 #ifdef ASSERT
1311 // Leave a bread crumb trail pointing to the original node:
1312 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) {
1313 dest->set_debug_orig(source);
1314 }
1315 #endif
1316
1317 if (node_note_array() == nullptr)
1318 return false; // Not collecting any notes now.
1319
1320 // This is a copy onto a pre-existing node, which may already have notes.
1321 // If both nodes have notes, do not overwrite any pre-existing notes.
1322 Node_Notes* source_notes = node_notes_at(source->_idx);
1323 if (source_notes == nullptr || source_notes->is_clear()) return false;
1324 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1325 if (dest_notes == nullptr || dest_notes->is_clear()) {
1326 return set_node_notes_at(dest->_idx, source_notes);
1327 }
1328
1329 Node_Notes merged_notes = (*source_notes);
1330 // The order of operations here ensures that dest notes will win...
1331 merged_notes.update_from(dest_notes);
1332 return set_node_notes_at(dest->_idx, &merged_notes);
1333 }
1334
1335
1336 //--------------------------allow_range_check_smearing-------------------------
1337 // Gating condition for coalescing similar range checks.
1338 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1339 // single covering check that is at least as strong as any of them.
1340 // If the optimization succeeds, the simplified (strengthened) range check
1341 // will always succeed. If it fails, we will deopt, and then give up
1342 // on the optimization.
1343 bool Compile::allow_range_check_smearing() const {
1344 // If this method has already thrown a range-check,
1345 // assume it was because we already tried range smearing
1346 // and it failed.
1347 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1348 return !already_trapped;
1349 }
1350
1351
1352 //------------------------------flatten_alias_type-----------------------------
1353 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1354 assert(do_aliasing(), "Aliasing should be enabled");
1355 int offset = tj->offset();
1356 TypePtr::PTR ptr = tj->ptr();
1357
1358 // Known instance (scalarizable allocation) alias only with itself.
1359 bool is_known_inst = tj->isa_oopptr() != nullptr &&
1360 tj->is_oopptr()->is_known_instance();
1361
1362 // Process weird unsafe references.
1363 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1364 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1365 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1366 tj = TypeOopPtr::BOTTOM;
1367 ptr = tj->ptr();
1368 offset = tj->offset();
1369 }
1370
1371 // Array pointers need some flattening
1372 const TypeAryPtr* ta = tj->isa_aryptr();
1373 if (ta && ta->is_stable()) {
1374 // Erase stability property for alias analysis.
1375 tj = ta = ta->cast_to_stable(false);
1376 }
1377 if (ta && ta->is_not_flat()) {
1378 // Erase not flat property for alias analysis.
1379 tj = ta = ta->cast_to_not_flat(false);
1380 }
1381 if (ta && ta->is_not_null_free()) {
1382 // Erase not null free property for alias analysis.
1383 tj = ta = ta->cast_to_not_null_free(false);
1384 }
1385
1386 if( ta && is_known_inst ) {
1387 if ( offset != Type::OffsetBot &&
1388 offset > arrayOopDesc::length_offset_in_bytes() ) {
1389 offset = Type::OffsetBot; // Flatten constant access into array body only
1390 tj = ta = ta->
1391 remove_speculative()->
1392 cast_to_ptr_type(ptr)->
1393 with_offset(offset);
1394 }
1395 } else if (ta) {
1396 // For arrays indexed by constant indices, we flatten the alias
1397 // space to include all of the array body. Only the header, klass
1398 // and array length can be accessed un-aliased.
1399 // For flat inline type array, each field has its own slice so
1400 // we must include the field offset.
1401 if( offset != Type::OffsetBot ) {
1402 if( ta->const_oop() ) { // MethodData* or Method*
1403 offset = Type::OffsetBot; // Flatten constant access into array body
1404 tj = ta = ta->
1405 remove_speculative()->
1406 cast_to_ptr_type(ptr)->
1407 cast_to_exactness(false)->
1408 with_offset(offset);
1409 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1410 // range is OK as-is.
1411 tj = ta = TypeAryPtr::RANGE;
1412 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1413 tj = TypeInstPtr::KLASS; // all klass loads look alike
1414 ta = TypeAryPtr::RANGE; // generic ignored junk
1415 ptr = TypePtr::BotPTR;
1416 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1417 tj = TypeInstPtr::MARK;
1418 ta = TypeAryPtr::RANGE; // generic ignored junk
1419 ptr = TypePtr::BotPTR;
1420 } else { // Random constant offset into array body
1421 offset = Type::OffsetBot; // Flatten constant access into array body
1422 tj = ta = ta->
1423 remove_speculative()->
1424 cast_to_ptr_type(ptr)->
1425 cast_to_exactness(false)->
1426 with_offset(offset);
1427 }
1428 }
1429 // Arrays of fixed size alias with arrays of unknown size.
1430 if (ta->size() != TypeInt::POS) {
1431 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1432 tj = ta = ta->
1433 remove_speculative()->
1434 cast_to_ptr_type(ptr)->
1435 with_ary(tary)->
1436 cast_to_exactness(false);
1437 }
1438 // Arrays of known objects become arrays of unknown objects.
1439 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1440 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1441 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1442 }
1443 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1444 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1445 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset());
1446 }
1447 // Initially all flattened array accesses share a single slice
1448 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) {
1449 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true);
1450 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1451 }
1452 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1453 // cannot be distinguished by bytecode alone.
1454 if (ta->elem() == TypeInt::BOOL) {
1455 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1456 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1457 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1458 }
1459 // During the 2nd round of IterGVN, NotNull castings are removed.
1460 // Make sure the Bottom and NotNull variants alias the same.
1461 // Also, make sure exact and non-exact variants alias the same.
1462 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) {
1463 tj = ta = ta->
1464 remove_speculative()->
1465 cast_to_ptr_type(TypePtr::BotPTR)->
1466 cast_to_exactness(false)->
1467 with_offset(offset);
1468 }
1469 }
1470
1471 // Oop pointers need some flattening
1472 const TypeInstPtr *to = tj->isa_instptr();
1473 if (to && to != TypeOopPtr::BOTTOM) {
1474 ciInstanceKlass* ik = to->instance_klass();
1475 if( ptr == TypePtr::Constant ) {
1476 if (ik != ciEnv::current()->Class_klass() ||
1477 offset < ik->layout_helper_size_in_bytes()) {
1478 // No constant oop pointers (such as Strings); they alias with
1479 // unknown strings.
1480 assert(!is_known_inst, "not scalarizable allocation");
1481 tj = to = to->
1482 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1483 remove_speculative()->
1484 cast_to_ptr_type(TypePtr::BotPTR)->
1485 cast_to_exactness(false);
1486 }
1487 } else if( is_known_inst ) {
1488 tj = to; // Keep NotNull and klass_is_exact for instance type
1489 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1490 // During the 2nd round of IterGVN, NotNull castings are removed.
1491 // Make sure the Bottom and NotNull variants alias the same.
1492 // Also, make sure exact and non-exact variants alias the same.
1493 tj = to = to->
1494 remove_speculative()->
1495 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1496 cast_to_ptr_type(TypePtr::BotPTR)->
1497 cast_to_exactness(false);
1498 }
1499 if (to->speculative() != nullptr) {
1500 tj = to = to->remove_speculative();
1501 }
1502 // Canonicalize the holder of this field
1503 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1504 // First handle header references such as a LoadKlassNode, even if the
1505 // object's klass is unloaded at compile time (4965979).
1506 if (!is_known_inst) { // Do it only for non-instance types
1507 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset));
1508 }
1509 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1510 // Static fields are in the space above the normal instance
1511 // fields in the java.lang.Class instance.
1512 if (ik != ciEnv::current()->Class_klass()) {
1513 to = nullptr;
1514 tj = TypeOopPtr::BOTTOM;
1515 offset = tj->offset();
1516 }
1517 } else {
1518 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1519 assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1520 assert(tj->offset() == offset, "no change to offset expected");
1521 bool xk = to->klass_is_exact();
1522 int instance_id = to->instance_id();
1523
1524 // If the input type's class is the holder: if exact, the type only includes interfaces implemented by the holder
1525 // but if not exact, it may include extra interfaces: build new type from the holder class to make sure only
1526 // its interfaces are included.
1527 if (xk && ik->equals(canonical_holder)) {
1528 assert(tj == TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id), "exact type should be canonical type");
1529 } else {
1530 assert(xk || !is_known_inst, "Known instance should be exact type");
1531 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, is_known_inst, nullptr, Type::Offset(offset), instance_id);
1532 }
1533 }
1534 }
1535
1536 // Klass pointers to object array klasses need some flattening
1537 const TypeKlassPtr *tk = tj->isa_klassptr();
1538 if( tk ) {
1539 // If we are referencing a field within a Klass, we need
1540 // to assume the worst case of an Object. Both exact and
1541 // inexact types must flatten to the same alias class so
1542 // use NotNull as the PTR.
1543 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1544 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1545 env()->Object_klass(),
1546 Type::Offset(offset));
1547 }
1548
1549 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1550 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1551 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs
1552 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset));
1553 } else {
1554 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_flat(), tk->is_null_free());
1555 }
1556 }
1557 // Check for precise loads from the primary supertype array and force them
1558 // to the supertype cache alias index. Check for generic array loads from
1559 // the primary supertype array and also force them to the supertype cache
1560 // alias index. Since the same load can reach both, we need to merge
1561 // these 2 disparate memories into the same alias class. Since the
1562 // primary supertype array is read-only, there's no chance of confusion
1563 // where we bypass an array load and an array store.
1564 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1565 if (offset == Type::OffsetBot ||
1566 (offset >= primary_supers_offset &&
1567 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1568 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1569 offset = in_bytes(Klass::secondary_super_cache_offset());
1570 tj = tk = tk->with_offset(offset);
1571 }
1572 }
1573
1574 // Flatten all Raw pointers together.
1575 if (tj->base() == Type::RawPtr)
1576 tj = TypeRawPtr::BOTTOM;
1577
1578 if (tj->base() == Type::AnyPtr)
1579 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1580
1581 offset = tj->offset();
1582 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1583
1584 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1585 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1586 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1587 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1588 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1589 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1590 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1591 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1592 assert( tj->ptr() != TypePtr::TopPTR &&
1593 tj->ptr() != TypePtr::AnyNull &&
1594 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1595 // assert( tj->ptr() != TypePtr::Constant ||
1596 // tj->base() == Type::RawPtr ||
1597 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1598
1599 return tj;
1600 }
1601
1602 void Compile::AliasType::Init(int i, const TypePtr* at) {
1603 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1604 _index = i;
1605 _adr_type = at;
1606 _field = nullptr;
1607 _element = nullptr;
1608 _is_rewritable = true; // default
1609 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr;
1610 if (atoop != nullptr && atoop->is_known_instance()) {
1611 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1612 _general_index = Compile::current()->get_alias_index(gt);
1613 } else {
1614 _general_index = 0;
1615 }
1616 }
1617
1618 BasicType Compile::AliasType::basic_type() const {
1619 if (element() != nullptr) {
1620 const Type* element = adr_type()->is_aryptr()->elem();
1621 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1622 } if (field() != nullptr) {
1623 return field()->layout_type();
1624 } else {
1625 return T_ILLEGAL; // unknown
1626 }
1627 }
1628
1629 //---------------------------------print_on------------------------------------
1630 #ifndef PRODUCT
1631 void Compile::AliasType::print_on(outputStream* st) {
1632 if (index() < 10)
1633 st->print("@ <%d> ", index());
1634 else st->print("@ <%d>", index());
1635 st->print(is_rewritable() ? " " : " RO");
1636 int offset = adr_type()->offset();
1637 if (offset == Type::OffsetBot)
1638 st->print(" +any");
1639 else st->print(" +%-3d", offset);
1640 st->print(" in ");
1641 adr_type()->dump_on(st);
1642 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1643 if (field() != nullptr && tjp) {
1644 if (tjp->is_instptr()->instance_klass() != field()->holder() ||
1645 tjp->offset() != field()->offset_in_bytes()) {
1646 st->print(" != ");
1647 field()->print();
1648 st->print(" ***");
1649 }
1650 }
1651 }
1652
1653 void print_alias_types() {
1654 Compile* C = Compile::current();
1655 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1656 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1657 C->alias_type(idx)->print_on(tty);
1658 tty->cr();
1659 }
1660 }
1661 #endif
1662
1663
1664 //----------------------------probe_alias_cache--------------------------------
1665 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1666 intptr_t key = (intptr_t) adr_type;
1667 key ^= key >> logAliasCacheSize;
1668 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1669 }
1670
1671
1672 //-----------------------------grow_alias_types--------------------------------
1673 void Compile::grow_alias_types() {
1674 const int old_ats = _max_alias_types; // how many before?
1675 const int new_ats = old_ats; // how many more?
1676 const int grow_ats = old_ats+new_ats; // how many now?
1677 _max_alias_types = grow_ats;
1678 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1679 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1680 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1681 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1682 }
1683
1684
1685 //--------------------------------find_alias_type------------------------------
1686 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1687 if (!do_aliasing()) {
1688 return alias_type(AliasIdxBot);
1689 }
1690
1691 AliasCacheEntry* ace = nullptr;
1692 if (!uncached) {
1693 ace = probe_alias_cache(adr_type);
1694 if (ace->_adr_type == adr_type) {
1695 return alias_type(ace->_index);
1696 }
1697 }
1698
1699 // Handle special cases.
1700 if (adr_type == nullptr) return alias_type(AliasIdxTop);
1701 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1702
1703 // Do it the slow way.
1704 const TypePtr* flat = flatten_alias_type(adr_type);
1705
1706 #ifdef ASSERT
1707 {
1708 ResourceMark rm;
1709 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1710 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1711 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1712 Type::str(adr_type));
1713 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1714 const TypeOopPtr* foop = flat->is_oopptr();
1715 // Scalarizable allocations have exact klass always.
1716 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1717 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1718 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1719 Type::str(foop), Type::str(xoop));
1720 }
1721 }
1722 #endif
1723
1724 int idx = AliasIdxTop;
1725 for (int i = 0; i < num_alias_types(); i++) {
1726 if (alias_type(i)->adr_type() == flat) {
1727 idx = i;
1728 break;
1729 }
1730 }
1731
1732 if (idx == AliasIdxTop) {
1733 if (no_create) return nullptr;
1734 // Grow the array if necessary.
1735 if (_num_alias_types == _max_alias_types) grow_alias_types();
1736 // Add a new alias type.
1737 idx = _num_alias_types++;
1738 _alias_types[idx]->Init(idx, flat);
1739 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1740 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1741 if (flat->isa_instptr()) {
1742 if (flat->offset() == java_lang_Class::klass_offset()
1743 && flat->is_instptr()->instance_klass() == env()->Class_klass())
1744 alias_type(idx)->set_rewritable(false);
1745 }
1746 ciField* field = nullptr;
1747 if (flat->isa_aryptr()) {
1748 #ifdef ASSERT
1749 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1750 // (T_BYTE has the weakest alignment and size restrictions...)
1751 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1752 #endif
1753 const Type* elemtype = flat->is_aryptr()->elem();
1754 if (flat->offset() == TypePtr::OffsetBot) {
1755 alias_type(idx)->set_element(elemtype);
1756 }
1757 int field_offset = flat->is_aryptr()->field_offset().get();
1758 if (flat->is_flat() &&
1759 field_offset != Type::OffsetBot) {
1760 ciInlineKlass* vk = elemtype->inline_klass();
1761 field_offset += vk->payload_offset();
1762 field = vk->get_field_by_offset(field_offset, false);
1763 }
1764 }
1765 if (flat->isa_klassptr()) {
1766 if (UseCompactObjectHeaders) {
1767 if (flat->offset() == in_bytes(Klass::prototype_header_offset()))
1768 alias_type(idx)->set_rewritable(false);
1769 }
1770 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1771 alias_type(idx)->set_rewritable(false);
1772 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1773 alias_type(idx)->set_rewritable(false);
1774 if (flat->offset() == in_bytes(Klass::misc_flags_offset()))
1775 alias_type(idx)->set_rewritable(false);
1776 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1777 alias_type(idx)->set_rewritable(false);
1778 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1779 alias_type(idx)->set_rewritable(false);
1780 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1781 alias_type(idx)->set_rewritable(false);
1782 }
1783 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1784 // but the base pointer type is not distinctive enough to identify
1785 // references into JavaThread.)
1786
1787 // Check for final fields.
1788 const TypeInstPtr* tinst = flat->isa_instptr();
1789 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1790 if (tinst->const_oop() != nullptr &&
1791 tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1792 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1793 // static field
1794 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1795 field = k->get_field_by_offset(tinst->offset(), true);
1796 } else if (tinst->is_inlinetypeptr()) {
1797 // Inline type field
1798 ciInlineKlass* vk = tinst->inline_klass();
1799 field = vk->get_field_by_offset(tinst->offset(), false);
1800 } else {
1801 ciInstanceKlass *k = tinst->instance_klass();
1802 field = k->get_field_by_offset(tinst->offset(), false);
1803 }
1804 }
1805 assert(field == nullptr ||
1806 original_field == nullptr ||
1807 (field->holder() == original_field->holder() &&
1808 field->offset_in_bytes() == original_field->offset_in_bytes() &&
1809 field->is_static() == original_field->is_static()), "wrong field?");
1810 // Set field() and is_rewritable() attributes.
1811 if (field != nullptr) {
1812 alias_type(idx)->set_field(field);
1813 if (flat->isa_aryptr()) {
1814 // Fields of flat arrays are rewritable although they are declared final
1815 assert(flat->is_flat(), "must be a flat array");
1816 alias_type(idx)->set_rewritable(true);
1817 }
1818 }
1819 }
1820
1821 // Fill the cache for next time.
1822 if (!uncached) {
1823 ace->_adr_type = adr_type;
1824 ace->_index = idx;
1825 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1826
1827 // Might as well try to fill the cache for the flattened version, too.
1828 AliasCacheEntry* face = probe_alias_cache(flat);
1829 if (face->_adr_type == nullptr) {
1830 face->_adr_type = flat;
1831 face->_index = idx;
1832 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1833 }
1834 }
1835
1836 return alias_type(idx);
1837 }
1838
1839
1840 Compile::AliasType* Compile::alias_type(ciField* field) {
1841 const TypeOopPtr* t;
1842 if (field->is_static())
1843 t = TypeInstPtr::make(field->holder()->java_mirror());
1844 else
1845 t = TypeOopPtr::make_from_klass_raw(field->holder());
1846 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1847 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1848 return atp;
1849 }
1850
1851
1852 //------------------------------have_alias_type--------------------------------
1853 bool Compile::have_alias_type(const TypePtr* adr_type) {
1854 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1855 if (ace->_adr_type == adr_type) {
1856 return true;
1857 }
1858
1859 // Handle special cases.
1860 if (adr_type == nullptr) return true;
1861 if (adr_type == TypePtr::BOTTOM) return true;
1862
1863 return find_alias_type(adr_type, true, nullptr) != nullptr;
1864 }
1865
1866 //-----------------------------must_alias--------------------------------------
1867 // True if all values of the given address type are in the given alias category.
1868 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1869 if (alias_idx == AliasIdxBot) return true; // the universal category
1870 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP
1871 if (alias_idx == AliasIdxTop) return false; // the empty category
1872 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1873
1874 // the only remaining possible overlap is identity
1875 int adr_idx = get_alias_index(adr_type);
1876 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1877 assert(adr_idx == alias_idx ||
1878 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1879 && adr_type != TypeOopPtr::BOTTOM),
1880 "should not be testing for overlap with an unsafe pointer");
1881 return adr_idx == alias_idx;
1882 }
1883
1884 //------------------------------can_alias--------------------------------------
1885 // True if any values of the given address type are in the given alias category.
1886 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1887 if (alias_idx == AliasIdxTop) return false; // the empty category
1888 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP
1889 // Known instance doesn't alias with bottom memory
1890 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category
1891 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1892
1893 // the only remaining possible overlap is identity
1894 int adr_idx = get_alias_index(adr_type);
1895 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1896 return adr_idx == alias_idx;
1897 }
1898
1899 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their
1900 // uncommon traps if no Runtime Predicates were created from the Parse Predicates.
1901 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) {
1902 if (parse_predicate_count() == 0) {
1903 return;
1904 }
1905 for (int i = 0; i < parse_predicate_count(); i++) {
1906 ParsePredicateNode* parse_predicate = _parse_predicates.at(i);
1907 parse_predicate->mark_useless(igvn);
1908 }
1909 _parse_predicates.clear();
1910 }
1911
1912 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1913 if (!n->for_post_loop_opts_igvn()) {
1914 assert(!_for_post_loop_igvn.contains(n), "duplicate");
1915 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1916 _for_post_loop_igvn.append(n);
1917 }
1918 }
1919
1920 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1921 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1922 _for_post_loop_igvn.remove(n);
1923 }
1924
1925 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1926 // Verify that all previous optimizations produced a valid graph
1927 // at least to this point, even if no loop optimizations were done.
1928 PhaseIdealLoop::verify(igvn);
1929
1930 C->set_post_loop_opts_phase(); // no more loop opts allowed
1931
1932 assert(!C->major_progress(), "not cleared");
1933
1934 if (_for_post_loop_igvn.length() > 0) {
1935 while (_for_post_loop_igvn.length() > 0) {
1936 Node* n = _for_post_loop_igvn.pop();
1937 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1938 igvn._worklist.push(n);
1939 }
1940 igvn.optimize();
1941 if (failing()) return;
1942 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1943 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now");
1944
1945 // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1946 if (C->major_progress()) {
1947 C->clear_major_progress(); // ensure that major progress is now clear
1948 }
1949 }
1950 }
1951
1952 void Compile::add_inline_type(Node* n) {
1953 assert(n->is_InlineType(), "unexpected node");
1954 _inline_type_nodes.push(n);
1955 }
1956
1957 void Compile::remove_inline_type(Node* n) {
1958 assert(n->is_InlineType(), "unexpected node");
1959 if (_inline_type_nodes.contains(n)) {
1960 _inline_type_nodes.remove(n);
1961 }
1962 }
1963
1964 // Does the return value keep otherwise useless inline type allocations alive?
1965 static bool return_val_keeps_allocations_alive(Node* ret_val) {
1966 ResourceMark rm;
1967 Unique_Node_List wq;
1968 wq.push(ret_val);
1969 bool some_allocations = false;
1970 for (uint i = 0; i < wq.size(); i++) {
1971 Node* n = wq.at(i);
1972 if (n->outcnt() > 1) {
1973 // Some other use for the allocation
1974 return false;
1975 } else if (n->is_InlineType()) {
1976 wq.push(n->in(1));
1977 } else if (n->is_Phi()) {
1978 for (uint j = 1; j < n->req(); j++) {
1979 wq.push(n->in(j));
1980 }
1981 } else if (n->is_CheckCastPP() &&
1982 n->in(1)->is_Proj() &&
1983 n->in(1)->in(0)->is_Allocate()) {
1984 some_allocations = true;
1985 } else if (n->is_CheckCastPP()) {
1986 wq.push(n->in(1));
1987 }
1988 }
1989 return some_allocations;
1990 }
1991
1992 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) {
1993 // Make sure that the return value does not keep an otherwise unused allocation alive
1994 if (tf()->returns_inline_type_as_fields()) {
1995 Node* ret = nullptr;
1996 for (uint i = 1; i < root()->req(); i++) {
1997 Node* in = root()->in(i);
1998 if (in->Opcode() == Op_Return) {
1999 assert(ret == nullptr, "only one return");
2000 ret = in;
2001 }
2002 }
2003 if (ret != nullptr) {
2004 Node* ret_val = ret->in(TypeFunc::Parms);
2005 if (igvn.type(ret_val)->isa_oopptr() &&
2006 return_val_keeps_allocations_alive(ret_val)) {
2007 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn));
2008 assert(ret_val->outcnt() == 0, "should be dead now");
2009 igvn.remove_dead_node(ret_val);
2010 }
2011 }
2012 }
2013 if (_inline_type_nodes.length() == 0) {
2014 return;
2015 }
2016 // Scalarize inline types in safepoint debug info.
2017 // Delay this until all inlining is over to avoid getting inconsistent debug info.
2018 set_scalarize_in_safepoints(true);
2019 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) {
2020 InlineTypeNode* vt = _inline_type_nodes.at(i)->as_InlineType();
2021 vt->make_scalar_in_safepoints(&igvn);
2022 igvn.record_for_igvn(vt);
2023 }
2024 if (remove) {
2025 // Remove inline type nodes by replacing them with their oop input
2026 while (_inline_type_nodes.length() > 0) {
2027 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType();
2028 if (vt->outcnt() == 0) {
2029 igvn.remove_dead_node(vt);
2030 continue;
2031 }
2032 for (DUIterator i = vt->outs(); vt->has_out(i); i++) {
2033 DEBUG_ONLY(bool must_be_buffered = false);
2034 Node* u = vt->out(i);
2035 // Check if any users are blackholes. If so, rewrite them to use either the
2036 // allocated buffer, or individual components, instead of the inline type node
2037 // that goes away.
2038 if (u->is_Blackhole()) {
2039 BlackholeNode* bh = u->as_Blackhole();
2040
2041 // Unlink the old input
2042 int idx = bh->find_edge(vt);
2043 assert(idx != -1, "The edge should be there");
2044 bh->del_req(idx);
2045 --i;
2046
2047 if (vt->is_allocated(&igvn)) {
2048 // Already has the allocated instance, blackhole that
2049 bh->add_req(vt->get_oop());
2050 } else {
2051 // Not allocated yet, blackhole the components
2052 for (uint c = 0; c < vt->field_count(); c++) {
2053 bh->add_req(vt->field_value(c));
2054 }
2055 }
2056
2057 // Node modified, record for IGVN
2058 igvn.record_for_igvn(bh);
2059 }
2060 #ifdef ASSERT
2061 // Verify that inline type is buffered when replacing by oop
2062 else if (u->is_InlineType()) {
2063 // InlineType uses don't need buffering because they are about to be replaced as well
2064 } else if (u->is_Phi()) {
2065 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through
2066 } else {
2067 must_be_buffered = true;
2068 }
2069 if (must_be_buffered && !vt->is_allocated(&igvn)) {
2070 vt->dump(0);
2071 u->dump(0);
2072 assert(false, "Should have been buffered");
2073 }
2074 #endif
2075 }
2076 igvn.replace_node(vt, vt->get_oop());
2077 }
2078 }
2079 igvn.optimize();
2080 }
2081
2082 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) {
2083 if (!_has_flat_accesses) {
2084 return;
2085 }
2086 // Initially, all flat array accesses share the same slice to
2087 // keep dependencies with Object[] array accesses (that could be
2088 // to a flat array) correct. We're done with parsing so we
2089 // now know all flat array accesses in this compile
2090 // unit. Let's move flat array accesses to their own slice,
2091 // one per element field. This should help memory access
2092 // optimizations.
2093 ResourceMark rm;
2094 Unique_Node_List wq;
2095 wq.push(root());
2096
2097 Node_List mergememnodes;
2098 Node_List memnodes;
2099
2100 // Alias index currently shared by all flat memory accesses
2101 int index = get_alias_index(TypeAryPtr::INLINES);
2102
2103 // Find MergeMem nodes and flat array accesses
2104 for (uint i = 0; i < wq.size(); i++) {
2105 Node* n = wq.at(i);
2106 if (n->is_Mem()) {
2107 const TypePtr* adr_type = nullptr;
2108 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2109 if (adr_type == TypeAryPtr::INLINES) {
2110 memnodes.push(n);
2111 }
2112 } else if (n->is_MergeMem()) {
2113 MergeMemNode* mm = n->as_MergeMem();
2114 if (mm->memory_at(index) != mm->base_memory()) {
2115 mergememnodes.push(n);
2116 }
2117 }
2118 for (uint j = 0; j < n->req(); j++) {
2119 Node* m = n->in(j);
2120 if (m != nullptr) {
2121 wq.push(m);
2122 }
2123 }
2124 }
2125
2126 if (memnodes.size() > 0) {
2127 _flat_accesses_share_alias = false;
2128
2129 // We are going to change the slice for the flat array
2130 // accesses so we need to clear the cache entries that refer to
2131 // them.
2132 for (uint i = 0; i < AliasCacheSize; i++) {
2133 AliasCacheEntry* ace = &_alias_cache[i];
2134 if (ace->_adr_type != nullptr &&
2135 ace->_adr_type->is_flat()) {
2136 ace->_adr_type = nullptr;
2137 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop
2138 }
2139 }
2140
2141 // Find what aliases we are going to add
2142 int start_alias = num_alias_types()-1;
2143 int stop_alias = 0;
2144
2145 for (uint i = 0; i < memnodes.size(); i++) {
2146 Node* m = memnodes.at(i);
2147 const TypePtr* adr_type = nullptr;
2148 adr_type = m->adr_type();
2149 #ifdef ASSERT
2150 m->as_Mem()->set_adr_type(adr_type);
2151 #endif
2152 int idx = get_alias_index(adr_type);
2153 start_alias = MIN2(start_alias, idx);
2154 stop_alias = MAX2(stop_alias, idx);
2155 }
2156
2157 assert(stop_alias >= start_alias, "should have expanded aliases");
2158
2159 Node_Stack stack(0);
2160 #ifdef ASSERT
2161 VectorSet seen(Thread::current()->resource_area());
2162 #endif
2163 // Now let's fix the memory graph so each flat array access
2164 // is moved to the right slice. Start from the MergeMem nodes.
2165 uint last = unique();
2166 for (uint i = 0; i < mergememnodes.size(); i++) {
2167 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2168 Node* n = current->memory_at(index);
2169 MergeMemNode* mm = nullptr;
2170 do {
2171 // Follow memory edges through memory accesses, phis and
2172 // narrow membars and push nodes on the stack. Once we hit
2173 // bottom memory, we pop element off the stack one at a
2174 // time, in reverse order, and move them to the right slice
2175 // by changing their memory edges.
2176 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) {
2177 assert(!seen.test_set(n->_idx), "");
2178 // Uses (a load for instance) will need to be moved to the
2179 // right slice as well and will get a new memory state
2180 // that we don't know yet. The use could also be the
2181 // backedge of a loop. We put a place holder node between
2182 // the memory node and its uses. We replace that place
2183 // holder with the correct memory state once we know it,
2184 // i.e. when nodes are popped off the stack. Using the
2185 // place holder make the logic work in the presence of
2186 // loops.
2187 if (n->outcnt() > 1) {
2188 Node* place_holder = nullptr;
2189 assert(!n->has_out_with(Op_Node), "");
2190 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2191 Node* u = n->out(k);
2192 if (u != current && u->_idx < last) {
2193 bool success = false;
2194 for (uint l = 0; l < u->req(); l++) {
2195 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2196 continue;
2197 }
2198 Node* in = u->in(l);
2199 if (in == n) {
2200 if (place_holder == nullptr) {
2201 place_holder = new Node(1);
2202 place_holder->init_req(0, n);
2203 }
2204 igvn.replace_input_of(u, l, place_holder);
2205 success = true;
2206 }
2207 }
2208 if (success) {
2209 --k;
2210 }
2211 }
2212 }
2213 }
2214 if (n->is_Phi()) {
2215 stack.push(n, 1);
2216 n = n->in(1);
2217 } else if (n->is_Mem()) {
2218 stack.push(n, n->req());
2219 n = n->in(MemNode::Memory);
2220 } else {
2221 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2222 stack.push(n, n->req());
2223 n = n->in(0)->in(TypeFunc::Memory);
2224 }
2225 } else {
2226 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2227 // Build a new MergeMem node to carry the new memory state
2228 // as we build it. IGVN should fold extraneous MergeMem
2229 // nodes.
2230 mm = MergeMemNode::make(n);
2231 igvn.register_new_node_with_optimizer(mm);
2232 while (stack.size() > 0) {
2233 Node* m = stack.node();
2234 uint idx = stack.index();
2235 if (m->is_Mem()) {
2236 // Move memory node to its new slice
2237 const TypePtr* adr_type = m->adr_type();
2238 int alias = get_alias_index(adr_type);
2239 Node* prev = mm->memory_at(alias);
2240 igvn.replace_input_of(m, MemNode::Memory, prev);
2241 mm->set_memory_at(alias, m);
2242 } else if (m->is_Phi()) {
2243 // We need as many new phis as there are new aliases
2244 igvn.replace_input_of(m, idx, mm);
2245 if (idx == m->req()-1) {
2246 Node* r = m->in(0);
2247 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2248 const TypePtr* adr_type = get_adr_type(j);
2249 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2250 continue;
2251 }
2252 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2253 igvn.register_new_node_with_optimizer(phi);
2254 for (uint k = 1; k < m->req(); k++) {
2255 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2256 }
2257 mm->set_memory_at(j, phi);
2258 }
2259 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2260 igvn.register_new_node_with_optimizer(base_phi);
2261 for (uint k = 1; k < m->req(); k++) {
2262 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2263 }
2264 mm->set_base_memory(base_phi);
2265 }
2266 } else {
2267 // This is a MemBarCPUOrder node from
2268 // Parse::array_load()/Parse::array_store(), in the
2269 // branch that handles flat arrays hidden under
2270 // an Object[] array. We also need one new membar per
2271 // new alias to keep the unknown access that the
2272 // membars protect properly ordered with accesses to
2273 // known flat array.
2274 assert(m->is_Proj(), "projection expected");
2275 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2276 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2277 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2278 const TypePtr* adr_type = get_adr_type(j);
2279 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) {
2280 continue;
2281 }
2282 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr);
2283 igvn.register_new_node_with_optimizer(mb);
2284 Node* mem = mm->memory_at(j);
2285 mb->init_req(TypeFunc::Control, ctrl);
2286 mb->init_req(TypeFunc::Memory, mem);
2287 ctrl = new ProjNode(mb, TypeFunc::Control);
2288 igvn.register_new_node_with_optimizer(ctrl);
2289 mem = new ProjNode(mb, TypeFunc::Memory);
2290 igvn.register_new_node_with_optimizer(mem);
2291 mm->set_memory_at(j, mem);
2292 }
2293 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2294 }
2295 if (idx < m->req()-1) {
2296 idx += 1;
2297 stack.set_index(idx);
2298 n = m->in(idx);
2299 break;
2300 }
2301 // Take care of place holder nodes
2302 if (m->has_out_with(Op_Node)) {
2303 Node* place_holder = m->find_out_with(Op_Node);
2304 if (place_holder != nullptr) {
2305 Node* mm_clone = mm->clone();
2306 igvn.register_new_node_with_optimizer(mm_clone);
2307 Node* hook = new Node(1);
2308 hook->init_req(0, mm);
2309 igvn.replace_node(place_holder, mm_clone);
2310 hook->destruct(&igvn);
2311 }
2312 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2313 }
2314 stack.pop();
2315 }
2316 }
2317 } while(stack.size() > 0);
2318 // Fix the memory state at the MergeMem we started from
2319 igvn.rehash_node_delayed(current);
2320 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2321 const TypePtr* adr_type = get_adr_type(j);
2322 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) {
2323 continue;
2324 }
2325 current->set_memory_at(j, mm);
2326 }
2327 current->set_memory_at(index, current->base_memory());
2328 }
2329 igvn.optimize();
2330 }
2331 print_method(PHASE_SPLIT_INLINES_ARRAY, 2);
2332 #ifdef ASSERT
2333 if (!_flat_accesses_share_alias) {
2334 wq.clear();
2335 wq.push(root());
2336 for (uint i = 0; i < wq.size(); i++) {
2337 Node* n = wq.at(i);
2338 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph");
2339 for (uint j = 0; j < n->req(); j++) {
2340 Node* m = n->in(j);
2341 if (m != nullptr) {
2342 wq.push(m);
2343 }
2344 }
2345 }
2346 }
2347 #endif
2348 }
2349
2350 void Compile::record_for_merge_stores_igvn(Node* n) {
2351 if (!n->for_merge_stores_igvn()) {
2352 assert(!_for_merge_stores_igvn.contains(n), "duplicate");
2353 n->add_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2354 _for_merge_stores_igvn.append(n);
2355 }
2356 }
2357
2358 void Compile::remove_from_merge_stores_igvn(Node* n) {
2359 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2360 _for_merge_stores_igvn.remove(n);
2361 }
2362
2363 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
2364 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
2365 // the stores, and we merge the wrong sequence of stores.
2366 // Example:
2367 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
2368 // Apply MergeStores:
2369 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
2370 // Remove more RangeChecks:
2371 // StoreI [ StoreL ] StoreI
2372 // But now it would have been better to do this instead:
2373 // [ StoreL ] [ StoreL ]
2374 //
2375 // Note: we allow stores to merge in this dedicated IGVN round, and any later IGVN round,
2376 // since we never unset _merge_stores_phase.
2377 void Compile::process_for_merge_stores_igvn(PhaseIterGVN& igvn) {
2378 C->set_merge_stores_phase();
2379
2380 if (_for_merge_stores_igvn.length() > 0) {
2381 while (_for_merge_stores_igvn.length() > 0) {
2382 Node* n = _for_merge_stores_igvn.pop();
2383 n->remove_flag(Node::NodeFlags::Flag_for_merge_stores_igvn);
2384 igvn._worklist.push(n);
2385 }
2386 igvn.optimize();
2387 if (failing()) return;
2388 assert(_for_merge_stores_igvn.length() == 0, "no more delayed nodes allowed");
2389 print_method(PHASE_AFTER_MERGE_STORES, 3);
2390 }
2391 }
2392
2393 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
2394 if (OptimizeUnstableIf) {
2395 _unstable_if_traps.append(trap);
2396 }
2397 }
2398
2399 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
2400 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
2401 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2402 Node* n = trap->uncommon_trap();
2403 if (!useful.member(n)) {
2404 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
2405 }
2406 }
2407 }
2408
2409 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
2410 // or fold-compares case. Return true if succeed or not found.
2411 //
2412 // In rare cases, the found trap has been processed. It is too late to delete it. Return
2413 // false and ask fold-compares to yield.
2414 //
2415 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
2416 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
2417 // when deoptimization does happen.
2418 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
2419 for (int i = 0; i < _unstable_if_traps.length(); ++i) {
2420 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2421 if (trap->uncommon_trap() == unc) {
2422 if (yield && trap->modified()) {
2423 return false;
2424 }
2425 _unstable_if_traps.delete_at(i);
2426 break;
2427 }
2428 }
2429 return true;
2430 }
2431
2432 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
2433 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
2434 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
2435 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
2436 UnstableIfTrap* trap = _unstable_if_traps.at(i);
2437 CallStaticJavaNode* unc = trap->uncommon_trap();
2438 int next_bci = trap->next_bci();
2439 bool modified = trap->modified();
2440
2441 if (next_bci != -1 && !modified) {
2442 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
2443 JVMState* jvms = unc->jvms();
2444 ciMethod* method = jvms->method();
2445 ciBytecodeStream iter(method);
2446
2447 iter.force_bci(jvms->bci());
2448 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
2449 Bytecodes::Code c = iter.cur_bc();
2450 Node* lhs = nullptr;
2451 Node* rhs = nullptr;
2452 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
2453 lhs = unc->peek_operand(0);
2454 rhs = unc->peek_operand(1);
2455 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
2456 lhs = unc->peek_operand(0);
2457 }
2458
2459 ResourceMark rm;
2460 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
2461 assert(live_locals.is_valid(), "broken liveness info");
2462 int len = (int)live_locals.size();
2463
2464 for (int i = 0; i < len; i++) {
2465 Node* local = unc->local(jvms, i);
2466 // kill local using the liveness of next_bci.
2467 // give up when the local looks like an operand to secure reexecution.
2468 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) {
2469 uint idx = jvms->locoff() + i;
2470 #ifdef ASSERT
2471 if (PrintOpto && Verbose) {
2472 tty->print("[unstable_if] kill local#%d: ", idx);
2473 local->dump();
2474 tty->cr();
2475 }
2476 #endif
2477 igvn.replace_input_of(unc, idx, top());
2478 modified = true;
2479 }
2480 }
2481 }
2482
2483 // keep the modified trap for late query
2484 if (modified) {
2485 trap->set_modified();
2486 } else {
2487 _unstable_if_traps.delete_at(i);
2488 }
2489 }
2490 igvn.optimize();
2491 }
2492
2493 // StringOpts and late inlining of string methods
2494 void Compile::inline_string_calls(bool parse_time) {
2495 {
2496 // remove useless nodes to make the usage analysis simpler
2497 ResourceMark rm;
2498 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2499 }
2500
2501 {
2502 ResourceMark rm;
2503 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2504 PhaseStringOpts pso(initial_gvn());
2505 print_method(PHASE_AFTER_STRINGOPTS, 3);
2506 }
2507
2508 // now inline anything that we skipped the first time around
2509 if (!parse_time) {
2510 _late_inlines_pos = _late_inlines.length();
2511 }
2512
2513 while (_string_late_inlines.length() > 0) {
2514 CallGenerator* cg = _string_late_inlines.pop();
2515 cg->do_late_inline();
2516 if (failing()) return;
2517 }
2518 _string_late_inlines.trunc_to(0);
2519 }
2520
2521 // Late inlining of boxing methods
2522 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2523 if (_boxing_late_inlines.length() > 0) {
2524 assert(has_boxed_value(), "inconsistent");
2525
2526 set_inlining_incrementally(true);
2527
2528 igvn_worklist()->ensure_empty(); // should be done with igvn
2529
2530 _late_inlines_pos = _late_inlines.length();
2531
2532 while (_boxing_late_inlines.length() > 0) {
2533 CallGenerator* cg = _boxing_late_inlines.pop();
2534 cg->do_late_inline();
2535 if (failing()) return;
2536 }
2537 _boxing_late_inlines.trunc_to(0);
2538
2539 inline_incrementally_cleanup(igvn);
2540
2541 set_inlining_incrementally(false);
2542 }
2543 }
2544
2545 bool Compile::inline_incrementally_one() {
2546 assert(IncrementalInline, "incremental inlining should be on");
2547
2548 TracePhase tp(_t_incrInline_inline);
2549
2550 set_inlining_progress(false);
2551 set_do_cleanup(false);
2552
2553 for (int i = 0; i < _late_inlines.length(); i++) {
2554 _late_inlines_pos = i+1;
2555 CallGenerator* cg = _late_inlines.at(i);
2556 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2557 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2558 cg->do_late_inline();
2559 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2560 if (failing()) {
2561 return false;
2562 } else if (inlining_progress()) {
2563 _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2564 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2565 break; // process one call site at a time
2566 }
2567 } else {
2568 // Ignore late inline direct calls when inlining is not allowed.
2569 // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2570 }
2571 }
2572 // Remove processed elements.
2573 _late_inlines.remove_till(_late_inlines_pos);
2574 _late_inlines_pos = 0;
2575
2576 assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2577
2578 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2579
2580 set_inlining_progress(false);
2581 set_do_cleanup(false);
2582
2583 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2584 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2585 }
2586
2587 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2588 {
2589 TracePhase tp(_t_incrInline_pru);
2590 ResourceMark rm;
2591 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist());
2592 }
2593 {
2594 TracePhase tp(_t_incrInline_igvn);
2595 igvn.reset_from_gvn(initial_gvn());
2596 igvn.optimize();
2597 if (failing()) return;
2598 }
2599 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2600 }
2601
2602 // Perform incremental inlining until bound on number of live nodes is reached
2603 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2604 TracePhase tp(_t_incrInline);
2605
2606 set_inlining_incrementally(true);
2607 uint low_live_nodes = 0;
2608
2609 while (_late_inlines.length() > 0) {
2610 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2611 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2612 TracePhase tp(_t_incrInline_ideal);
2613 // PhaseIdealLoop is expensive so we only try it once we are
2614 // out of live nodes and we only try it again if the previous
2615 // helped got the number of nodes down significantly
2616 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2617 if (failing()) return;
2618 low_live_nodes = live_nodes();
2619 _major_progress = true;
2620 }
2621
2622 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2623 bool do_print_inlining = print_inlining() || print_intrinsics();
2624 if (do_print_inlining || log() != nullptr) {
2625 // Print inlining message for candidates that we couldn't inline for lack of space.
2626 for (int i = 0; i < _late_inlines.length(); i++) {
2627 CallGenerator* cg = _late_inlines.at(i);
2628 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2629 if (do_print_inlining) {
2630 inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE, msg);
2631 }
2632 log_late_inline_failure(cg, msg);
2633 }
2634 }
2635 break; // finish
2636 }
2637 }
2638
2639 igvn_worklist()->ensure_empty(); // should be done with igvn
2640
2641 while (inline_incrementally_one()) {
2642 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2643 }
2644 if (failing()) return;
2645
2646 inline_incrementally_cleanup(igvn);
2647
2648 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2649
2650 if (failing()) return;
2651
2652 if (_late_inlines.length() == 0) {
2653 break; // no more progress
2654 }
2655 }
2656
2657 igvn_worklist()->ensure_empty(); // should be done with igvn
2658
2659 if (_string_late_inlines.length() > 0) {
2660 assert(has_stringbuilder(), "inconsistent");
2661
2662 inline_string_calls(false);
2663
2664 if (failing()) return;
2665
2666 inline_incrementally_cleanup(igvn);
2667 }
2668
2669 set_inlining_incrementally(false);
2670 }
2671
2672 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2673 // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2674 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2675 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr"
2676 // as if "inlining_incrementally() == true" were set.
2677 assert(inlining_incrementally() == false, "not allowed");
2678 #ifdef ASSERT
2679 Unique_Node_List* modified_nodes = _modified_nodes;
2680 _modified_nodes = nullptr;
2681 #endif
2682 assert(_late_inlines.length() > 0, "sanity");
2683
2684 while (_late_inlines.length() > 0) {
2685 igvn_worklist()->ensure_empty(); // should be done with igvn
2686
2687 while (inline_incrementally_one()) {
2688 assert(!failing_internal() || failure_is_artificial(), "inconsistent");
2689 }
2690 if (failing()) return;
2691
2692 inline_incrementally_cleanup(igvn);
2693 }
2694 DEBUG_ONLY( _modified_nodes = modified_nodes; )
2695 }
2696
2697 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2698 if (_loop_opts_cnt > 0) {
2699 while (major_progress() && (_loop_opts_cnt > 0)) {
2700 TracePhase tp(_t_idealLoop);
2701 PhaseIdealLoop::optimize(igvn, mode);
2702 _loop_opts_cnt--;
2703 if (failing()) return false;
2704 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2705 }
2706 }
2707 return true;
2708 }
2709
2710 // Remove edges from "root" to each SafePoint at a backward branch.
2711 // They were inserted during parsing (see add_safepoint()) to make
2712 // infinite loops without calls or exceptions visible to root, i.e.,
2713 // useful.
2714 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2715 Node *r = root();
2716 if (r != nullptr) {
2717 for (uint i = r->req(); i < r->len(); ++i) {
2718 Node *n = r->in(i);
2719 if (n != nullptr && n->is_SafePoint()) {
2720 r->rm_prec(i);
2721 if (n->outcnt() == 0) {
2722 igvn.remove_dead_node(n);
2723 }
2724 --i;
2725 }
2726 }
2727 // Parsing may have added top inputs to the root node (Path
2728 // leading to the Halt node proven dead). Make sure we get a
2729 // chance to clean them up.
2730 igvn._worklist.push(r);
2731 igvn.optimize();
2732 }
2733 }
2734
2735 //------------------------------Optimize---------------------------------------
2736 // Given a graph, optimize it.
2737 void Compile::Optimize() {
2738 TracePhase tp(_t_optimizer);
2739
2740 #ifndef PRODUCT
2741 if (env()->break_at_compile()) {
2742 BREAKPOINT;
2743 }
2744
2745 #endif
2746
2747 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2748 #ifdef ASSERT
2749 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2750 #endif
2751
2752 ResourceMark rm;
2753
2754 NOT_PRODUCT( verify_graph_edges(); )
2755
2756 print_method(PHASE_AFTER_PARSING, 1);
2757
2758 {
2759 // Iterative Global Value Numbering, including ideal transforms
2760 // Initialize IterGVN with types and values from parse-time GVN
2761 PhaseIterGVN igvn(initial_gvn());
2762 #ifdef ASSERT
2763 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2764 #endif
2765 {
2766 TracePhase tp(_t_iterGVN);
2767 igvn.optimize();
2768 }
2769
2770 if (failing()) return;
2771
2772 print_method(PHASE_ITER_GVN1, 2);
2773
2774 process_for_unstable_if_traps(igvn);
2775
2776 if (failing()) return;
2777
2778 inline_incrementally(igvn);
2779
2780 print_method(PHASE_INCREMENTAL_INLINE, 2);
2781
2782 if (failing()) return;
2783
2784 if (eliminate_boxing()) {
2785 // Inline valueOf() methods now.
2786 inline_boxing_calls(igvn);
2787
2788 if (failing()) return;
2789
2790 if (AlwaysIncrementalInline || StressIncrementalInlining) {
2791 inline_incrementally(igvn);
2792 }
2793
2794 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2795
2796 if (failing()) return;
2797 }
2798
2799 // Remove the speculative part of types and clean up the graph from
2800 // the extra CastPP nodes whose only purpose is to carry them. Do
2801 // that early so that optimizations are not disrupted by the extra
2802 // CastPP nodes.
2803 remove_speculative_types(igvn);
2804
2805 if (failing()) return;
2806
2807 // No more new expensive nodes will be added to the list from here
2808 // so keep only the actual candidates for optimizations.
2809 cleanup_expensive_nodes(igvn);
2810
2811 if (failing()) return;
2812
2813 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2814 if (EnableVectorSupport && has_vbox_nodes()) {
2815 TracePhase tp(_t_vector);
2816 PhaseVector pv(igvn);
2817 pv.optimize_vector_boxes();
2818 if (failing()) return;
2819 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2820 }
2821 assert(!has_vbox_nodes(), "sanity");
2822
2823 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2824 Compile::TracePhase tp(_t_renumberLive);
2825 igvn_worklist()->ensure_empty(); // should be done with igvn
2826 {
2827 ResourceMark rm;
2828 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist());
2829 }
2830 igvn.reset_from_gvn(initial_gvn());
2831 igvn.optimize();
2832 if (failing()) return;
2833 }
2834
2835 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2836 // safepoints
2837 remove_root_to_sfpts_edges(igvn);
2838
2839 // Process inline type nodes now that all inlining is over
2840 process_inline_types(igvn);
2841
2842 adjust_flat_array_access_aliases(igvn);
2843
2844 if (failing()) return;
2845
2846 // Perform escape analysis
2847 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2848 if (has_loops()) {
2849 // Cleanup graph (remove dead nodes).
2850 TracePhase tp(_t_idealLoop);
2851 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2852 if (failing()) return;
2853 }
2854 bool progress;
2855 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2856 do {
2857 ConnectionGraph::do_analysis(this, &igvn);
2858
2859 if (failing()) return;
2860
2861 int mcount = macro_count(); // Record number of allocations and locks before IGVN
2862
2863 // Optimize out fields loads from scalar replaceable allocations.
2864 igvn.optimize();
2865 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2866
2867 if (failing()) return;
2868
2869 if (congraph() != nullptr && macro_count() > 0) {
2870 TracePhase tp(_t_macroEliminate);
2871 PhaseMacroExpand mexp(igvn);
2872 mexp.eliminate_macro_nodes();
2873 if (failing()) return;
2874
2875 igvn.set_delay_transform(false);
2876 igvn.optimize();
2877 if (failing()) return;
2878
2879 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2880 }
2881
2882 ConnectionGraph::verify_ram_nodes(this, root());
2883 if (failing()) return;
2884
2885 progress = do_iterative_escape_analysis() &&
2886 (macro_count() < mcount) &&
2887 ConnectionGraph::has_candidates(this);
2888 // Try again if candidates exist and made progress
2889 // by removing some allocations and/or locks.
2890 } while (progress);
2891 }
2892
2893 // Loop transforms on the ideal graph. Range Check Elimination,
2894 // peeling, unrolling, etc.
2895
2896 // Set loop opts counter
2897 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2898 {
2899 TracePhase tp(_t_idealLoop);
2900 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2901 _loop_opts_cnt--;
2902 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2903 if (failing()) return;
2904 }
2905 // Loop opts pass if partial peeling occurred in previous pass
2906 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2907 TracePhase tp(_t_idealLoop);
2908 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2909 _loop_opts_cnt--;
2910 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2911 if (failing()) return;
2912 }
2913 // Loop opts pass for loop-unrolling before CCP
2914 if(major_progress() && (_loop_opts_cnt > 0)) {
2915 TracePhase tp(_t_idealLoop);
2916 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2917 _loop_opts_cnt--;
2918 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2919 }
2920 if (!failing()) {
2921 // Verify that last round of loop opts produced a valid graph
2922 PhaseIdealLoop::verify(igvn);
2923 }
2924 }
2925 if (failing()) return;
2926
2927 // Conditional Constant Propagation;
2928 print_method(PHASE_BEFORE_CCP1, 2);
2929 PhaseCCP ccp( &igvn );
2930 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2931 {
2932 TracePhase tp(_t_ccp);
2933 ccp.do_transform();
2934 }
2935 print_method(PHASE_CCP1, 2);
2936
2937 assert( true, "Break here to ccp.dump_old2new_map()");
2938
2939 // Iterative Global Value Numbering, including ideal transforms
2940 {
2941 TracePhase tp(_t_iterGVN2);
2942 igvn.reset_from_igvn(&ccp);
2943 igvn.optimize();
2944 }
2945 print_method(PHASE_ITER_GVN2, 2);
2946
2947 if (failing()) return;
2948
2949 // Loop transforms on the ideal graph. Range Check Elimination,
2950 // peeling, unrolling, etc.
2951 if (!optimize_loops(igvn, LoopOptsDefault)) {
2952 return;
2953 }
2954
2955 if (failing()) return;
2956
2957 C->clear_major_progress(); // ensure that major progress is now clear
2958
2959 process_for_post_loop_opts_igvn(igvn);
2960
2961 process_for_merge_stores_igvn(igvn);
2962
2963 if (failing()) return;
2964
2965 #ifdef ASSERT
2966 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2967 #endif
2968
2969 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2970
2971 if (_late_inlines.length() > 0) {
2972 // More opportunities to optimize virtual and MH calls.
2973 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2974 process_late_inline_calls_no_inline(igvn);
2975 }
2976
2977 {
2978 TracePhase tp(_t_macroExpand);
2979 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3);
2980 PhaseMacroExpand mex(igvn);
2981 if (mex.expand_macro_nodes()) {
2982 assert(failing(), "must bail out w/ explicit message");
2983 return;
2984 }
2985 print_method(PHASE_AFTER_MACRO_EXPANSION, 2);
2986 }
2987
2988 // Process inline type nodes again and remove them. From here
2989 // on we don't need to keep track of field values anymore.
2990 process_inline_types(igvn, /* remove= */ true);
2991
2992 {
2993 TracePhase tp(_t_barrierExpand);
2994 if (bs->expand_barriers(this, igvn)) {
2995 assert(failing(), "must bail out w/ explicit message");
2996 return;
2997 }
2998 print_method(PHASE_BARRIER_EXPANSION, 2);
2999 }
3000
3001 if (C->max_vector_size() > 0) {
3002 C->optimize_logic_cones(igvn);
3003 igvn.optimize();
3004 if (failing()) return;
3005 }
3006
3007 DEBUG_ONLY( _modified_nodes = nullptr; )
3008 DEBUG_ONLY( _late_inlines.clear(); )
3009
3010 assert(igvn._worklist.size() == 0, "not empty");
3011 } // (End scope of igvn; run destructor if necessary for asserts.)
3012
3013 check_no_dead_use();
3014
3015 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have
3016 // to remove hashes to unlock nodes for modifications.
3017 C->node_hash()->clear();
3018
3019 // A method with only infinite loops has no edges entering loops from root
3020 {
3021 TracePhase tp(_t_graphReshaping);
3022 if (final_graph_reshaping()) {
3023 assert(failing(), "must bail out w/ explicit message");
3024 return;
3025 }
3026 }
3027
3028 print_method(PHASE_OPTIMIZE_FINISHED, 2);
3029 DEBUG_ONLY(set_phase_optimize_finished();)
3030 }
3031
3032 #ifdef ASSERT
3033 void Compile::check_no_dead_use() const {
3034 ResourceMark rm;
3035 Unique_Node_List wq;
3036 wq.push(root());
3037 for (uint i = 0; i < wq.size(); ++i) {
3038 Node* n = wq.at(i);
3039 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3040 Node* u = n->fast_out(j);
3041 if (u->outcnt() == 0 && !u->is_Con()) {
3042 u->dump();
3043 fatal("no reachable node should have no use");
3044 }
3045 wq.push(u);
3046 }
3047 }
3048 }
3049 #endif
3050
3051 void Compile::inline_vector_reboxing_calls() {
3052 if (C->_vector_reboxing_late_inlines.length() > 0) {
3053 _late_inlines_pos = C->_late_inlines.length();
3054 while (_vector_reboxing_late_inlines.length() > 0) {
3055 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
3056 cg->do_late_inline();
3057 if (failing()) return;
3058 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
3059 }
3060 _vector_reboxing_late_inlines.trunc_to(0);
3061 }
3062 }
3063
3064 bool Compile::has_vbox_nodes() {
3065 if (C->_vector_reboxing_late_inlines.length() > 0) {
3066 return true;
3067 }
3068 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
3069 Node * n = C->macro_node(macro_idx);
3070 assert(n->is_macro(), "only macro nodes expected here");
3071 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
3072 return true;
3073 }
3074 }
3075 return false;
3076 }
3077
3078 //---------------------------- Bitwise operation packing optimization ---------------------------
3079
3080 static bool is_vector_unary_bitwise_op(Node* n) {
3081 return n->Opcode() == Op_XorV &&
3082 VectorNode::is_vector_bitwise_not_pattern(n);
3083 }
3084
3085 static bool is_vector_binary_bitwise_op(Node* n) {
3086 switch (n->Opcode()) {
3087 case Op_AndV:
3088 case Op_OrV:
3089 return true;
3090
3091 case Op_XorV:
3092 return !is_vector_unary_bitwise_op(n);
3093
3094 default:
3095 return false;
3096 }
3097 }
3098
3099 static bool is_vector_ternary_bitwise_op(Node* n) {
3100 return n->Opcode() == Op_MacroLogicV;
3101 }
3102
3103 static bool is_vector_bitwise_op(Node* n) {
3104 return is_vector_unary_bitwise_op(n) ||
3105 is_vector_binary_bitwise_op(n) ||
3106 is_vector_ternary_bitwise_op(n);
3107 }
3108
3109 static bool is_vector_bitwise_cone_root(Node* n) {
3110 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
3111 return false;
3112 }
3113 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3114 if (is_vector_bitwise_op(n->fast_out(i))) {
3115 return false;
3116 }
3117 }
3118 return true;
3119 }
3120
3121 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
3122 uint cnt = 0;
3123 if (is_vector_bitwise_op(n)) {
3124 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
3125 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3126 for (uint i = 1; i < inp_cnt; i++) {
3127 Node* in = n->in(i);
3128 bool skip = VectorNode::is_all_ones_vector(in);
3129 if (!skip && !inputs.member(in)) {
3130 inputs.push(in);
3131 cnt++;
3132 }
3133 }
3134 assert(cnt <= 1, "not unary");
3135 } else {
3136 uint last_req = inp_cnt;
3137 if (is_vector_ternary_bitwise_op(n)) {
3138 last_req = inp_cnt - 1; // skip last input
3139 }
3140 for (uint i = 1; i < last_req; i++) {
3141 Node* def = n->in(i);
3142 if (!inputs.member(def)) {
3143 inputs.push(def);
3144 cnt++;
3145 }
3146 }
3147 }
3148 } else { // not a bitwise operations
3149 if (!inputs.member(n)) {
3150 inputs.push(n);
3151 cnt++;
3152 }
3153 }
3154 return cnt;
3155 }
3156
3157 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
3158 Unique_Node_List useful_nodes;
3159 C->identify_useful_nodes(useful_nodes);
3160
3161 for (uint i = 0; i < useful_nodes.size(); i++) {
3162 Node* n = useful_nodes.at(i);
3163 if (is_vector_bitwise_cone_root(n)) {
3164 list.push(n);
3165 }
3166 }
3167 }
3168
3169 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
3170 const TypeVect* vt,
3171 Unique_Node_List& partition,
3172 Unique_Node_List& inputs) {
3173 assert(partition.size() == 2 || partition.size() == 3, "not supported");
3174 assert(inputs.size() == 2 || inputs.size() == 3, "not supported");
3175 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
3176
3177 Node* in1 = inputs.at(0);
3178 Node* in2 = inputs.at(1);
3179 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
3180
3181 uint func = compute_truth_table(partition, inputs);
3182
3183 Node* pn = partition.at(partition.size() - 1);
3184 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3185 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
3186 }
3187
3188 static uint extract_bit(uint func, uint pos) {
3189 return (func & (1 << pos)) >> pos;
3190 }
3191
3192 //
3193 // A macro logic node represents a truth table. It has 4 inputs,
3194 // First three inputs corresponds to 3 columns of a truth table
3195 // and fourth input captures the logic function.
3196 //
3197 // eg. fn = (in1 AND in2) OR in3;
3198 //
3199 // MacroNode(in1,in2,in3,fn)
3200 //
3201 // -----------------
3202 // in1 in2 in3 fn
3203 // -----------------
3204 // 0 0 0 0
3205 // 0 0 1 1
3206 // 0 1 0 0
3207 // 0 1 1 1
3208 // 1 0 0 0
3209 // 1 0 1 1
3210 // 1 1 0 1
3211 // 1 1 1 1
3212 //
3213
3214 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
3215 int res = 0;
3216 for (int i = 0; i < 8; i++) {
3217 int bit1 = extract_bit(in1, i);
3218 int bit2 = extract_bit(in2, i);
3219 int bit3 = extract_bit(in3, i);
3220
3221 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
3222 int func_bit = extract_bit(func, func_bit_pos);
3223
3224 res |= func_bit << i;
3225 }
3226 return res;
3227 }
3228
3229 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
3230 assert(n != nullptr, "");
3231 assert(eval_map.contains(n), "absent");
3232 return *(eval_map.get(n));
3233 }
3234
3235 static void eval_operands(Node* n,
3236 uint& func1, uint& func2, uint& func3,
3237 ResourceHashtable<Node*,uint>& eval_map) {
3238 assert(is_vector_bitwise_op(n), "");
3239
3240 if (is_vector_unary_bitwise_op(n)) {
3241 Node* opnd = n->in(1);
3242 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
3243 opnd = n->in(2);
3244 }
3245 func1 = eval_operand(opnd, eval_map);
3246 } else if (is_vector_binary_bitwise_op(n)) {
3247 func1 = eval_operand(n->in(1), eval_map);
3248 func2 = eval_operand(n->in(2), eval_map);
3249 } else {
3250 assert(is_vector_ternary_bitwise_op(n), "unknown operation");
3251 func1 = eval_operand(n->in(1), eval_map);
3252 func2 = eval_operand(n->in(2), eval_map);
3253 func3 = eval_operand(n->in(3), eval_map);
3254 }
3255 }
3256
3257 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
3258 assert(inputs.size() <= 3, "sanity");
3259 ResourceMark rm;
3260 uint res = 0;
3261 ResourceHashtable<Node*,uint> eval_map;
3262
3263 // Populate precomputed functions for inputs.
3264 // Each input corresponds to one column of 3 input truth-table.
3265 uint input_funcs[] = { 0xAA, // (_, _, c) -> c
3266 0xCC, // (_, b, _) -> b
3267 0xF0 }; // (a, _, _) -> a
3268 for (uint i = 0; i < inputs.size(); i++) {
3269 eval_map.put(inputs.at(i), input_funcs[2-i]);
3270 }
3271
3272 for (uint i = 0; i < partition.size(); i++) {
3273 Node* n = partition.at(i);
3274
3275 uint func1 = 0, func2 = 0, func3 = 0;
3276 eval_operands(n, func1, func2, func3, eval_map);
3277
3278 switch (n->Opcode()) {
3279 case Op_OrV:
3280 assert(func3 == 0, "not binary");
3281 res = func1 | func2;
3282 break;
3283 case Op_AndV:
3284 assert(func3 == 0, "not binary");
3285 res = func1 & func2;
3286 break;
3287 case Op_XorV:
3288 if (VectorNode::is_vector_bitwise_not_pattern(n)) {
3289 assert(func2 == 0 && func3 == 0, "not unary");
3290 res = (~func1) & 0xFF;
3291 } else {
3292 assert(func3 == 0, "not binary");
3293 res = func1 ^ func2;
3294 }
3295 break;
3296 case Op_MacroLogicV:
3297 // Ordering of inputs may change during evaluation of sub-tree
3298 // containing MacroLogic node as a child node, thus a re-evaluation
3299 // makes sure that function is evaluated in context of current
3300 // inputs.
3301 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
3302 break;
3303
3304 default: assert(false, "not supported: %s", n->Name());
3305 }
3306 assert(res <= 0xFF, "invalid");
3307 eval_map.put(n, res);
3308 }
3309 return res;
3310 }
3311
3312 // Criteria under which nodes gets packed into a macro logic node:-
3313 // 1) Parent and both child nodes are all unmasked or masked with
3314 // same predicates.
3315 // 2) Masked parent can be packed with left child if it is predicated
3316 // and both have same predicates.
3317 // 3) Masked parent can be packed with right child if its un-predicated
3318 // or has matching predication condition.
3319 // 4) An unmasked parent can be packed with an unmasked child.
3320 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
3321 assert(partition.size() == 0, "not empty");
3322 assert(inputs.size() == 0, "not empty");
3323 if (is_vector_ternary_bitwise_op(n)) {
3324 return false;
3325 }
3326
3327 bool is_unary_op = is_vector_unary_bitwise_op(n);
3328 if (is_unary_op) {
3329 assert(collect_unique_inputs(n, inputs) == 1, "not unary");
3330 return false; // too few inputs
3331 }
3332
3333 bool pack_left_child = true;
3334 bool pack_right_child = true;
3335
3336 bool left_child_LOP = is_vector_bitwise_op(n->in(1));
3337 bool right_child_LOP = is_vector_bitwise_op(n->in(2));
3338
3339 int left_child_input_cnt = 0;
3340 int right_child_input_cnt = 0;
3341
3342 bool parent_is_predicated = n->is_predicated_vector();
3343 bool left_child_predicated = n->in(1)->is_predicated_vector();
3344 bool right_child_predicated = n->in(2)->is_predicated_vector();
3345
3346 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr;
3347 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3348 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr;
3349
3350 do {
3351 if (pack_left_child && left_child_LOP &&
3352 ((!parent_is_predicated && !left_child_predicated) ||
3353 ((parent_is_predicated && left_child_predicated &&
3354 parent_pred == left_child_pred)))) {
3355 partition.push(n->in(1));
3356 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
3357 } else {
3358 inputs.push(n->in(1));
3359 left_child_input_cnt = 1;
3360 }
3361
3362 if (pack_right_child && right_child_LOP &&
3363 (!right_child_predicated ||
3364 (right_child_predicated && parent_is_predicated &&
3365 parent_pred == right_child_pred))) {
3366 partition.push(n->in(2));
3367 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
3368 } else {
3369 inputs.push(n->in(2));
3370 right_child_input_cnt = 1;
3371 }
3372
3373 if (inputs.size() > 3) {
3374 assert(partition.size() > 0, "");
3375 inputs.clear();
3376 partition.clear();
3377 if (left_child_input_cnt > right_child_input_cnt) {
3378 pack_left_child = false;
3379 } else {
3380 pack_right_child = false;
3381 }
3382 } else {
3383 break;
3384 }
3385 } while(true);
3386
3387 if(partition.size()) {
3388 partition.push(n);
3389 }
3390
3391 return (partition.size() == 2 || partition.size() == 3) &&
3392 (inputs.size() == 2 || inputs.size() == 3);
3393 }
3394
3395 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
3396 assert(is_vector_bitwise_op(n), "not a root");
3397
3398 visited.set(n->_idx);
3399
3400 // 1) Do a DFS walk over the logic cone.
3401 for (uint i = 1; i < n->req(); i++) {
3402 Node* in = n->in(i);
3403 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
3404 process_logic_cone_root(igvn, in, visited);
3405 }
3406 }
3407
3408 // 2) Bottom up traversal: Merge node[s] with
3409 // the parent to form macro logic node.
3410 Unique_Node_List partition;
3411 Unique_Node_List inputs;
3412 if (compute_logic_cone(n, partition, inputs)) {
3413 const TypeVect* vt = n->bottom_type()->is_vect();
3414 Node* pn = partition.at(partition.size() - 1);
3415 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr;
3416 if (mask == nullptr ||
3417 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
3418 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
3419 VectorNode::trace_new_vector(macro_logic, "MacroLogic");
3420 igvn.replace_node(n, macro_logic);
3421 }
3422 }
3423 }
3424
3425 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
3426 ResourceMark rm;
3427 if (Matcher::match_rule_supported(Op_MacroLogicV)) {
3428 Unique_Node_List list;
3429 collect_logic_cone_roots(list);
3430
3431 while (list.size() > 0) {
3432 Node* n = list.pop();
3433 const TypeVect* vt = n->bottom_type()->is_vect();
3434 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
3435 if (supported) {
3436 VectorSet visited(comp_arena());
3437 process_logic_cone_root(igvn, n, visited);
3438 }
3439 }
3440 }
3441 }
3442
3443 //------------------------------Code_Gen---------------------------------------
3444 // Given a graph, generate code for it
3445 void Compile::Code_Gen() {
3446 if (failing()) {
3447 return;
3448 }
3449
3450 // Perform instruction selection. You might think we could reclaim Matcher
3451 // memory PDQ, but actually the Matcher is used in generating spill code.
3452 // Internals of the Matcher (including some VectorSets) must remain live
3453 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
3454 // set a bit in reclaimed memory.
3455
3456 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3457 // nodes. Mapping is only valid at the root of each matched subtree.
3458 NOT_PRODUCT( verify_graph_edges(); )
3459
3460 Matcher matcher;
3461 _matcher = &matcher;
3462 {
3463 TracePhase tp(_t_matcher);
3464 matcher.match();
3465 if (failing()) {
3466 return;
3467 }
3468 }
3469 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
3470 // nodes. Mapping is only valid at the root of each matched subtree.
3471 NOT_PRODUCT( verify_graph_edges(); )
3472
3473 // If you have too many nodes, or if matching has failed, bail out
3474 check_node_count(0, "out of nodes matching instructions");
3475 if (failing()) {
3476 return;
3477 }
3478
3479 print_method(PHASE_MATCHING, 2);
3480
3481 // Build a proper-looking CFG
3482 PhaseCFG cfg(node_arena(), root(), matcher);
3483 if (failing()) {
3484 return;
3485 }
3486 _cfg = &cfg;
3487 {
3488 TracePhase tp(_t_scheduler);
3489 bool success = cfg.do_global_code_motion();
3490 if (!success) {
3491 return;
3492 }
3493
3494 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
3495 NOT_PRODUCT( verify_graph_edges(); )
3496 cfg.verify();
3497 if (failing()) {
3498 return;
3499 }
3500 }
3501
3502 PhaseChaitin regalloc(unique(), cfg, matcher, false);
3503 _regalloc = ®alloc;
3504 {
3505 TracePhase tp(_t_registerAllocation);
3506 // Perform register allocation. After Chaitin, use-def chains are
3507 // no longer accurate (at spill code) and so must be ignored.
3508 // Node->LRG->reg mappings are still accurate.
3509 _regalloc->Register_Allocate();
3510
3511 // Bail out if the allocator builds too many nodes
3512 if (failing()) {
3513 return;
3514 }
3515
3516 print_method(PHASE_REGISTER_ALLOCATION, 2);
3517 }
3518
3519 // Prior to register allocation we kept empty basic blocks in case the
3520 // the allocator needed a place to spill. After register allocation we
3521 // are not adding any new instructions. If any basic block is empty, we
3522 // can now safely remove it.
3523 {
3524 TracePhase tp(_t_blockOrdering);
3525 cfg.remove_empty_blocks();
3526 if (do_freq_based_layout()) {
3527 PhaseBlockLayout layout(cfg);
3528 } else {
3529 cfg.set_loop_alignment();
3530 }
3531 cfg.fixup_flow();
3532 cfg.remove_unreachable_blocks();
3533 cfg.verify_dominator_tree();
3534 print_method(PHASE_BLOCK_ORDERING, 3);
3535 }
3536
3537 // Apply peephole optimizations
3538 if( OptoPeephole ) {
3539 TracePhase tp(_t_peephole);
3540 PhasePeephole peep( _regalloc, cfg);
3541 peep.do_transform();
3542 print_method(PHASE_PEEPHOLE, 3);
3543 }
3544
3545 // Do late expand if CPU requires this.
3546 if (Matcher::require_postalloc_expand) {
3547 TracePhase tp(_t_postalloc_expand);
3548 cfg.postalloc_expand(_regalloc);
3549 print_method(PHASE_POSTALLOC_EXPAND, 3);
3550 }
3551
3552 #ifdef ASSERT
3553 {
3554 CompilationMemoryStatistic::do_test_allocations();
3555 if (failing()) return;
3556 }
3557 #endif
3558
3559 // Convert Nodes to instruction bits in a buffer
3560 {
3561 TracePhase tp(_t_output);
3562 PhaseOutput output;
3563 output.Output();
3564 if (failing()) return;
3565 output.install();
3566 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here
3567 }
3568
3569 // He's dead, Jim.
3570 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3571 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3572 }
3573
3574 //------------------------------Final_Reshape_Counts---------------------------
3575 // This class defines counters to help identify when a method
3576 // may/must be executed using hardware with only 24-bit precision.
3577 struct Final_Reshape_Counts : public StackObj {
3578 int _call_count; // count non-inlined 'common' calls
3579 int _float_count; // count float ops requiring 24-bit precision
3580 int _double_count; // count double ops requiring more precision
3581 int _java_call_count; // count non-inlined 'java' calls
3582 int _inner_loop_count; // count loops which need alignment
3583 VectorSet _visited; // Visitation flags
3584 Node_List _tests; // Set of IfNodes & PCTableNodes
3585
3586 Final_Reshape_Counts() :
3587 _call_count(0), _float_count(0), _double_count(0),
3588 _java_call_count(0), _inner_loop_count(0) { }
3589
3590 void inc_call_count () { _call_count ++; }
3591 void inc_float_count () { _float_count ++; }
3592 void inc_double_count() { _double_count++; }
3593 void inc_java_call_count() { _java_call_count++; }
3594 void inc_inner_loop_count() { _inner_loop_count++; }
3595
3596 int get_call_count () const { return _call_count ; }
3597 int get_float_count () const { return _float_count ; }
3598 int get_double_count() const { return _double_count; }
3599 int get_java_call_count() const { return _java_call_count; }
3600 int get_inner_loop_count() const { return _inner_loop_count; }
3601 };
3602
3603 //------------------------------final_graph_reshaping_impl----------------------
3604 // Implement items 1-5 from final_graph_reshaping below.
3605 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3606
3607 if ( n->outcnt() == 0 ) return; // dead node
3608 uint nop = n->Opcode();
3609
3610 // Check for 2-input instruction with "last use" on right input.
3611 // Swap to left input. Implements item (2).
3612 if( n->req() == 3 && // two-input instruction
3613 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3614 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3615 n->in(2)->outcnt() == 1 &&// right use IS a last use
3616 !n->in(2)->is_Con() ) { // right use is not a constant
3617 // Check for commutative opcode
3618 switch( nop ) {
3619 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3620 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3621 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3622 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3623 case Op_AndL: case Op_XorL: case Op_OrL:
3624 case Op_AndI: case Op_XorI: case Op_OrI: {
3625 // Move "last use" input to left by swapping inputs
3626 n->swap_edges(1, 2);
3627 break;
3628 }
3629 default:
3630 break;
3631 }
3632 }
3633
3634 #ifdef ASSERT
3635 if( n->is_Mem() ) {
3636 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3637 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw ||
3638 // oop will be recorded in oop map if load crosses safepoint
3639 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3640 LoadNode::is_immutable_value(n->in(MemNode::Address)))),
3641 "raw memory operations should have control edge");
3642 }
3643 if (n->is_MemBar()) {
3644 MemBarNode* mb = n->as_MemBar();
3645 if (mb->trailing_store() || mb->trailing_load_store()) {
3646 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3647 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3648 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3649 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3650 } else if (mb->leading()) {
3651 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3652 }
3653 }
3654 #endif
3655 // Count FPU ops and common calls, implements item (3)
3656 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3657 if (!gc_handled) {
3658 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3659 }
3660
3661 // Collect CFG split points
3662 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3663 frc._tests.push(n);
3664 }
3665 }
3666
3667 void Compile::handle_div_mod_op(Node* n, BasicType bt, bool is_unsigned) {
3668 if (!UseDivMod) {
3669 return;
3670 }
3671
3672 // Check if "a % b" and "a / b" both exist
3673 Node* d = n->find_similar(Op_DivIL(bt, is_unsigned));
3674 if (d == nullptr) {
3675 return;
3676 }
3677
3678 // Replace them with a fused divmod if supported
3679 if (Matcher::has_match_rule(Op_DivModIL(bt, is_unsigned))) {
3680 DivModNode* divmod = DivModNode::make(n, bt, is_unsigned);
3681 // If the divisor input for a Div (or Mod etc.) is not zero, then the control input of the Div is set to zero.
3682 // It could be that the divisor input is found not zero because its type is narrowed down by a CastII in the
3683 // subgraph for that input. Range check CastIIs are removed during final graph reshape. To preserve the dependency
3684 // carried by a CastII, precedence edges are added to the Div node. We need to transfer the precedence edges to the
3685 // DivMod node so the dependency is not lost.
3686 divmod->add_prec_from(n);
3687 divmod->add_prec_from(d);
3688 d->subsume_by(divmod->div_proj(), this);
3689 n->subsume_by(divmod->mod_proj(), this);
3690 } else {
3691 // Replace "a % b" with "a - ((a / b) * b)"
3692 Node* mult = MulNode::make(d, d->in(2), bt);
3693 Node* sub = SubNode::make(d->in(1), mult, bt);
3694 n->subsume_by(sub, this);
3695 }
3696 }
3697
3698 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3699 switch( nop ) {
3700 // Count all float operations that may use FPU
3701 case Op_AddF:
3702 case Op_SubF:
3703 case Op_MulF:
3704 case Op_DivF:
3705 case Op_NegF:
3706 case Op_ModF:
3707 case Op_ConvI2F:
3708 case Op_ConF:
3709 case Op_CmpF:
3710 case Op_CmpF3:
3711 case Op_StoreF:
3712 case Op_LoadF:
3713 // case Op_ConvL2F: // longs are split into 32-bit halves
3714 frc.inc_float_count();
3715 break;
3716
3717 case Op_ConvF2D:
3718 case Op_ConvD2F:
3719 frc.inc_float_count();
3720 frc.inc_double_count();
3721 break;
3722
3723 // Count all double operations that may use FPU
3724 case Op_AddD:
3725 case Op_SubD:
3726 case Op_MulD:
3727 case Op_DivD:
3728 case Op_NegD:
3729 case Op_ModD:
3730 case Op_ConvI2D:
3731 case Op_ConvD2I:
3732 // case Op_ConvL2D: // handled by leaf call
3733 // case Op_ConvD2L: // handled by leaf call
3734 case Op_ConD:
3735 case Op_CmpD:
3736 case Op_CmpD3:
3737 case Op_StoreD:
3738 case Op_LoadD:
3739 case Op_LoadD_unaligned:
3740 frc.inc_double_count();
3741 break;
3742 case Op_Opaque1: // Remove Opaque Nodes before matching
3743 n->subsume_by(n->in(1), this);
3744 break;
3745 case Op_CallStaticJava:
3746 case Op_CallJava:
3747 case Op_CallDynamicJava:
3748 frc.inc_java_call_count(); // Count java call site;
3749 case Op_CallRuntime:
3750 case Op_CallLeaf:
3751 case Op_CallLeafVector:
3752 case Op_CallLeafNoFP: {
3753 assert (n->is_Call(), "");
3754 CallNode *call = n->as_Call();
3755 // Count call sites where the FP mode bit would have to be flipped.
3756 // Do not count uncommon runtime calls:
3757 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3758 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3759 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3760 frc.inc_call_count(); // Count the call site
3761 } else { // See if uncommon argument is shared
3762 Node *n = call->in(TypeFunc::Parms);
3763 int nop = n->Opcode();
3764 // Clone shared simple arguments to uncommon calls, item (1).
3765 if (n->outcnt() > 1 &&
3766 !n->is_Proj() &&
3767 nop != Op_CreateEx &&
3768 nop != Op_CheckCastPP &&
3769 nop != Op_DecodeN &&
3770 nop != Op_DecodeNKlass &&
3771 !n->is_Mem() &&
3772 !n->is_Phi()) {
3773 Node *x = n->clone();
3774 call->set_req(TypeFunc::Parms, x);
3775 }
3776 }
3777 break;
3778 }
3779 case Op_StoreB:
3780 case Op_StoreC:
3781 case Op_StoreI:
3782 case Op_StoreL:
3783 case Op_StoreLSpecial:
3784 case Op_CompareAndSwapB:
3785 case Op_CompareAndSwapS:
3786 case Op_CompareAndSwapI:
3787 case Op_CompareAndSwapL:
3788 case Op_CompareAndSwapP:
3789 case Op_CompareAndSwapN:
3790 case Op_WeakCompareAndSwapB:
3791 case Op_WeakCompareAndSwapS:
3792 case Op_WeakCompareAndSwapI:
3793 case Op_WeakCompareAndSwapL:
3794 case Op_WeakCompareAndSwapP:
3795 case Op_WeakCompareAndSwapN:
3796 case Op_CompareAndExchangeB:
3797 case Op_CompareAndExchangeS:
3798 case Op_CompareAndExchangeI:
3799 case Op_CompareAndExchangeL:
3800 case Op_CompareAndExchangeP:
3801 case Op_CompareAndExchangeN:
3802 case Op_GetAndAddS:
3803 case Op_GetAndAddB:
3804 case Op_GetAndAddI:
3805 case Op_GetAndAddL:
3806 case Op_GetAndSetS:
3807 case Op_GetAndSetB:
3808 case Op_GetAndSetI:
3809 case Op_GetAndSetL:
3810 case Op_GetAndSetP:
3811 case Op_GetAndSetN:
3812 case Op_StoreP:
3813 case Op_StoreN:
3814 case Op_StoreNKlass:
3815 case Op_LoadB:
3816 case Op_LoadUB:
3817 case Op_LoadUS:
3818 case Op_LoadI:
3819 case Op_LoadKlass:
3820 case Op_LoadNKlass:
3821 case Op_LoadL:
3822 case Op_LoadL_unaligned:
3823 case Op_LoadP:
3824 case Op_LoadN:
3825 case Op_LoadRange:
3826 case Op_LoadS:
3827 break;
3828
3829 case Op_AddP: { // Assert sane base pointers
3830 Node *addp = n->in(AddPNode::Address);
3831 assert( !addp->is_AddP() ||
3832 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3833 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3834 "Base pointers must match (addp %u)", addp->_idx );
3835 #ifdef _LP64
3836 if ((UseCompressedOops || UseCompressedClassPointers) &&
3837 addp->Opcode() == Op_ConP &&
3838 addp == n->in(AddPNode::Base) &&
3839 n->in(AddPNode::Offset)->is_Con()) {
3840 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3841 // on the platform and on the compressed oops mode.
3842 // Use addressing with narrow klass to load with offset on x86.
3843 // Some platforms can use the constant pool to load ConP.
3844 // Do this transformation here since IGVN will convert ConN back to ConP.
3845 const Type* t = addp->bottom_type();
3846 bool is_oop = t->isa_oopptr() != nullptr;
3847 bool is_klass = t->isa_klassptr() != nullptr;
3848
3849 if ((is_oop && UseCompressedOops && Matcher::const_oop_prefer_decode() ) ||
3850 (is_klass && UseCompressedClassPointers && Matcher::const_klass_prefer_decode() &&
3851 t->isa_klassptr()->exact_klass()->is_in_encoding_range())) {
3852 Node* nn = nullptr;
3853
3854 int op = is_oop ? Op_ConN : Op_ConNKlass;
3855
3856 // Look for existing ConN node of the same exact type.
3857 Node* r = root();
3858 uint cnt = r->outcnt();
3859 for (uint i = 0; i < cnt; i++) {
3860 Node* m = r->raw_out(i);
3861 if (m!= nullptr && m->Opcode() == op &&
3862 m->bottom_type()->make_ptr() == t) {
3863 nn = m;
3864 break;
3865 }
3866 }
3867 if (nn != nullptr) {
3868 // Decode a narrow oop to match address
3869 // [R12 + narrow_oop_reg<<3 + offset]
3870 if (is_oop) {
3871 nn = new DecodeNNode(nn, t);
3872 } else {
3873 nn = new DecodeNKlassNode(nn, t);
3874 }
3875 // Check for succeeding AddP which uses the same Base.
3876 // Otherwise we will run into the assertion above when visiting that guy.
3877 for (uint i = 0; i < n->outcnt(); ++i) {
3878 Node *out_i = n->raw_out(i);
3879 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3880 out_i->set_req(AddPNode::Base, nn);
3881 #ifdef ASSERT
3882 for (uint j = 0; j < out_i->outcnt(); ++j) {
3883 Node *out_j = out_i->raw_out(j);
3884 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3885 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3886 }
3887 #endif
3888 }
3889 }
3890 n->set_req(AddPNode::Base, nn);
3891 n->set_req(AddPNode::Address, nn);
3892 if (addp->outcnt() == 0) {
3893 addp->disconnect_inputs(this);
3894 }
3895 }
3896 }
3897 }
3898 #endif
3899 break;
3900 }
3901
3902 case Op_CastPP: {
3903 // Remove CastPP nodes to gain more freedom during scheduling but
3904 // keep the dependency they encode as control or precedence edges
3905 // (if control is set already) on memory operations. Some CastPP
3906 // nodes don't have a control (don't carry a dependency): skip
3907 // those.
3908 if (n->in(0) != nullptr) {
3909 ResourceMark rm;
3910 Unique_Node_List wq;
3911 wq.push(n);
3912 for (uint next = 0; next < wq.size(); ++next) {
3913 Node *m = wq.at(next);
3914 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3915 Node* use = m->fast_out(i);
3916 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3917 use->ensure_control_or_add_prec(n->in(0));
3918 } else {
3919 switch(use->Opcode()) {
3920 case Op_AddP:
3921 case Op_DecodeN:
3922 case Op_DecodeNKlass:
3923 case Op_CheckCastPP:
3924 case Op_CastPP:
3925 wq.push(use);
3926 break;
3927 }
3928 }
3929 }
3930 }
3931 }
3932 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3933 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3934 Node* in1 = n->in(1);
3935 const Type* t = n->bottom_type();
3936 Node* new_in1 = in1->clone();
3937 new_in1->as_DecodeN()->set_type(t);
3938
3939 if (!Matcher::narrow_oop_use_complex_address()) {
3940 //
3941 // x86, ARM and friends can handle 2 adds in addressing mode
3942 // and Matcher can fold a DecodeN node into address by using
3943 // a narrow oop directly and do implicit null check in address:
3944 //
3945 // [R12 + narrow_oop_reg<<3 + offset]
3946 // NullCheck narrow_oop_reg
3947 //
3948 // On other platforms (Sparc) we have to keep new DecodeN node and
3949 // use it to do implicit null check in address:
3950 //
3951 // decode_not_null narrow_oop_reg, base_reg
3952 // [base_reg + offset]
3953 // NullCheck base_reg
3954 //
3955 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3956 // to keep the information to which null check the new DecodeN node
3957 // corresponds to use it as value in implicit_null_check().
3958 //
3959 new_in1->set_req(0, n->in(0));
3960 }
3961
3962 n->subsume_by(new_in1, this);
3963 if (in1->outcnt() == 0) {
3964 in1->disconnect_inputs(this);
3965 }
3966 } else {
3967 n->subsume_by(n->in(1), this);
3968 if (n->outcnt() == 0) {
3969 n->disconnect_inputs(this);
3970 }
3971 }
3972 break;
3973 }
3974 case Op_CastII: {
3975 n->as_CastII()->remove_range_check_cast(this);
3976 break;
3977 }
3978 #ifdef _LP64
3979 case Op_CmpP:
3980 // Do this transformation here to preserve CmpPNode::sub() and
3981 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3982 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3983 Node* in1 = n->in(1);
3984 Node* in2 = n->in(2);
3985 if (!in1->is_DecodeNarrowPtr()) {
3986 in2 = in1;
3987 in1 = n->in(2);
3988 }
3989 assert(in1->is_DecodeNarrowPtr(), "sanity");
3990
3991 Node* new_in2 = nullptr;
3992 if (in2->is_DecodeNarrowPtr()) {
3993 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3994 new_in2 = in2->in(1);
3995 } else if (in2->Opcode() == Op_ConP) {
3996 const Type* t = in2->bottom_type();
3997 if (t == TypePtr::NULL_PTR) {
3998 assert(in1->is_DecodeN(), "compare klass to null?");
3999 // Don't convert CmpP null check into CmpN if compressed
4000 // oops implicit null check is not generated.
4001 // This will allow to generate normal oop implicit null check.
4002 if (Matcher::gen_narrow_oop_implicit_null_checks())
4003 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
4004 //
4005 // This transformation together with CastPP transformation above
4006 // will generated code for implicit null checks for compressed oops.
4007 //
4008 // The original code after Optimize()
4009 //
4010 // LoadN memory, narrow_oop_reg
4011 // decode narrow_oop_reg, base_reg
4012 // CmpP base_reg, nullptr
4013 // CastPP base_reg // NotNull
4014 // Load [base_reg + offset], val_reg
4015 //
4016 // after these transformations will be
4017 //
4018 // LoadN memory, narrow_oop_reg
4019 // CmpN narrow_oop_reg, nullptr
4020 // decode_not_null narrow_oop_reg, base_reg
4021 // Load [base_reg + offset], val_reg
4022 //
4023 // and the uncommon path (== nullptr) will use narrow_oop_reg directly
4024 // since narrow oops can be used in debug info now (see the code in
4025 // final_graph_reshaping_walk()).
4026 //
4027 // At the end the code will be matched to
4028 // on x86:
4029 //
4030 // Load_narrow_oop memory, narrow_oop_reg
4031 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
4032 // NullCheck narrow_oop_reg
4033 //
4034 // and on sparc:
4035 //
4036 // Load_narrow_oop memory, narrow_oop_reg
4037 // decode_not_null narrow_oop_reg, base_reg
4038 // Load [base_reg + offset], val_reg
4039 // NullCheck base_reg
4040 //
4041 } else if (t->isa_oopptr()) {
4042 new_in2 = ConNode::make(t->make_narrowoop());
4043 } else if (t->isa_klassptr()) {
4044 new_in2 = ConNode::make(t->make_narrowklass());
4045 }
4046 }
4047 if (new_in2 != nullptr) {
4048 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
4049 n->subsume_by(cmpN, this);
4050 if (in1->outcnt() == 0) {
4051 in1->disconnect_inputs(this);
4052 }
4053 if (in2->outcnt() == 0) {
4054 in2->disconnect_inputs(this);
4055 }
4056 }
4057 }
4058 break;
4059
4060 case Op_DecodeN:
4061 case Op_DecodeNKlass:
4062 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
4063 // DecodeN could be pinned when it can't be fold into
4064 // an address expression, see the code for Op_CastPP above.
4065 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
4066 break;
4067
4068 case Op_EncodeP:
4069 case Op_EncodePKlass: {
4070 Node* in1 = n->in(1);
4071 if (in1->is_DecodeNarrowPtr()) {
4072 n->subsume_by(in1->in(1), this);
4073 } else if (in1->Opcode() == Op_ConP) {
4074 const Type* t = in1->bottom_type();
4075 if (t == TypePtr::NULL_PTR) {
4076 assert(t->isa_oopptr(), "null klass?");
4077 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
4078 } else if (t->isa_oopptr()) {
4079 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
4080 } else if (t->isa_klassptr()) {
4081 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
4082 }
4083 }
4084 if (in1->outcnt() == 0) {
4085 in1->disconnect_inputs(this);
4086 }
4087 break;
4088 }
4089
4090 case Op_Proj: {
4091 if (OptimizeStringConcat || IncrementalInline) {
4092 ProjNode* proj = n->as_Proj();
4093 if (proj->_is_io_use) {
4094 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
4095 // Separate projections were used for the exception path which
4096 // are normally removed by a late inline. If it wasn't inlined
4097 // then they will hang around and should just be replaced with
4098 // the original one. Merge them.
4099 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
4100 if (non_io_proj != nullptr) {
4101 proj->subsume_by(non_io_proj , this);
4102 }
4103 }
4104 }
4105 break;
4106 }
4107
4108 case Op_Phi:
4109 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
4110 // The EncodeP optimization may create Phi with the same edges
4111 // for all paths. It is not handled well by Register Allocator.
4112 Node* unique_in = n->in(1);
4113 assert(unique_in != nullptr, "");
4114 uint cnt = n->req();
4115 for (uint i = 2; i < cnt; i++) {
4116 Node* m = n->in(i);
4117 assert(m != nullptr, "");
4118 if (unique_in != m)
4119 unique_in = nullptr;
4120 }
4121 if (unique_in != nullptr) {
4122 n->subsume_by(unique_in, this);
4123 }
4124 }
4125 break;
4126
4127 #endif
4128
4129 case Op_ModI:
4130 handle_div_mod_op(n, T_INT, false);
4131 break;
4132
4133 case Op_ModL:
4134 handle_div_mod_op(n, T_LONG, false);
4135 break;
4136
4137 case Op_UModI:
4138 handle_div_mod_op(n, T_INT, true);
4139 break;
4140
4141 case Op_UModL:
4142 handle_div_mod_op(n, T_LONG, true);
4143 break;
4144
4145 case Op_LoadVector:
4146 case Op_StoreVector:
4147 #ifdef ASSERT
4148 // Add VerifyVectorAlignment node between adr and load / store.
4149 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) {
4150 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() :
4151 n->as_StoreVector()->must_verify_alignment();
4152 if (must_verify_alignment) {
4153 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() :
4154 n->as_StoreVector()->memory_size();
4155 // The memory access should be aligned to the vector width in bytes.
4156 // However, the underlying array is possibly less well aligned, but at least
4157 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in
4158 // a loop we can expect at least the following alignment:
4159 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes);
4160 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range");
4161 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2");
4162 // Create mask from alignment. e.g. 0b1000 -> 0b0111
4163 jlong mask = guaranteed_alignment - 1;
4164 Node* mask_con = ConLNode::make(mask);
4165 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con);
4166 n->set_req(MemNode::Address, va);
4167 }
4168 }
4169 #endif
4170 break;
4171
4172 case Op_LoadVectorGather:
4173 case Op_StoreVectorScatter:
4174 case Op_LoadVectorGatherMasked:
4175 case Op_StoreVectorScatterMasked:
4176 case Op_VectorCmpMasked:
4177 case Op_VectorMaskGen:
4178 case Op_LoadVectorMasked:
4179 case Op_StoreVectorMasked:
4180 break;
4181
4182 case Op_AddReductionVI:
4183 case Op_AddReductionVL:
4184 case Op_AddReductionVF:
4185 case Op_AddReductionVD:
4186 case Op_MulReductionVI:
4187 case Op_MulReductionVL:
4188 case Op_MulReductionVF:
4189 case Op_MulReductionVD:
4190 case Op_MinReductionV:
4191 case Op_MaxReductionV:
4192 case Op_AndReductionV:
4193 case Op_OrReductionV:
4194 case Op_XorReductionV:
4195 break;
4196
4197 case Op_PackB:
4198 case Op_PackS:
4199 case Op_PackI:
4200 case Op_PackF:
4201 case Op_PackL:
4202 case Op_PackD:
4203 if (n->req()-1 > 2) {
4204 // Replace many operand PackNodes with a binary tree for matching
4205 PackNode* p = (PackNode*) n;
4206 Node* btp = p->binary_tree_pack(1, n->req());
4207 n->subsume_by(btp, this);
4208 }
4209 break;
4210 case Op_Loop:
4211 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
4212 case Op_CountedLoop:
4213 case Op_LongCountedLoop:
4214 case Op_OuterStripMinedLoop:
4215 if (n->as_Loop()->is_inner_loop()) {
4216 frc.inc_inner_loop_count();
4217 }
4218 n->as_Loop()->verify_strip_mined(0);
4219 break;
4220 case Op_LShiftI:
4221 case Op_RShiftI:
4222 case Op_URShiftI:
4223 case Op_LShiftL:
4224 case Op_RShiftL:
4225 case Op_URShiftL:
4226 if (Matcher::need_masked_shift_count) {
4227 // The cpu's shift instructions don't restrict the count to the
4228 // lower 5/6 bits. We need to do the masking ourselves.
4229 Node* in2 = n->in(2);
4230 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
4231 const TypeInt* t = in2->find_int_type();
4232 if (t != nullptr && t->is_con()) {
4233 juint shift = t->get_con();
4234 if (shift > mask) { // Unsigned cmp
4235 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
4236 }
4237 } else {
4238 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) {
4239 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
4240 n->set_req(2, shift);
4241 }
4242 }
4243 if (in2->outcnt() == 0) { // Remove dead node
4244 in2->disconnect_inputs(this);
4245 }
4246 }
4247 break;
4248 case Op_MemBarStoreStore:
4249 case Op_MemBarRelease:
4250 // Break the link with AllocateNode: it is no longer useful and
4251 // confuses register allocation.
4252 if (n->req() > MemBarNode::Precedent) {
4253 n->set_req(MemBarNode::Precedent, top());
4254 }
4255 break;
4256 case Op_MemBarAcquire: {
4257 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
4258 // At parse time, the trailing MemBarAcquire for a volatile load
4259 // is created with an edge to the load. After optimizations,
4260 // that input may be a chain of Phis. If those phis have no
4261 // other use, then the MemBarAcquire keeps them alive and
4262 // register allocation can be confused.
4263 dead_nodes.push(n->in(MemBarNode::Precedent));
4264 n->set_req(MemBarNode::Precedent, top());
4265 }
4266 break;
4267 }
4268 case Op_Blackhole:
4269 break;
4270 case Op_RangeCheck: {
4271 RangeCheckNode* rc = n->as_RangeCheck();
4272 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
4273 n->subsume_by(iff, this);
4274 frc._tests.push(iff);
4275 break;
4276 }
4277 case Op_ConvI2L: {
4278 if (!Matcher::convi2l_type_required) {
4279 // Code generation on some platforms doesn't need accurate
4280 // ConvI2L types. Widening the type can help remove redundant
4281 // address computations.
4282 n->as_Type()->set_type(TypeLong::INT);
4283 ResourceMark rm;
4284 Unique_Node_List wq;
4285 wq.push(n);
4286 for (uint next = 0; next < wq.size(); next++) {
4287 Node *m = wq.at(next);
4288
4289 for(;;) {
4290 // Loop over all nodes with identical inputs edges as m
4291 Node* k = m->find_similar(m->Opcode());
4292 if (k == nullptr) {
4293 break;
4294 }
4295 // Push their uses so we get a chance to remove node made
4296 // redundant
4297 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
4298 Node* u = k->fast_out(i);
4299 if (u->Opcode() == Op_LShiftL ||
4300 u->Opcode() == Op_AddL ||
4301 u->Opcode() == Op_SubL ||
4302 u->Opcode() == Op_AddP) {
4303 wq.push(u);
4304 }
4305 }
4306 // Replace all nodes with identical edges as m with m
4307 k->subsume_by(m, this);
4308 }
4309 }
4310 }
4311 break;
4312 }
4313 case Op_CmpUL: {
4314 if (!Matcher::has_match_rule(Op_CmpUL)) {
4315 // No support for unsigned long comparisons
4316 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
4317 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
4318 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
4319 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
4320 Node* andl = new AndLNode(orl, remove_sign_mask);
4321 Node* cmp = new CmpLNode(andl, n->in(2));
4322 n->subsume_by(cmp, this);
4323 }
4324 break;
4325 }
4326 #ifdef ASSERT
4327 case Op_InlineType: {
4328 n->dump(-1);
4329 assert(false, "inline type node was not removed");
4330 break;
4331 }
4332 case Op_ConNKlass: {
4333 const TypePtr* tp = n->as_Type()->type()->make_ptr();
4334 ciKlass* klass = tp->is_klassptr()->exact_klass();
4335 assert(klass->is_in_encoding_range(), "klass cannot be compressed");
4336 break;
4337 }
4338 #endif
4339 default:
4340 assert(!n->is_Call(), "");
4341 assert(!n->is_Mem(), "");
4342 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4343 break;
4344 }
4345 }
4346
4347 //------------------------------final_graph_reshaping_walk---------------------
4348 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4349 // requires that the walk visits a node's inputs before visiting the node.
4350 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
4351 Unique_Node_List sfpt;
4352
4353 frc._visited.set(root->_idx); // first, mark node as visited
4354 uint cnt = root->req();
4355 Node *n = root;
4356 uint i = 0;
4357 while (true) {
4358 if (i < cnt) {
4359 // Place all non-visited non-null inputs onto stack
4360 Node* m = n->in(i);
4361 ++i;
4362 if (m != nullptr && !frc._visited.test_set(m->_idx)) {
4363 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) {
4364 // compute worst case interpreter size in case of a deoptimization
4365 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4366
4367 sfpt.push(m);
4368 }
4369 cnt = m->req();
4370 nstack.push(n, i); // put on stack parent and next input's index
4371 n = m;
4372 i = 0;
4373 }
4374 } else {
4375 // Now do post-visit work
4376 final_graph_reshaping_impl(n, frc, dead_nodes);
4377 if (nstack.is_empty())
4378 break; // finished
4379 n = nstack.node(); // Get node from stack
4380 cnt = n->req();
4381 i = nstack.index();
4382 nstack.pop(); // Shift to the next node on stack
4383 }
4384 }
4385
4386 // Skip next transformation if compressed oops are not used.
4387 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4388 (!UseCompressedOops && !UseCompressedClassPointers))
4389 return;
4390
4391 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4392 // It could be done for an uncommon traps or any safepoints/calls
4393 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4394 while (sfpt.size() > 0) {
4395 n = sfpt.pop();
4396 JVMState *jvms = n->as_SafePoint()->jvms();
4397 assert(jvms != nullptr, "sanity");
4398 int start = jvms->debug_start();
4399 int end = n->req();
4400 bool is_uncommon = (n->is_CallStaticJava() &&
4401 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4402 for (int j = start; j < end; j++) {
4403 Node* in = n->in(j);
4404 if (in->is_DecodeNarrowPtr()) {
4405 bool safe_to_skip = true;
4406 if (!is_uncommon ) {
4407 // Is it safe to skip?
4408 for (uint i = 0; i < in->outcnt(); i++) {
4409 Node* u = in->raw_out(i);
4410 if (!u->is_SafePoint() ||
4411 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4412 safe_to_skip = false;
4413 }
4414 }
4415 }
4416 if (safe_to_skip) {
4417 n->set_req(j, in->in(1));
4418 }
4419 if (in->outcnt() == 0) {
4420 in->disconnect_inputs(this);
4421 }
4422 }
4423 }
4424 }
4425 }
4426
4427 //------------------------------final_graph_reshaping--------------------------
4428 // Final Graph Reshaping.
4429 //
4430 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4431 // and not commoned up and forced early. Must come after regular
4432 // optimizations to avoid GVN undoing the cloning. Clone constant
4433 // inputs to Loop Phis; these will be split by the allocator anyways.
4434 // Remove Opaque nodes.
4435 // (2) Move last-uses by commutative operations to the left input to encourage
4436 // Intel update-in-place two-address operations and better register usage
4437 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4438 // calls canonicalizing them back.
4439 // (3) Count the number of double-precision FP ops, single-precision FP ops
4440 // and call sites. On Intel, we can get correct rounding either by
4441 // forcing singles to memory (requires extra stores and loads after each
4442 // FP bytecode) or we can set a rounding mode bit (requires setting and
4443 // clearing the mode bit around call sites). The mode bit is only used
4444 // if the relative frequency of single FP ops to calls is low enough.
4445 // This is a key transform for SPEC mpeg_audio.
4446 // (4) Detect infinite loops; blobs of code reachable from above but not
4447 // below. Several of the Code_Gen algorithms fail on such code shapes,
4448 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4449 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4450 // Detection is by looking for IfNodes where only 1 projection is
4451 // reachable from below or CatchNodes missing some targets.
4452 // (5) Assert for insane oop offsets in debug mode.
4453
4454 bool Compile::final_graph_reshaping() {
4455 // an infinite loop may have been eliminated by the optimizer,
4456 // in which case the graph will be empty.
4457 if (root()->req() == 1) {
4458 // Do not compile method that is only a trivial infinite loop,
4459 // since the content of the loop may have been eliminated.
4460 record_method_not_compilable("trivial infinite loop");
4461 return true;
4462 }
4463
4464 // Expensive nodes have their control input set to prevent the GVN
4465 // from freely commoning them. There's no GVN beyond this point so
4466 // no need to keep the control input. We want the expensive nodes to
4467 // be freely moved to the least frequent code path by gcm.
4468 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4469 for (int i = 0; i < expensive_count(); i++) {
4470 _expensive_nodes.at(i)->set_req(0, nullptr);
4471 }
4472
4473 Final_Reshape_Counts frc;
4474
4475 // Visit everybody reachable!
4476 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4477 Node_Stack nstack(live_nodes() >> 1);
4478 Unique_Node_List dead_nodes;
4479 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
4480
4481 // Check for unreachable (from below) code (i.e., infinite loops).
4482 for( uint i = 0; i < frc._tests.size(); i++ ) {
4483 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4484 // Get number of CFG targets.
4485 // Note that PCTables include exception targets after calls.
4486 uint required_outcnt = n->required_outcnt();
4487 if (n->outcnt() != required_outcnt) {
4488 // Check for a few special cases. Rethrow Nodes never take the
4489 // 'fall-thru' path, so expected kids is 1 less.
4490 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4491 if (n->in(0)->in(0)->is_Call()) {
4492 CallNode* call = n->in(0)->in(0)->as_Call();
4493 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4494 required_outcnt--; // Rethrow always has 1 less kid
4495 } else if (call->req() > TypeFunc::Parms &&
4496 call->is_CallDynamicJava()) {
4497 // Check for null receiver. In such case, the optimizer has
4498 // detected that the virtual call will always result in a null
4499 // pointer exception. The fall-through projection of this CatchNode
4500 // will not be populated.
4501 Node* arg0 = call->in(TypeFunc::Parms);
4502 if (arg0->is_Type() &&
4503 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4504 required_outcnt--;
4505 }
4506 } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
4507 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4508 // Check for illegal array length. In such case, the optimizer has
4509 // detected that the allocation attempt will always result in an
4510 // exception. There is no fall-through projection of this CatchNode .
4511 assert(call->is_CallStaticJava(), "static call expected");
4512 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4513 uint valid_length_test_input = call->req() - 1;
4514 Node* valid_length_test = call->in(valid_length_test_input);
4515 call->del_req(valid_length_test_input);
4516 if (valid_length_test->find_int_con(1) == 0) {
4517 required_outcnt--;
4518 }
4519 dead_nodes.push(valid_length_test);
4520 assert(n->outcnt() == required_outcnt, "malformed control flow");
4521 continue;
4522 }
4523 }
4524 }
4525
4526 // Recheck with a better notion of 'required_outcnt'
4527 if (n->outcnt() != required_outcnt) {
4528 record_method_not_compilable("malformed control flow");
4529 return true; // Not all targets reachable!
4530 }
4531 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
4532 CallNode* call = n->in(0)->in(0)->as_Call();
4533 if (call->entry_point() == OptoRuntime::new_array_Java() ||
4534 call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
4535 assert(call->is_CallStaticJava(), "static call expected");
4536 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
4537 uint valid_length_test_input = call->req() - 1;
4538 dead_nodes.push(call->in(valid_length_test_input));
4539 call->del_req(valid_length_test_input); // valid length test useless now
4540 }
4541 }
4542 // Check that I actually visited all kids. Unreached kids
4543 // must be infinite loops.
4544 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4545 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4546 record_method_not_compilable("infinite loop");
4547 return true; // Found unvisited kid; must be unreach
4548 }
4549
4550 // Here so verification code in final_graph_reshaping_walk()
4551 // always see an OuterStripMinedLoopEnd
4552 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4553 IfNode* init_iff = n->as_If();
4554 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4555 n->subsume_by(iff, this);
4556 }
4557 }
4558
4559 while (dead_nodes.size() > 0) {
4560 Node* m = dead_nodes.pop();
4561 if (m->outcnt() == 0 && m != top()) {
4562 for (uint j = 0; j < m->req(); j++) {
4563 Node* in = m->in(j);
4564 if (in != nullptr) {
4565 dead_nodes.push(in);
4566 }
4567 }
4568 m->disconnect_inputs(this);
4569 }
4570 }
4571
4572 #ifdef IA32
4573 // If original bytecodes contained a mixture of floats and doubles
4574 // check if the optimizer has made it homogeneous, item (3).
4575 if (UseSSE == 0 &&
4576 frc.get_float_count() > 32 &&
4577 frc.get_double_count() == 0 &&
4578 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4579 set_24_bit_selection_and_mode(false, true);
4580 }
4581 #endif // IA32
4582
4583 set_java_calls(frc.get_java_call_count());
4584 set_inner_loops(frc.get_inner_loop_count());
4585
4586 // No infinite loops, no reason to bail out.
4587 return false;
4588 }
4589
4590 //-----------------------------too_many_traps----------------------------------
4591 // Report if there are too many traps at the current method and bci.
4592 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4593 bool Compile::too_many_traps(ciMethod* method,
4594 int bci,
4595 Deoptimization::DeoptReason reason) {
4596 ciMethodData* md = method->method_data();
4597 if (md->is_empty()) {
4598 // Assume the trap has not occurred, or that it occurred only
4599 // because of a transient condition during start-up in the interpreter.
4600 return false;
4601 }
4602 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4603 if (md->has_trap_at(bci, m, reason) != 0) {
4604 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4605 // Also, if there are multiple reasons, or if there is no per-BCI record,
4606 // assume the worst.
4607 if (log())
4608 log()->elem("observe trap='%s' count='%d'",
4609 Deoptimization::trap_reason_name(reason),
4610 md->trap_count(reason));
4611 return true;
4612 } else {
4613 // Ignore method/bci and see if there have been too many globally.
4614 return too_many_traps(reason, md);
4615 }
4616 }
4617
4618 // Less-accurate variant which does not require a method and bci.
4619 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4620 ciMethodData* logmd) {
4621 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4622 // Too many traps globally.
4623 // Note that we use cumulative trap_count, not just md->trap_count.
4624 if (log()) {
4625 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason);
4626 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4627 Deoptimization::trap_reason_name(reason),
4628 mcount, trap_count(reason));
4629 }
4630 return true;
4631 } else {
4632 // The coast is clear.
4633 return false;
4634 }
4635 }
4636
4637 //--------------------------too_many_recompiles--------------------------------
4638 // Report if there are too many recompiles at the current method and bci.
4639 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4640 // Is not eager to return true, since this will cause the compiler to use
4641 // Action_none for a trap point, to avoid too many recompilations.
4642 bool Compile::too_many_recompiles(ciMethod* method,
4643 int bci,
4644 Deoptimization::DeoptReason reason) {
4645 ciMethodData* md = method->method_data();
4646 if (md->is_empty()) {
4647 // Assume the trap has not occurred, or that it occurred only
4648 // because of a transient condition during start-up in the interpreter.
4649 return false;
4650 }
4651 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4652 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4653 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4654 Deoptimization::DeoptReason per_bc_reason
4655 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4656 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr;
4657 if ((per_bc_reason == Deoptimization::Reason_none
4658 || md->has_trap_at(bci, m, reason) != 0)
4659 // The trap frequency measure we care about is the recompile count:
4660 && md->trap_recompiled_at(bci, m)
4661 && md->overflow_recompile_count() >= bc_cutoff) {
4662 // Do not emit a trap here if it has already caused recompilations.
4663 // Also, if there are multiple reasons, or if there is no per-BCI record,
4664 // assume the worst.
4665 if (log())
4666 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4667 Deoptimization::trap_reason_name(reason),
4668 md->trap_count(reason),
4669 md->overflow_recompile_count());
4670 return true;
4671 } else if (trap_count(reason) != 0
4672 && decompile_count() >= m_cutoff) {
4673 // Too many recompiles globally, and we have seen this sort of trap.
4674 // Use cumulative decompile_count, not just md->decompile_count.
4675 if (log())
4676 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4677 Deoptimization::trap_reason_name(reason),
4678 md->trap_count(reason), trap_count(reason),
4679 md->decompile_count(), decompile_count());
4680 return true;
4681 } else {
4682 // The coast is clear.
4683 return false;
4684 }
4685 }
4686
4687 // Compute when not to trap. Used by matching trap based nodes and
4688 // NullCheck optimization.
4689 void Compile::set_allowed_deopt_reasons() {
4690 _allowed_reasons = 0;
4691 if (is_method_compilation()) {
4692 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4693 assert(rs < BitsPerInt, "recode bit map");
4694 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4695 _allowed_reasons |= nth_bit(rs);
4696 }
4697 }
4698 }
4699 }
4700
4701 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4702 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4703 }
4704
4705 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4706 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4707 }
4708
4709 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4710 if (holder->is_initialized()) {
4711 return false;
4712 }
4713 if (holder->is_being_initialized()) {
4714 if (accessing_method->holder() == holder) {
4715 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4716 // <init>, or a static method. In all those cases, there was an initialization
4717 // barrier on the holder klass passed.
4718 if (accessing_method->is_class_initializer() ||
4719 accessing_method->is_object_constructor() ||
4720 accessing_method->is_static()) {
4721 return false;
4722 }
4723 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4724 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4725 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4726 // child class can become fully initialized while its parent class is still being initialized.
4727 if (accessing_method->is_class_initializer()) {
4728 return false;
4729 }
4730 }
4731 ciMethod* root = method(); // the root method of compilation
4732 if (root != accessing_method) {
4733 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4734 }
4735 }
4736 return true;
4737 }
4738
4739 #ifndef PRODUCT
4740 //------------------------------verify_bidirectional_edges---------------------
4741 // For each input edge to a node (ie - for each Use-Def edge), verify that
4742 // there is a corresponding Def-Use edge.
4743 void Compile::verify_bidirectional_edges(Unique_Node_List& visited, const Unique_Node_List* root_and_safepoints) const {
4744 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4745 uint stack_size = live_nodes() >> 4;
4746 Node_List nstack(MAX2(stack_size, (uint) OptoNodeListSize));
4747 if (root_and_safepoints != nullptr) {
4748 assert(root_and_safepoints->member(_root), "root is not in root_and_safepoints");
4749 for (uint i = 0, limit = root_and_safepoints->size(); i < limit; i++) {
4750 Node* root_or_safepoint = root_and_safepoints->at(i);
4751 // If the node is a safepoint, let's check if it still has a control input
4752 // Lack of control input signifies that this node was killed by CCP or
4753 // recursively by remove_globally_dead_node and it shouldn't be a starting
4754 // point.
4755 if (!root_or_safepoint->is_SafePoint() || root_or_safepoint->in(0) != nullptr) {
4756 nstack.push(root_or_safepoint);
4757 }
4758 }
4759 } else {
4760 nstack.push(_root);
4761 }
4762
4763 while (nstack.size() > 0) {
4764 Node* n = nstack.pop();
4765 if (visited.member(n)) {
4766 continue;
4767 }
4768 visited.push(n);
4769
4770 // Walk over all input edges, checking for correspondence
4771 uint length = n->len();
4772 for (uint i = 0; i < length; i++) {
4773 Node* in = n->in(i);
4774 if (in != nullptr && !visited.member(in)) {
4775 nstack.push(in); // Put it on stack
4776 }
4777 if (in != nullptr && !in->is_top()) {
4778 // Count instances of `next`
4779 int cnt = 0;
4780 for (uint idx = 0; idx < in->_outcnt; idx++) {
4781 if (in->_out[idx] == n) {
4782 cnt++;
4783 }
4784 }
4785 assert(cnt > 0, "Failed to find Def-Use edge.");
4786 // Check for duplicate edges
4787 // walk the input array downcounting the input edges to n
4788 for (uint j = 0; j < length; j++) {
4789 if (n->in(j) == in) {
4790 cnt--;
4791 }
4792 }
4793 assert(cnt == 0, "Mismatched edge count.");
4794 } else if (in == nullptr) {
4795 assert(i == 0 || i >= n->req() ||
4796 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4797 (n->is_Allocate() && i >= AllocateNode::InlineType) ||
4798 (n->is_Unlock() && i == (n->req() - 1)) ||
4799 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4800 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges");
4801 } else {
4802 assert(in->is_top(), "sanity");
4803 // Nothing to check.
4804 }
4805 }
4806 }
4807 }
4808
4809 //------------------------------verify_graph_edges---------------------------
4810 // Walk the Graph and verify that there is a one-to-one correspondence
4811 // between Use-Def edges and Def-Use edges in the graph.
4812 void Compile::verify_graph_edges(bool no_dead_code, const Unique_Node_List* root_and_safepoints) const {
4813 if (VerifyGraphEdges) {
4814 Unique_Node_List visited;
4815
4816 // Call graph walk to check edges
4817 verify_bidirectional_edges(visited, root_and_safepoints);
4818 if (no_dead_code) {
4819 // Now make sure that no visited node is used by an unvisited node.
4820 bool dead_nodes = false;
4821 Unique_Node_List checked;
4822 while (visited.size() > 0) {
4823 Node* n = visited.pop();
4824 checked.push(n);
4825 for (uint i = 0; i < n->outcnt(); i++) {
4826 Node* use = n->raw_out(i);
4827 if (checked.member(use)) continue; // already checked
4828 if (visited.member(use)) continue; // already in the graph
4829 if (use->is_Con()) continue; // a dead ConNode is OK
4830 // At this point, we have found a dead node which is DU-reachable.
4831 if (!dead_nodes) {
4832 tty->print_cr("*** Dead nodes reachable via DU edges:");
4833 dead_nodes = true;
4834 }
4835 use->dump(2);
4836 tty->print_cr("---");
4837 checked.push(use); // No repeats; pretend it is now checked.
4838 }
4839 }
4840 assert(!dead_nodes, "using nodes must be reachable from root");
4841 }
4842 }
4843 }
4844 #endif
4845
4846 // The Compile object keeps track of failure reasons separately from the ciEnv.
4847 // This is required because there is not quite a 1-1 relation between the
4848 // ciEnv and its compilation task and the Compile object. Note that one
4849 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4850 // to backtrack and retry without subsuming loads. Other than this backtracking
4851 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4852 // by the logic in C2Compiler.
4853 void Compile::record_failure(const char* reason DEBUG_ONLY(COMMA bool allow_multiple_failures)) {
4854 if (log() != nullptr) {
4855 log()->elem("failure reason='%s' phase='compile'", reason);
4856 }
4857 if (_failure_reason.get() == nullptr) {
4858 // Record the first failure reason.
4859 _failure_reason.set(reason);
4860 if (CaptureBailoutInformation) {
4861 _first_failure_details = new CompilationFailureInfo(reason);
4862 }
4863 } else {
4864 assert(!StressBailout || allow_multiple_failures, "should have handled previous failure.");
4865 }
4866
4867 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4868 C->print_method(PHASE_FAILURE, 1);
4869 }
4870 _root = nullptr; // flush the graph, too
4871 }
4872
4873 Compile::TracePhase::TracePhase(const char* name, PhaseTraceId id)
4874 : TraceTime(name, &Phase::timers[id], CITime, CITimeVerbose),
4875 _compile(Compile::current()),
4876 _log(nullptr),
4877 _dolog(CITimeVerbose)
4878 {
4879 assert(_compile != nullptr, "sanity check");
4880 assert(id != PhaseTraceId::_t_none, "Don't use none");
4881 if (_dolog) {
4882 _log = _compile->log();
4883 }
4884 if (_log != nullptr) {
4885 _log->begin_head("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4886 _log->stamp();
4887 _log->end_head();
4888 }
4889
4890 // Inform memory statistic, if enabled
4891 if (CompilationMemoryStatistic::enabled()) {
4892 CompilationMemoryStatistic::on_phase_start((int)id, name);
4893 }
4894 }
4895
4896 Compile::TracePhase::TracePhase(PhaseTraceId id)
4897 : TracePhase(Phase::get_phase_trace_id_text(id), id) {}
4898
4899 Compile::TracePhase::~TracePhase() {
4900
4901 // Inform memory statistic, if enabled
4902 if (CompilationMemoryStatistic::enabled()) {
4903 CompilationMemoryStatistic::on_phase_end();
4904 }
4905
4906 if (_compile->failing_internal()) {
4907 if (_log != nullptr) {
4908 _log->done("phase");
4909 }
4910 return; // timing code, not stressing bailouts.
4911 }
4912 #ifdef ASSERT
4913 if (PrintIdealNodeCount) {
4914 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4915 phase_name(), _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk());
4916 }
4917
4918 if (VerifyIdealNodeCount) {
4919 _compile->print_missing_nodes();
4920 }
4921 #endif
4922
4923 if (_log != nullptr) {
4924 _log->done("phase name='%s' nodes='%d' live='%d'", phase_name(), _compile->unique(), _compile->live_nodes());
4925 }
4926 }
4927
4928 //----------------------------static_subtype_check-----------------------------
4929 // Shortcut important common cases when superklass is exact:
4930 // (0) superklass is java.lang.Object (can occur in reflective code)
4931 // (1) subklass is already limited to a subtype of superklass => always ok
4932 // (2) subklass does not overlap with superklass => always fail
4933 // (3) superklass has NO subtypes and we can check with a simple compare.
4934 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4935 if (skip) {
4936 return SSC_full_test; // Let caller generate the general case.
4937 }
4938
4939 if (subk->is_java_subtype_of(superk)) {
4940 return SSC_always_true; // (0) and (1) this test cannot fail
4941 }
4942
4943 if (!subk->maybe_java_subtype_of(superk)) {
4944 return SSC_always_false; // (2) true path dead; no dynamic test needed
4945 }
4946
4947 const Type* superelem = superk;
4948 if (superk->isa_aryklassptr()) {
4949 int ignored;
4950 superelem = superk->is_aryklassptr()->base_element_type(ignored);
4951
4952 // Do not fold the subtype check to an array klass pointer comparison for null-able inline type arrays
4953 // because null-free [LMyValue <: null-able [LMyValue but the klasses are different. Perform a full test.
4954 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() &&
4955 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) {
4956 return SSC_full_test;
4957 }
4958 }
4959
4960 if (superelem->isa_instklassptr()) {
4961 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4962 if (!ik->has_subklass()) {
4963 if (!ik->is_final()) {
4964 // Add a dependency if there is a chance of a later subclass.
4965 dependencies()->assert_leaf_type(ik);
4966 }
4967 if (!superk->maybe_java_subtype_of(subk)) {
4968 return SSC_always_false;
4969 }
4970 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4971 }
4972 } else {
4973 // A primitive array type has no subtypes.
4974 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4975 }
4976
4977 return SSC_full_test;
4978 }
4979
4980 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4981 #ifdef _LP64
4982 // The scaled index operand to AddP must be a clean 64-bit value.
4983 // Java allows a 32-bit int to be incremented to a negative
4984 // value, which appears in a 64-bit register as a large
4985 // positive number. Using that large positive number as an
4986 // operand in pointer arithmetic has bad consequences.
4987 // On the other hand, 32-bit overflow is rare, and the possibility
4988 // can often be excluded, if we annotate the ConvI2L node with
4989 // a type assertion that its value is known to be a small positive
4990 // number. (The prior range check has ensured this.)
4991 // This assertion is used by ConvI2LNode::Ideal.
4992 int index_max = max_jint - 1; // array size is max_jint, index is one less
4993 if (sizetype != nullptr) index_max = sizetype->_hi - 1;
4994 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4995 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4996 #endif
4997 return idx;
4998 }
4999
5000 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
5001 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
5002 if (ctrl != nullptr) {
5003 // Express control dependency by a CastII node with a narrow type.
5004 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
5005 // node from floating above the range check during loop optimizations. Otherwise, the
5006 // ConvI2L node may be eliminated independently of the range check, causing the data path
5007 // to become TOP while the control path is still there (although it's unreachable).
5008 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
5009 value = phase->transform(value);
5010 }
5011 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
5012 return phase->transform(new ConvI2LNode(value, ltype));
5013 }
5014
5015 void Compile::dump_print_inlining() {
5016 inline_printer()->print_on(tty);
5017 }
5018
5019 void Compile::log_late_inline(CallGenerator* cg) {
5020 if (log() != nullptr) {
5021 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
5022 cg->unique_id());
5023 JVMState* p = cg->call_node()->jvms();
5024 while (p != nullptr) {
5025 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
5026 p = p->caller();
5027 }
5028 log()->tail("late_inline");
5029 }
5030 }
5031
5032 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
5033 log_late_inline(cg);
5034 if (log() != nullptr) {
5035 log()->inline_fail(msg);
5036 }
5037 }
5038
5039 void Compile::log_inline_id(CallGenerator* cg) {
5040 if (log() != nullptr) {
5041 // The LogCompilation tool needs a unique way to identify late
5042 // inline call sites. This id must be unique for this call site in
5043 // this compilation. Try to have it unique across compilations as
5044 // well because it can be convenient when grepping through the log
5045 // file.
5046 // Distinguish OSR compilations from others in case CICountOSR is
5047 // on.
5048 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
5049 cg->set_unique_id(id);
5050 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
5051 }
5052 }
5053
5054 void Compile::log_inline_failure(const char* msg) {
5055 if (C->log() != nullptr) {
5056 C->log()->inline_fail(msg);
5057 }
5058 }
5059
5060
5061 // Dump inlining replay data to the stream.
5062 // Don't change thread state and acquire any locks.
5063 void Compile::dump_inline_data(outputStream* out) {
5064 InlineTree* inl_tree = ilt();
5065 if (inl_tree != nullptr) {
5066 out->print(" inline %d", inl_tree->count());
5067 inl_tree->dump_replay_data(out);
5068 }
5069 }
5070
5071 void Compile::dump_inline_data_reduced(outputStream* out) {
5072 assert(ReplayReduce, "");
5073
5074 InlineTree* inl_tree = ilt();
5075 if (inl_tree == nullptr) {
5076 return;
5077 }
5078 // Enable iterative replay file reduction
5079 // Output "compile" lines for depth 1 subtrees,
5080 // simulating that those trees were compiled
5081 // instead of inlined.
5082 for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
5083 InlineTree* sub = inl_tree->subtrees().at(i);
5084 if (sub->inline_level() != 1) {
5085 continue;
5086 }
5087
5088 ciMethod* method = sub->method();
5089 int entry_bci = -1;
5090 int comp_level = env()->task()->comp_level();
5091 out->print("compile ");
5092 method->dump_name_as_ascii(out);
5093 out->print(" %d %d", entry_bci, comp_level);
5094 out->print(" inline %d", sub->count());
5095 sub->dump_replay_data(out, -1);
5096 out->cr();
5097 }
5098 }
5099
5100 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
5101 if (n1->Opcode() < n2->Opcode()) return -1;
5102 else if (n1->Opcode() > n2->Opcode()) return 1;
5103
5104 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
5105 for (uint i = 1; i < n1->req(); i++) {
5106 if (n1->in(i) < n2->in(i)) return -1;
5107 else if (n1->in(i) > n2->in(i)) return 1;
5108 }
5109
5110 return 0;
5111 }
5112
5113 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
5114 Node* n1 = *n1p;
5115 Node* n2 = *n2p;
5116
5117 return cmp_expensive_nodes(n1, n2);
5118 }
5119
5120 void Compile::sort_expensive_nodes() {
5121 if (!expensive_nodes_sorted()) {
5122 _expensive_nodes.sort(cmp_expensive_nodes);
5123 }
5124 }
5125
5126 bool Compile::expensive_nodes_sorted() const {
5127 for (int i = 1; i < _expensive_nodes.length(); i++) {
5128 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
5129 return false;
5130 }
5131 }
5132 return true;
5133 }
5134
5135 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5136 if (_expensive_nodes.length() == 0) {
5137 return false;
5138 }
5139
5140 assert(OptimizeExpensiveOps, "optimization off?");
5141
5142 // Take this opportunity to remove dead nodes from the list
5143 int j = 0;
5144 for (int i = 0; i < _expensive_nodes.length(); i++) {
5145 Node* n = _expensive_nodes.at(i);
5146 if (!n->is_unreachable(igvn)) {
5147 assert(n->is_expensive(), "should be expensive");
5148 _expensive_nodes.at_put(j, n);
5149 j++;
5150 }
5151 }
5152 _expensive_nodes.trunc_to(j);
5153
5154 // Then sort the list so that similar nodes are next to each other
5155 // and check for at least two nodes of identical kind with same data
5156 // inputs.
5157 sort_expensive_nodes();
5158
5159 for (int i = 0; i < _expensive_nodes.length()-1; i++) {
5160 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
5161 return true;
5162 }
5163 }
5164
5165 return false;
5166 }
5167
5168 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5169 if (_expensive_nodes.length() == 0) {
5170 return;
5171 }
5172
5173 assert(OptimizeExpensiveOps, "optimization off?");
5174
5175 // Sort to bring similar nodes next to each other and clear the
5176 // control input of nodes for which there's only a single copy.
5177 sort_expensive_nodes();
5178
5179 int j = 0;
5180 int identical = 0;
5181 int i = 0;
5182 bool modified = false;
5183 for (; i < _expensive_nodes.length()-1; i++) {
5184 assert(j <= i, "can't write beyond current index");
5185 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
5186 identical++;
5187 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5188 continue;
5189 }
5190 if (identical > 0) {
5191 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5192 identical = 0;
5193 } else {
5194 Node* n = _expensive_nodes.at(i);
5195 igvn.replace_input_of(n, 0, nullptr);
5196 igvn.hash_insert(n);
5197 modified = true;
5198 }
5199 }
5200 if (identical > 0) {
5201 _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
5202 } else if (_expensive_nodes.length() >= 1) {
5203 Node* n = _expensive_nodes.at(i);
5204 igvn.replace_input_of(n, 0, nullptr);
5205 igvn.hash_insert(n);
5206 modified = true;
5207 }
5208 _expensive_nodes.trunc_to(j);
5209 if (modified) {
5210 igvn.optimize();
5211 }
5212 }
5213
5214 void Compile::add_expensive_node(Node * n) {
5215 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
5216 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5217 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5218 if (OptimizeExpensiveOps) {
5219 _expensive_nodes.append(n);
5220 } else {
5221 // Clear control input and let IGVN optimize expensive nodes if
5222 // OptimizeExpensiveOps is off.
5223 n->set_req(0, nullptr);
5224 }
5225 }
5226
5227 /**
5228 * Track coarsened Lock and Unlock nodes.
5229 */
5230
5231 class Lock_List : public Node_List {
5232 uint _origin_cnt;
5233 public:
5234 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
5235 uint origin_cnt() const { return _origin_cnt; }
5236 };
5237
5238 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
5239 int length = locks.length();
5240 if (length > 0) {
5241 // Have to keep this list until locks elimination during Macro nodes elimination.
5242 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
5243 AbstractLockNode* alock = locks.at(0);
5244 BoxLockNode* box = alock->box_node()->as_BoxLock();
5245 for (int i = 0; i < length; i++) {
5246 AbstractLockNode* lock = locks.at(i);
5247 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
5248 locks_list->push(lock);
5249 BoxLockNode* this_box = lock->box_node()->as_BoxLock();
5250 if (this_box != box) {
5251 // Locking regions (BoxLock) could be Unbalanced here:
5252 // - its coarsened locks were eliminated in earlier
5253 // macro nodes elimination followed by loop unroll
5254 // - it is OSR locking region (no Lock node)
5255 // Preserve Unbalanced status in such cases.
5256 if (!this_box->is_unbalanced()) {
5257 this_box->set_coarsened();
5258 }
5259 if (!box->is_unbalanced()) {
5260 box->set_coarsened();
5261 }
5262 }
5263 }
5264 _coarsened_locks.append(locks_list);
5265 }
5266 }
5267
5268 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
5269 int count = coarsened_count();
5270 for (int i = 0; i < count; i++) {
5271 Node_List* locks_list = _coarsened_locks.at(i);
5272 for (uint j = 0; j < locks_list->size(); j++) {
5273 Node* lock = locks_list->at(j);
5274 assert(lock->is_AbstractLock(), "sanity");
5275 if (!useful.member(lock)) {
5276 locks_list->yank(lock);
5277 }
5278 }
5279 }
5280 }
5281
5282 void Compile::remove_coarsened_lock(Node* n) {
5283 if (n->is_AbstractLock()) {
5284 int count = coarsened_count();
5285 for (int i = 0; i < count; i++) {
5286 Node_List* locks_list = _coarsened_locks.at(i);
5287 locks_list->yank(n);
5288 }
5289 }
5290 }
5291
5292 bool Compile::coarsened_locks_consistent() {
5293 int count = coarsened_count();
5294 for (int i = 0; i < count; i++) {
5295 bool unbalanced = false;
5296 bool modified = false; // track locks kind modifications
5297 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
5298 uint size = locks_list->size();
5299 if (size == 0) {
5300 unbalanced = false; // All locks were eliminated - good
5301 } else if (size != locks_list->origin_cnt()) {
5302 unbalanced = true; // Some locks were removed from list
5303 } else {
5304 for (uint j = 0; j < size; j++) {
5305 Node* lock = locks_list->at(j);
5306 // All nodes in group should have the same state (modified or not)
5307 if (!lock->as_AbstractLock()->is_coarsened()) {
5308 if (j == 0) {
5309 // first on list was modified, the rest should be too for consistency
5310 modified = true;
5311 } else if (!modified) {
5312 // this lock was modified but previous locks on the list were not
5313 unbalanced = true;
5314 break;
5315 }
5316 } else if (modified) {
5317 // previous locks on list were modified but not this lock
5318 unbalanced = true;
5319 break;
5320 }
5321 }
5322 }
5323 if (unbalanced) {
5324 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
5325 #ifdef ASSERT
5326 if (PrintEliminateLocks) {
5327 tty->print_cr("=== unbalanced coarsened locks ===");
5328 for (uint l = 0; l < size; l++) {
5329 locks_list->at(l)->dump();
5330 }
5331 }
5332 #endif
5333 record_failure(C2Compiler::retry_no_locks_coarsening());
5334 return false;
5335 }
5336 }
5337 return true;
5338 }
5339
5340 // Mark locking regions (identified by BoxLockNode) as unbalanced if
5341 // locks coarsening optimization removed Lock/Unlock nodes from them.
5342 // Such regions become unbalanced because coarsening only removes part
5343 // of Lock/Unlock nodes in region. As result we can't execute other
5344 // locks elimination optimizations which assume all code paths have
5345 // corresponding pair of Lock/Unlock nodes - they are balanced.
5346 void Compile::mark_unbalanced_boxes() const {
5347 int count = coarsened_count();
5348 for (int i = 0; i < count; i++) {
5349 Node_List* locks_list = _coarsened_locks.at(i);
5350 uint size = locks_list->size();
5351 if (size > 0) {
5352 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock();
5353 BoxLockNode* box = alock->box_node()->as_BoxLock();
5354 if (alock->is_coarsened()) {
5355 // coarsened_locks_consistent(), which is called before this method, verifies
5356 // that the rest of Lock/Unlock nodes on locks_list are also coarsened.
5357 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5358 for (uint j = 1; j < size; j++) {
5359 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here");
5360 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock();
5361 if (box != this_box) {
5362 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated");
5363 box->set_unbalanced();
5364 this_box->set_unbalanced();
5365 }
5366 }
5367 }
5368 }
5369 }
5370 }
5371
5372 /**
5373 * Remove the speculative part of types and clean up the graph
5374 */
5375 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5376 if (UseTypeSpeculation) {
5377 Unique_Node_List worklist;
5378 worklist.push(root());
5379 int modified = 0;
5380 // Go over all type nodes that carry a speculative type, drop the
5381 // speculative part of the type and enqueue the node for an igvn
5382 // which may optimize it out.
5383 for (uint next = 0; next < worklist.size(); ++next) {
5384 Node *n = worklist.at(next);
5385 if (n->is_Type()) {
5386 TypeNode* tn = n->as_Type();
5387 const Type* t = tn->type();
5388 const Type* t_no_spec = t->remove_speculative();
5389 if (t_no_spec != t) {
5390 bool in_hash = igvn.hash_delete(n);
5391 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
5392 tn->set_type(t_no_spec);
5393 igvn.hash_insert(n);
5394 igvn._worklist.push(n); // give it a chance to go away
5395 modified++;
5396 }
5397 }
5398 // Iterate over outs - endless loops is unreachable from below
5399 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5400 Node *m = n->fast_out(i);
5401 if (not_a_node(m)) {
5402 continue;
5403 }
5404 worklist.push(m);
5405 }
5406 }
5407 // Drop the speculative part of all types in the igvn's type table
5408 igvn.remove_speculative_types();
5409 if (modified > 0) {
5410 igvn.optimize();
5411 if (failing()) return;
5412 }
5413 #ifdef ASSERT
5414 // Verify that after the IGVN is over no speculative type has resurfaced
5415 worklist.clear();
5416 worklist.push(root());
5417 for (uint next = 0; next < worklist.size(); ++next) {
5418 Node *n = worklist.at(next);
5419 const Type* t = igvn.type_or_null(n);
5420 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types");
5421 if (n->is_Type()) {
5422 t = n->as_Type()->type();
5423 assert(t == t->remove_speculative(), "no more speculative types");
5424 }
5425 // Iterate over outs - endless loops is unreachable from below
5426 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5427 Node *m = n->fast_out(i);
5428 if (not_a_node(m)) {
5429 continue;
5430 }
5431 worklist.push(m);
5432 }
5433 }
5434 igvn.check_no_speculative_types();
5435 #endif
5436 }
5437 }
5438
5439 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5440 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5441 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5442 if (!EnableValhalla || ta == nullptr || tb == nullptr ||
5443 ta->is_zero_type() || tb->is_zero_type() ||
5444 !ta->can_be_inline_type() || !tb->can_be_inline_type()) {
5445 // Use old acmp if one operand is null or not an inline type
5446 return new CmpPNode(a, b);
5447 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) {
5448 // We know that one operand is an inline type. Therefore,
5449 // new acmp will only return true if both operands are nullptr.
5450 // Check if both operands are null by or'ing the oops.
5451 a = phase->transform(new CastP2XNode(nullptr, a));
5452 b = phase->transform(new CastP2XNode(nullptr, b));
5453 a = phase->transform(new OrXNode(a, b));
5454 return new CmpXNode(a, phase->MakeConX(0));
5455 }
5456 // Use new acmp
5457 return nullptr;
5458 }
5459
5460 // Auxiliary methods to support randomized stressing/fuzzing.
5461
5462 void Compile::initialize_stress_seed(const DirectiveSet* directive) {
5463 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) {
5464 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
5465 FLAG_SET_ERGO(StressSeed, _stress_seed);
5466 } else {
5467 _stress_seed = StressSeed;
5468 }
5469 if (_log != nullptr) {
5470 _log->elem("stress_test seed='%u'", _stress_seed);
5471 }
5472 }
5473
5474 int Compile::random() {
5475 _stress_seed = os::next_random(_stress_seed);
5476 return static_cast<int>(_stress_seed);
5477 }
5478
5479 // This method can be called the arbitrary number of times, with current count
5480 // as the argument. The logic allows selecting a single candidate from the
5481 // running list of candidates as follows:
5482 // int count = 0;
5483 // Cand* selected = null;
5484 // while(cand = cand->next()) {
5485 // if (randomized_select(++count)) {
5486 // selected = cand;
5487 // }
5488 // }
5489 //
5490 // Including count equalizes the chances any candidate is "selected".
5491 // This is useful when we don't have the complete list of candidates to choose
5492 // from uniformly. In this case, we need to adjust the randomicity of the
5493 // selection, or else we will end up biasing the selection towards the latter
5494 // candidates.
5495 //
5496 // Quick back-envelope calculation shows that for the list of n candidates
5497 // the equal probability for the candidate to persist as "best" can be
5498 // achieved by replacing it with "next" k-th candidate with the probability
5499 // of 1/k. It can be easily shown that by the end of the run, the
5500 // probability for any candidate is converged to 1/n, thus giving the
5501 // uniform distribution among all the candidates.
5502 //
5503 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5504 #define RANDOMIZED_DOMAIN_POW 29
5505 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5506 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5507 bool Compile::randomized_select(int count) {
5508 assert(count > 0, "only positive");
5509 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5510 }
5511
5512 #ifdef ASSERT
5513 // Failures are geometrically distributed with probability 1/StressBailoutMean.
5514 bool Compile::fail_randomly() {
5515 if ((random() % StressBailoutMean) != 0) {
5516 return false;
5517 }
5518 record_failure("StressBailout");
5519 return true;
5520 }
5521
5522 bool Compile::failure_is_artificial() {
5523 return C->failure_reason_is("StressBailout");
5524 }
5525 #endif
5526
5527 CloneMap& Compile::clone_map() { return _clone_map; }
5528 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5529
5530 void NodeCloneInfo::dump_on(outputStream* st) const {
5531 st->print(" {%d:%d} ", idx(), gen());
5532 }
5533
5534 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5535 uint64_t val = value(old->_idx);
5536 NodeCloneInfo cio(val);
5537 assert(val != 0, "old node should be in the map");
5538 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5539 insert(nnn->_idx, cin.get());
5540 #ifndef PRODUCT
5541 if (is_debug()) {
5542 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5543 }
5544 #endif
5545 }
5546
5547 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5548 NodeCloneInfo cio(value(old->_idx));
5549 if (cio.get() == 0) {
5550 cio.set(old->_idx, 0);
5551 insert(old->_idx, cio.get());
5552 #ifndef PRODUCT
5553 if (is_debug()) {
5554 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5555 }
5556 #endif
5557 }
5558 clone(old, nnn, gen);
5559 }
5560
5561 int CloneMap::max_gen() const {
5562 int g = 0;
5563 DictI di(_dict);
5564 for(; di.test(); ++di) {
5565 int t = gen(di._key);
5566 if (g < t) {
5567 g = t;
5568 #ifndef PRODUCT
5569 if (is_debug()) {
5570 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5571 }
5572 #endif
5573 }
5574 }
5575 return g;
5576 }
5577
5578 void CloneMap::dump(node_idx_t key, outputStream* st) const {
5579 uint64_t val = value(key);
5580 if (val != 0) {
5581 NodeCloneInfo ni(val);
5582 ni.dump_on(st);
5583 }
5584 }
5585
5586 void Compile::shuffle_macro_nodes() {
5587 if (_macro_nodes.length() < 2) {
5588 return;
5589 }
5590 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) {
5591 uint j = C->random() % (i + 1);
5592 swap(_macro_nodes.at(i), _macro_nodes.at(j));
5593 }
5594 }
5595
5596 // Move Allocate nodes to the start of the list
5597 void Compile::sort_macro_nodes() {
5598 int count = macro_count();
5599 int allocates = 0;
5600 for (int i = 0; i < count; i++) {
5601 Node* n = macro_node(i);
5602 if (n->is_Allocate()) {
5603 if (i != allocates) {
5604 Node* tmp = macro_node(allocates);
5605 _macro_nodes.at_put(allocates, n);
5606 _macro_nodes.at_put(i, tmp);
5607 }
5608 allocates++;
5609 }
5610 }
5611 }
5612
5613 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5614 if (failing_internal()) { return; } // failing_internal to not stress bailouts from printing code.
5615 EventCompilerPhase event(UNTIMED);
5616 if (event.should_commit()) {
5617 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5618 }
5619 #ifndef PRODUCT
5620 ResourceMark rm;
5621 stringStream ss;
5622 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5623 int iter = ++_igv_phase_iter[cpt];
5624 if (iter > 1) {
5625 ss.print(" %d", iter);
5626 }
5627 if (n != nullptr) {
5628 ss.print(": %d %s", n->_idx, NodeClassNames[n->Opcode()]);
5629 if (n->is_Call()) {
5630 CallNode* call = n->as_Call();
5631 if (call->_name != nullptr) {
5632 // E.g. uncommon traps etc.
5633 ss.print(" - %s", call->_name);
5634 } else if (call->is_CallJava()) {
5635 CallJavaNode* call_java = call->as_CallJava();
5636 if (call_java->method() != nullptr) {
5637 ss.print(" -");
5638 call_java->method()->print_short_name(&ss);
5639 }
5640 }
5641 }
5642 }
5643
5644 const char* name = ss.as_string();
5645 if (should_print_igv(level)) {
5646 _igv_printer->print_graph(name);
5647 }
5648 if (should_print_phase(cpt)) {
5649 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5650 }
5651 #endif
5652 C->_latest_stage_start_counter.stamp();
5653 }
5654
5655 // Only used from CompileWrapper
5656 void Compile::begin_method() {
5657 #ifndef PRODUCT
5658 if (_method != nullptr && should_print_igv(1)) {
5659 _igv_printer->begin_method();
5660 }
5661 #endif
5662 C->_latest_stage_start_counter.stamp();
5663 }
5664
5665 // Only used from CompileWrapper
5666 void Compile::end_method() {
5667 EventCompilerPhase event(UNTIMED);
5668 if (event.should_commit()) {
5669 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5670 }
5671
5672 #ifndef PRODUCT
5673 if (_method != nullptr && should_print_igv(1)) {
5674 _igv_printer->end_method();
5675 }
5676 #endif
5677 }
5678
5679 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5680 #ifndef PRODUCT
5681 if (_directive->should_print_phase(cpt)) {
5682 return true;
5683 }
5684 #endif
5685 return false;
5686 }
5687
5688 #ifndef PRODUCT
5689 void Compile::init_igv() {
5690 if (_igv_printer == nullptr) {
5691 _igv_printer = IdealGraphPrinter::printer();
5692 _igv_printer->set_compile(this);
5693 }
5694 }
5695 #endif
5696
5697 bool Compile::should_print_igv(const int level) {
5698 #ifndef PRODUCT
5699 if (PrintIdealGraphLevel < 0) { // disabled by the user
5700 return false;
5701 }
5702
5703 bool need = directive()->IGVPrintLevelOption >= level;
5704 if (need) {
5705 Compile::init_igv();
5706 }
5707 return need;
5708 #else
5709 return false;
5710 #endif
5711 }
5712
5713 #ifndef PRODUCT
5714 IdealGraphPrinter* Compile::_debug_file_printer = nullptr;
5715 IdealGraphPrinter* Compile::_debug_network_printer = nullptr;
5716
5717 // Called from debugger. Prints method to the default file with the default phase name.
5718 // This works regardless of any Ideal Graph Visualizer flags set or not.
5719 void igv_print() {
5720 Compile::current()->igv_print_method_to_file();
5721 }
5722
5723 // Same as igv_print() above but with a specified phase name.
5724 void igv_print(const char* phase_name) {
5725 Compile::current()->igv_print_method_to_file(phase_name);
5726 }
5727
5728 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5729 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5730 // This works regardless of any Ideal Graph Visualizer flags set or not.
5731 void igv_print(bool network) {
5732 if (network) {
5733 Compile::current()->igv_print_method_to_network();
5734 } else {
5735 Compile::current()->igv_print_method_to_file();
5736 }
5737 }
5738
5739 // Same as igv_print(bool network) above but with a specified phase name.
5740 void igv_print(bool network, const char* phase_name) {
5741 if (network) {
5742 Compile::current()->igv_print_method_to_network(phase_name);
5743 } else {
5744 Compile::current()->igv_print_method_to_file(phase_name);
5745 }
5746 }
5747
5748 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5749 void igv_print_default() {
5750 Compile::current()->print_method(PHASE_DEBUG, 0);
5751 }
5752
5753 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5754 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5755 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5756 void igv_append() {
5757 Compile::current()->igv_print_method_to_file("Debug", true);
5758 }
5759
5760 // Same as igv_append() above but with a specified phase name.
5761 void igv_append(const char* phase_name) {
5762 Compile::current()->igv_print_method_to_file(phase_name, true);
5763 }
5764
5765 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5766 const char* file_name = "custom_debug.xml";
5767 if (_debug_file_printer == nullptr) {
5768 _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5769 } else {
5770 _debug_file_printer->update_compiled_method(C->method());
5771 }
5772 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5773 _debug_file_printer->print_graph(phase_name);
5774 }
5775
5776 void Compile::igv_print_method_to_network(const char* phase_name) {
5777 ResourceMark rm;
5778 GrowableArray<const Node*> empty_list;
5779 igv_print_graph_to_network(phase_name, (Node*) C->root(), empty_list);
5780 }
5781
5782 void Compile::igv_print_graph_to_network(const char* name, Node* node, GrowableArray<const Node*>& visible_nodes) {
5783 if (_debug_network_printer == nullptr) {
5784 _debug_network_printer = new IdealGraphPrinter(C);
5785 } else {
5786 _debug_network_printer->update_compiled_method(C->method());
5787 }
5788 tty->print_cr("Method printed over network stream to IGV");
5789 _debug_network_printer->print(name, C->root(), visible_nodes);
5790 }
5791 #endif
5792
5793 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5794 if (type != nullptr && phase->type(value)->higher_equal(type)) {
5795 return value;
5796 }
5797 Node* result = nullptr;
5798 if (bt == T_BYTE) {
5799 result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5800 result = new RShiftINode(result, phase->intcon(24));
5801 } else if (bt == T_BOOLEAN) {
5802 result = new AndINode(value, phase->intcon(0xFF));
5803 } else if (bt == T_CHAR) {
5804 result = new AndINode(value,phase->intcon(0xFFFF));
5805 } else if (bt == T_FLOAT) {
5806 result = new MoveI2FNode(value);
5807 } else {
5808 assert(bt == T_SHORT, "unexpected narrow type");
5809 result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5810 result = new RShiftINode(result, phase->intcon(16));
5811 }
5812 if (transform_res) {
5813 result = phase->transform(result);
5814 }
5815 return result;
5816 }
5817
5818 void Compile::record_method_not_compilable_oom() {
5819 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit());
5820 }